# Agents Finance — full content > A directory of real use cases of people and companies using AI and AI agents in accounting, treasury, consolidation, annual accounts, controlling and reporting, and in the most widely used financial ERPs, plus the tutorials, guides, prompts and resources people use today. Generated at build time. 18 entries sorted by type (real use cases first) and, within each type, newest first. Figures and claims belong to the source named in each entry; the level of evidence indicates reliability. Spanish edition: https://agentsfinance.io/es/llms-full.txt --- ## Ten weeks keeping the books of Swedish limited companies with Claude Code and Codex, with wins and mistakes URL: https://agentsfinance.io/entries/case-igdrasil-swedish-bookkeeping-claude-code-codex-2026/ Source date: 2026-09-27 (September 27, 2026) Type: Use case Who does it: Philip Eriksson (co-founder of Igdrasil) (Individual) — Co-founder of Igdrasil, a Swedish accounting platform. He uses his own platform as the agents’ tool, so this is not an independent source. Sector: Swedish limited companies in software, e-commerce, digital health, services and food Tools: Claude Code, Codex (OpenAI), Igdrasil (own CLI and MCP server), BAS chart of accounts and Swedish VAT rules Source language: English ERP: general Process: Accounting, Treasury Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: Igdrasil — Can AI do your bookkeeping? Lessons from Claude Code and Codex (Philip Eriksson) — https://igdrasil.se/en/resources/guides/can-ai-do-your-bookkeeping/ Verified: 2026-10-17 ### Summary Philip Eriksson, co-founder of Igdrasil, explains how, from mid-July 2026, he kept the books of several Swedish limited companies by talking to Claude Code and Codex. He provides volumes taken from his logs, a VAT error caught before filing and a table of agent failures. He is the author of an accounting tool. ### What it does The agents read supporting documents, propose an account and VAT treatment, match bank transactions to documents, review VAT before filing and post entries in batches. The author gives tasks in natural language (“see what is still unmatched in the bank”), the agent prepares drafts in a queue, he reviews and sets limits (“do not post supplier invoices, only receipts”), and then the agent posts and rereads the ledger to check. Accounting policy decisions, such as where a cloud expense goes, are made by a person once and the agent applies them to later rows. ### Reported result or impact Stated by the author, with data from his sessions between 15 July and 25 September 2026: 53 sessions with Codex (about 111 active hours) and 16 with Claude Code (about 25 hours), with batches of 195, 291 and 582 entries and about 1,570 in total in Claude Code. One agent caught an error in a VAT return days before the deadline and another discovered bank rows imported twice. His estimate, which he himself calls approximate, is that a weekend of work became an afternoon. Part of the volume was old backlog, and many entries were simple. ### What has really been tested The author did it himself with real companies, according to the logs he cites. Some figures were counted by the agent. The text is published by the vendor of the tool the agents use, so it should be read as an interested testimonial. He acknowledges failures: an agent posted 14 invoices to the wrong account, added test entries to the live books, treated rows that were only linked to drafts as posted and called a payroll import “verified” that landed in the wrong month. We have tested none of this. ### How to replicate it The author recommends starting in a test company, asking first only for read-only access and proposals, setting the rules before posting anything (series, recurring expense accounts, what to expect), letting the agent prepare and a person approve, asking for a reconciliation read from the ledger instead of a summary and widening the scope once a type of row has gone well several times. Include a duplicate check as a fixed step before posting and verify yourself everything that goes to the Swedish tax authority. The regulatory part is specific to Sweden. ### What to keep in mind - Responsibility for the books being correct stays with the company, not the agent. The text frames this for Sweden; in other countries you will need to check local rules. - The data the agent reads is sent to its AI provider (Anthropic or OpenAI). If it includes personal data, review data-protection rules such as the GDPR. - A document can contain hidden instructions to divert the agent (prompt injection); that is why the author prefers the agent to prepare in a queue and a person to approve. --- ## UNHCR uses AI to prepare its treasury cash forecast: what it did, what it gained and how to replicate it URL: https://agentsfinance.io/entries/case-unhcr-treasury-cash-forecast-ai-2026/ Source date: 2026-09-16 (September 16, 2026) Type: Use case Who does it: UNHCR, the UN Refugee Agency — treasury unit (Organization or institution) — Told by Hikaru Kozuki, senior treasury specialist and head of UNHCR's treasury unit. The project was developed with the United Nations International Computing Centre (UNICC), its technology partner for payments infrastructure. No commercial interest is declared: they do not sell the tool. Sector: International humanitarian organisation (United Nations system) Tools: AI cash forecasting model developed with UNICC (the source does not detail the type of model), ERP data: accounts payable and future-dated purchase orders, Almost ten years of cash flow history, Automated dashboard for management charts and tables, Workday (deeper integration planned as a next step) Source language: English ERP: general Process: Treasury, Reporting Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: EuroFinance — From the floor at EuroFinance 2026: UNHCR asks AI to do the work behind the cash forecast — https://www.eurofinance.com/news/from-the-floor-at-eurofinance-2026-unhcr-asks-ai-to-do-the-work-behind-the-cash-forecast/ Additional source: EuroFinance — Streamlining treasury operations: UNHCR's move towards integrated financial and treasury management solutions — https://www.eurofinance.com/news/streamlining-treasury-operations-unhcrs-move-towards-integrated-financial-and-treasury-management-solutions/ Additional source: UNICC — Digital Hub of Treasury Solutions (DHoTS) — https://www.unicc.org/what-we-do/services-overview/dhots/ Additional source: EuroFinance — How two treasurers automated cash management (ASML and Cloud Software Group) — https://www.eurofinance.com/news/how-two-treasurers-automated-cash-management/ Verified: 2026-10-17 Cover image: Four-step diagram of UNHCR's cash forecast: accounts payable, future orders and history data; date normalisation; model and dashboard; and human review and decision in treasury. ### Summary The treasury team at UNHCR, the UN Refugee Agency, explained at EuroFinance 2026 how it uses AI to collect, normalise and project its cash flows, with treasurers still responsible for reviewing and deciding. It is a real case told by its head; the 40% to 60% reduction in work is his estimate for once the model is stable, not an audited result. ### What it does UNHCR's treasury team stopped pulling data from several systems into spreadsheets by hand for its cash forecast. With UNICC it mapped the main cash flow drivers (accounts payable, future-dated orders), brought in almost ten years of history, agreed which date represents the cash event in each category and built a forecasting layer that combines current ERP commitments in the short term with budget-adjusted historical patterns in the long term, plus scenario assumptions. The presentation (charts and tables) is generated automatically. For a while the AI forecast ran in parallel with the manual one and both were compared with actuals. Treasurers still review the output and decide. ### Reported result or impact According to Kozuki, once the model is stable the work of collecting and consolidating data could fall by 40% to 60%; this is an estimate, not a published measurement. He states that they already spend less time collecting and normalising data, that the automatic dashboard reduces the effort of preparing charts and tables, and that the conversation with management has shifted from debating spreadsheets to debating assumptions. The process is not yet fully automated and some data still requires human intervention. No forecast accuracy metrics are published. ### What has really been tested UNHCR uses it in its treasury according to its head, who described it at an industry conference; it is reported by EuroFinance, the event's organiser. The model type, code, error metrics and cost are not published. We have not reproduced the system and have no access to its data; the steps to replicate it in this article are our own adaptation of the lessons the source gives. ### How to replicate it Start with the decision you want to improve (short-term liquidity, timing of FX purchases, investments), not with the technology. Run a small pilot with reliable, high-impact data. Before touching any model, write a data dictionary: cash flow categories, which date rules in each, how currencies and restricted funds are treated, and who owns each field. Use ERP commitments for the short term and history for the long term. Run the automatic forecast in parallel with the manual one for several cycles and compare both with actuals. Automate the dashboard and leave review and decision with the treasury team. ### Context: why a humanitarian agency needed a different way to forecast its cash Treasury forecasting is one of the processes where the gap between what AI promises and what actually works shows most. That is why a case told in detail by the person who runs it deserves attention: the treasury unit of UNHCR, the UN Refugee Agency. Hikaru Kozuki, senior treasury specialist and head of that unit, explained it at the EuroFinance International Treasury Management 2026 conference, and EuroFinance published the summary on 16 September 2026. The starting point is familiar to any treasurer. The team collected information by hand from different source systems, consolidated it in spreadsheets and then validated and adjusted it. Kozuki sums it up with a question: why were they spending so much time collecting data instead of analysing the result of the forecast? By the time the number was ready, the reality it was meant to reflect could have changed. In a humanitarian organisation, he recalls, the situation does not wait. Two pressures were added to that. Since mid-2025, changes in donor support and budget cuts made history a less reliable guide for anticipating cash needs. And UNHCR's headcount had shrunk without the workload falling in the same proportion. The forecast had to take into account future-dated transactions, historical trends, seasonality, currencies and changes in the operating environment: too many variables to keep going with a manual process. This is not a project that starts from zero. According to an earlier EuroFinance article, in September 2023 UNHCR modernised its treasury management systems and bank connectivity and integrated them with its cloud ERP and the Digital Hub of Treasury Solutions (DHoTS), a treasury platform shared by several UN agencies with technical support from UNICC. That base of integrated data is what makes the next step possible. ### Who does it and with what - **Protagonist:** UNHCR's treasury unit, with Kozuki as head and spokesperson for the case. - **Technology partner:** the United Nations International Computing Centre (UNICC), which already managed bank, payroll, supplier payment and donor receipt information for UNHCR. That prior relationship settled much of the concern about data security. - **Data:** accounts payable, future-dated purchase orders and almost ten years of history, plus scenario assumptions. - **Output:** a cash forecast and an automatic dashboard with the charts and tables for management. - **What the source does not say:** which type of AI model is used (statistical, machine learning or a language model), which tools it is built with, how much it cost or what forecast error it has. It is best not to fill those gaps with assumptions. ### How they built it, step by step [Image: Diagram of the forecast flow: data sources, data governance, forecasting layer combining short-term ERP and long-term history, automatic dashboard and human review in treasury; next steps are Workday, per-user scenarios and agents that suggest transactions.] The order matters as much as the pieces. Kozuki describes a deliberately iterative process in several phases. ### 1. Map the cash drivers and their sources The work started with the data, not the model. The team identified the main cash flow drivers (among them accounts payable and future-dated purchase orders) and which system each one came from. Treasurers took part from the beginning: they helped define which questions the forecast had to answer, which cash flow categories were relevant and what level of accuracy was needed. ### 2. Normalise: decide which date is the cash event This is where the most interesting problem in the case appeared. The different systems did not describe the same fact in the same way. A single transaction could have an approval date, an accounting date, a payment date, a date of receipt by the payee and a bank reconciliation date, and any of them could be used to place the cash movement in time. [Image: Timeline with five dates for the same transaction: approval, accounting date, payment date, receipt by the payee and bank reconciliation; the payment date is highlighted as an example.] To compare, in Kozuki's words, apples with apples, the project became partly an exercise in data governance: deciding which fields mattered, setting common definitions and establishing which date represented the real cash event in each category. ### 3. Build the forecasting layer and automate the presentation The resulting model combines current information with historical patterns. For the short term it leans more on ERP data, which is reliable because it reflects commitments already made; for the long term it uses historical trends adjusted to the budget conditions of the moment, plus scenario assumptions. The team also wanted the presentation to produce itself: instead of finishing the analysis and then building charts and dashboards by hand, the aim was to join collection, forecasting and visualisation in a single flow. ### 4. Run in parallel and earn trust The biggest difficulty, according to Kozuki, was not convincing anyone that AI could process data, but getting treasurers to trust the result. “It can’t be a black box,” he says: the team has to understand how the model arrives at its figures because it is the one explaining them to management. That is why, for a while, the AI forecast coexisted with the manual one. The team compared both with actual data and saw where the model was more or less accurate. Responsibility remained human: AI produces the forecast, but treasurers review it and decide. ### A concrete example for your team The most replicable part of the case is not the model but the data dictionary that comes first. This table is an **illustrative example** of how a mid-sized treasury team might start; it is not UNHCR data. | Cash flow category | Source system | Date that rules | Horizon where it is used | Data owner | | --- | --- | --- | --- | --- | | Supplier payments | ERP, accounts payable | Scheduled payment date | 0–13 weeks | Accounts payable manager | | Approved purchase orders | ERP, purchasing | Scheduled payment date according to terms | 4–26 weeks | Purchasing | | Payroll | Payroll system | Bank debit date | Whole horizon | Human resources | | Customer or donor receipts | ERP, accounts receivable | Bank value date | 0–13 weeks with data, then history | Collections manager | | Taxes | Tax calendar | Bank debit date | Whole horizon | Tax | With that table agreed, the first pilot can be limited to two or three categories with clean data and a single currency, as Kozuki himself recommends when he says that today he would start with something smaller. ### How to replicate it in your company These steps are our adaptation of the lessons the source gives; they do not describe UNHCR's internal system. 1. **Choose the decision before the tool.** Kozuki insists that AI is only a tool: first you need to know which decision you want to improve (short-term liquidity, timing of FX transactions, level of investments). That determines which data, which model and which controls you need. 2. **Take an inventory of cash drivers.** For each category, note which system it comes from, how often it is updated and how good its quality is. 3. **Write the data dictionary before modelling.** What counts as a forecastable flow, which date rules, how you treat currencies, how you record restricted funds and who answers for each field. It is what Kozuki says he would do even earlier if he started today. 4. **Separate horizons.** Use ERP commitments (registered invoices, approved orders) for the short term and adjusted history for the long term. Do not ask the model to guess what you already know. 5. **Start with methods you can explain.** A case from the same source illustrates it: at ASML, according to EuroFinance, the team tested twenty open-source Python algorithms against five years of actuals and retrains the model every month with a person reviewing its performance. What matters is not the algorithm but being able to explain why the result is what it is. 6. **Run in parallel for several cycles.** Keep the manual forecast while you compare both with actuals. Measure error by category and horizon, not just in total. 7. **Automate the presentation.** A dashboard that updates itself frees up time even if the model is not perfect. 8. **Leave the decision and execution in human hands.** UNHCR is exploring agents that suggest when to buy or sell currencies or place investments, but the transactions would be executed by people. ### Stated results and their source | Item | What is stated | Type of data | Source | | --- | --- | --- | --- | | Work of collecting and consolidating data | Could fall by 40% to 60% once the model is stable | Estimate by the head of the unit | EuroFinance, September 2026 | | Preparation of charts and tables | Less effort thanks to the automatic dashboard | Qualitative statement | EuroFinance, September 2026 | | Use of the team's time | Less time collecting and normalising, more analysing | Qualitative statement | EuroFinance, September 2026 | | Forecast accuracy | Not published | — | — | Kozuki stresses that the goal is not a perfect forecast but to make good decisions easier under uncertainty. According to him, seeing possible liquidity gaps and surpluses earlier helps decide precautionary balances and investment levels, and the conversation with management now centres on assumptions: which donor receipts are firm, which payments could be delayed and where risks appear. ### Limitations and risks - **Evidence:** it is a real case, but told by its protagonist at a conference and summarised by the organiser. There are no audited figures or error metrics. - **Incomplete process:** the source acknowledges that it is not fully automated and that some data still needs human intervention. - **Dependence on context:** UNHCR already had integrated treasury systems, almost ten years of history and a trusted technology partner. Without that base, the data effort will be larger. - **Less reliable history:** the case itself arises because the 2025 funding changes made the past a worse predictor of the future. A model based on history needs scenario assumptions and human review. - **Connection cost:** Kozuki points out that the main brake on integrating further with the ERP, including Workday, is the budget for connection costs, not technical capacity. - **Data security:** treasury data is sensitive. UNHCR solved it by relying on a partner that already held it; other companies will have to go through IT security, as Cloud Software Group describes in the EuroFinance article on ASML. ### Common mistakes this case helps avoid - **Starting with the model.** Without common definitions, AI amplifies inconsistency instead of resolving it, Kozuki warns. - **Mixing dates.** Using the accounting date in one category and the payment date in another produces timing gaps that look like model errors. - **Covering the whole organisation at once.** A narrow pilot shows value sooner and builds trust. - **Dropping the manual forecast too soon.** Without a comparison period against actuals you will not know where the model fails. - **Expecting AI to decide.** In this case AI gathers, processes and presents; interpreting uncertainty and deciding remains the treasurers’ job. ### Conclusions The UNHCR case is valuable precisely because it is unspectacular. It does not promise an autonomous treasury: it shows that the biggest job is putting the data in order, agreeing which date rules and earning the team’s trust with a parallel period. AI comes afterwards, to remove hours of collection and build the dashboard, and leaves treasurers the part that really adds value: discussing assumptions and deciding. If you want to replicate it, the first deliverable is not a model but a table of definitions and a small pilot. --- ## A CFO explains how their team reconciles prepaid expenses with an AI agent: from 2 hours to 5 minutes URL: https://agentsfinance.io/entries/case-onlycfo-cfo-prepaid-reconciliation-agent-2026/ Source date: 2026-05-15 (May 15, 2026) — Publication date of the article on OnlyCFO; the author says they had previously published it in their other newsletter, CFOpilot. Type: Use case Who does it: Author of OnlyCFO (a company CFO, no name visible on the page) (Individual) — Runs an accounting team and writes about AI in finance; the article does not name their company. Tools: AI agent with a “skill” (a SKILL.md of about 200 lines) and a YAML configuration file, MCP connector to the ERP (the article does not say which one), MCP connector to the shared drive (Google Drive, Box or SharePoint), Slack as trigger and as interface, Excel spreadsheets Source language: English ERP: general Process: Accounting Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: OnlyCFO — How to Build AI Agents for the Accounting Close — https://www.onlycfo.io/p/building-ai-agents-in-accounting Verified: 2026-10-17 Cover image: Four-step diagram: a Slack command reaches the AI agent, which reads the ERP general ledger and prepares a draft workpaper and journal entry that a human preparer and reviewer post. ### Summary The author of the OnlyCFO newsletter, who runs an accounting team, describes the agent they built to reconcile prepaid expenses against their ERP’s general ledger. According to them, it went from about 2 hours of human work to about 5 minutes, and the same pattern took a day off the monthly close. ### What it does An agent prepares the monthly prepaid expense reconciliation. When it receives the command from Slack, it queries the ERP general ledger directly through MCP, finds the previous month’s workpaper, copies it to the current close folder and updates it with the new items and the amortization month, keeping the Excel formulas. When it finishes, it posts a summary in Slack with the workpaper, a CSV with the proposed journal entry and a mention of the preparer. The agent does not post the entry: there is still a human preparer and a reviewer. ### Reported result or impact Stated by the author: from about 2 hours of human work to about 5 minutes, with more than 90% of the steps done by the agent, and the same pattern applied to other processes “took a full day” off the monthly close. The first run took more time, and afterwards the team kept refining the skill and the configuration. The article gives no per-process figures. ### What has really been tested According to the author, their team uses it in production. They do not publish the contents of the skill or the configuration, do not identify the ERP and do not measure the savings of the other processes. We have not reproduced the flow. A reader commented on the risk of prompt injection: the agent reads external text (supplier names, invoice descriptions), has access to sensitive data and can write to systems; human review comes after the agent has already acted. ### How to replicate it Pick a repetitive reconciliation with a monthly workpaper. Split what the author describes into two pieces: a skill (the procedure, which hardly changes) and a configuration file (ledger accounts, supplier-to-account map, materiality threshold, partial-month convention, who to notify). Connect the agent to the ERP and the workpaper folder in read mode at first, and let it prepare the draft without posting. Keep a human preparer and reviewer, and adjust the skill every time you correct something that repeats. ### The flow, step by step [Image: Flow diagram: Slack command, MCP read of the ERP general ledger, update of the previous month’s workpaper and Slack summary; a human posts the entry.] The diagram summarizes the steps the author describes: the command arrives from Slack, the agent queries the ERP general ledger through MCP, starts from the previous month’s workpaper and posts the result in Slack with a notice to the preparer. The two pieces (the skill and the configuration file) are the part their team keeps fine-tuning. ### Important nuances - This is one person’s experience at their company, told in a newsletter. The figures have not been audited. - The article mentions a specific ERP only as an example when defining what MCP is; it does not claim it is theirs. - That the agent does not post the entry is a control decision by the author, not a technical limitation stated in the text. --- ## A tax and accounting advisory firm in Madrid automates invoice intake with n8n and ChatGPT URL: https://agentsfinance.io/entries/case-fisconsulting-madrid-advisory-n8n-invoices-2025/ Source date: 2025-07 (July 2025) — The repository gives “July 2025” as the date of the work (deposited on 21 July 2025). Type: Use case Who does it: FISCONSULTING, tax and accounting advisory firm in Madrid (project by Álvaro Díaz Blanco, UPM) (Accounting or advisory firm) — Final-year thesis of the Degree in Information Systems at the E.T.S.I. de Sistemas Informáticos (UPM), supervised by José Eugenio Naranjo Hernández. Sector: Tax and accounting advisory Tools: n8n (on a Hostinger cloud server), Google Drive and Google Sheets, PDF.co, ChatGPT and Whisper, WhatsApp and Google Calendar, ALTAI (the program the sheet is imported into) Source language: Spanish ERP: general Process: Accounting Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: Archivo Digital UPM — Analysis and proposal of automation to optimize processes at a tax and accounting advisory firm (Álvaro Díaz Blanco; document in Spanish) — https://oa.upm.es/90055/ Verified: 2026-10-17 ### Summary A final-year thesis at UPM describes two automations built in n8n for FISCONSULTING, an advisory firm in Madrid: one that extracts and classifies invoices from Google Drive with ChatGPT, and another that handles appointments over WhatsApp. The author states that the time per invoice drops from 2 minutes to about 30 seconds. ### What it does The first flow detects new invoices in Google Drive, extracts their data with PDF.co and ChatGPT, normalizes the information, renames and classifies the files by year, quarter and type, removes duplicates and updates a Google Sheets spreadsheet prepared for import into ALTAI. The second, a “secretary agent”, receives appointment requests by WhatsApp as text or audio, interprets them with ChatGPT and Whisper, checks Google Calendar and schedules or changes the appointment, with automatic confirmations to the client. The work also covers access control in Drive, n8n logs and data retention in line with the GDPR and Spain’s LOPDGDD. ### Reported result or impact Stated by the author: processing time per invoice drops from 2 minutes to about 30 seconds (75% less) with fewer transcription errors; the secretary agent handles about 140 meetings and 50 inquiries a month with no manual intervention; and manual processing costs fall by more than 80%. The abstract also says the firm could take on more clients without expanding its staff. ### What has really been tested It is an academic work on a real firm, with implementation at the advisory firm itself according to its abstract. We have read only the repository record, not the full 11 MB PDF, so we could not check how the figures were measured or how many months of use are behind them. We have not reproduced the flows. ### How to replicate it Start from an invoice folder with a fixed structure (year, quarter, type). In n8n, build a flow that triggers when a new file arrives, extracts the text, asks a model for the invoice fields in a fixed format and writes a row to a sheet with the columns your accounting program expects. Check for duplicates before writing, keep logs of each run and review a sample by hand during the first weeks. Restrict access to the folder and validate the processing of client data with your data-protection advisor. ### Important nuances - The repository abstract talks about “agents”, but in practice they are automation flows in n8n that call a model to read documents and understand messages. - The license given in the repository is Creative Commons Attribution-NonCommercial. This entry summarizes the content in our own words and links to the original. - Invoices contain third parties’ personal data: before replicating this, review with whom the data is shared when using external extraction and AI services. --- ## U.S. AutoForce saves more than 30 hours a month on reconciliations with Copilot for Finance URL: https://agentsfinance.io/entries/us-autoforce-copilot-finance-bank-reconciliations/ Source date: 2025-01-28 (January 28, 2025) — It dates from January 2025, earlier than the rest of the entries, but it remains one of the few named cases with stated figures. Type: Use case Who does it: U.S. AutoForce (U.S. Venture) (Company) — A division of U.S. Venture; its finance team has used Dynamics 365 Finance since 2021, according to the source. Tools: Dynamics 365 Finance, Microsoft 365 Copilot (finance agent in Excel) Source language: English ERP: Microsoft Dynamics 365 Process: Accounting, Treasury Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: Microsoft Customer Stories — U.S. Venture gains efficiencies and insights with Dynamics 365 and Microsoft 365 Copilot for Finance — https://www.microsoft.com/en/customers/story/20600-us-venture-microsoft-365-copilot-for-finance Additional source: Microsoft Dynamics 365 Blog — Reinventing business process with AI: Agents in record to report (30 September 2025) — https://www.microsoft.com/en-us/dynamics-365/blog/business-leader/2025/09/30/reinventing-business-process-with-ai-agents-in-record-to-report/ Verified: 2026-10-17 ### Summary U.S. AutoForce, a division of U.S. Venture that uses Dynamics 365, automated its recurring credit card and bank statement reconciliations with Microsoft Copilot for Finance. According to Microsoft, it took four weeks to roll out and its accounting team saves more than 30 hours a month. ### What it does The finance team deployed the Microsoft 365 Copilot finance agent, which works in Excel, on two recurring processes: reconciling credit card transactions and reconciling bank statements. The company has used Dynamics 365 Finance since 2021. According to Microsoft’s blog, for this kind of reconciliation the matching logic and tolerances are defined in Excel and templates are saved to repeat the process. ### Reported result or impact Stated by the source: more than 30 hours a month saved by the accounting team and, in the words of a finance manager, an 80% cut in the time spent on US bank reconciliations after three months of use. Next steps mentioned: supplier statements (1.5 hours each by hand) and cash reconciliation. ### What has really been tested The company itself has used it, according to the testimonial Microsoft publishes. We have not tested the tool or audited the figures; the story is told by the vendor. ### How to replicate it Pick a repetitive monthly reconciliation based on Excel files, such as a corporate card or a specific bank. Note how long it takes today and how many items are left unmatched, define the matching rules and tolerances, and test on a month that is already closed so you can compare the result with a person’s. Save the flow as a template and review exceptions by hand. Measure after three months, as this company did, before taking the figure as good. ### Important nuances - This is a testimonial published by the vendor itself. The figures are stated, and we could not verify them independently. - The 80% figure refers to a specific process (US bank reconciliations) and the 30 hours is the accounting team’s monthly saving; they are not an overall finance saving. - The story also describes an invoice automation with EDI and OCR (10,000 to 15,000 invoices a day) that is not the AI agent: these are separate improvements and should not be mixed up. --- ## An accountant’s tutorial: reconciling accounts with Claude Cowork using a folder, an instructions file and a junior-style review URL: https://agentsfinance.io/entries/tutorial-reconcile-accounts-claude-cowork-close-folder-2026/ Source date: 2026-03-31 (March 31, 2026) — LinkedIn shows “6 months ago” with no exact date. The date is inferred from the post identifier and matches the creation of the GitHub repository on the same day. Type: Tutorial Who does it: Tyler Whitlock, CPA, MBA (Individual) — A certified accountant who posts about Python, databases and “vibe coding” for accountants. Tools: Claude Cowork, CSV, XML and PDF exports from the company’s tools, Instructions file (claude.md), GitHub repository Source language: English ERP: general Process: Accounting Level of evidence: Real case — An identified person or organization states that they have used it. The results are stated by the source; we have not audited them. Main source: LinkedIn — Tyler Whitlock, CPA, MBA: Reconciling Accounts with Claude Cowork (walkthrough) — https://www.linkedin.com/posts/tylerwhitlock50_tb-activity-7444829615598088192-BbQp Additional source: GitHub — tylerwhitlock50/month_end_claude (folder structure and reconciliation instructions) — https://github.com/tylerwhitlock50/month_end_claude Verified: 2026-10-17 ### Summary Tyler Whitlock, a certified public accountant (CPA), explains on LinkedIn the six steps he used to tie a close’s accounts to the trial balance with Claude Cowork. He says 2–3 days of work were reduced to an hour of review, and he published the folder structure and instructions on GitHub. ### What it does It is a method for an accounting professional to use an assistant with file access to prepare close reconciliations. It is more like putting an intern to work with a file system, clear instructions and a review cycle than chatting with a chatbot. ### Reported result or impact Stated by the post’s author: from 2–3 days of heavy, interrupted work to a little over an hour of review. The author insists that it is not an AI “doing the accounting” but an orderly environment with rules and a review cycle. Accounts such as the bad-debt provision needed several rounds of comments. ### What has really been tested The author did it himself with the data from his March close; he does not publish the data or the account figures. The repository contains the process template and examples of instructions, not the real data, and it does not declare a formal license (it only invites using and adapting the template). We have not reproduced the flow. ### How to replicate it Try first with a close that has already been reviewed and with one or two accounts, to compare the result with a person’s work. Remove client names and personal data from the folders before giving an AI assistant access and check your provider’s terms. Follow the steps below and do not post anything based on the assistant’s work without a person reviewing it. 1. Create a working folder for the close (for example, MARCH_RECON) and give Claude access only to it. Inside, add subfolders for documentation, evidence, inputs, reports, scripts and the trial balance. 2. Drop into the inputs folder the CSV, XML and PDF files you export from your programs. Rename only the files that are opaque. 3. Write a short text file with what you want: review the documents and tie them to the trial balance, use a fixed tag (#TB:account:balance) so you can check them later with a script, create one file per account explaining how it was reconciled and keep a summary and a to-do list. 4. Give that file to Claude and ask it to rewrite it and to ask you whatever questions it needs. Save the final version in the folder as claude.md. 5. Tell it to start reconciling. 6. Review the result as you would a junior accountant’s, ask for changes and have Claude document each new instruction for next time. Fix logic and edge cases and iterate. ### What to keep in mind - The method does not depend on a specific ERP: it starts from exports of the programs you use. - The useful idea is the fixed tag so that a script can check that each reconciled balance matches the trial balance. It is a mechanical check that does not depend on the assistant being right. - In the LinkedIn post, the accountant adds that the repository includes his instructions; we do not copy those texts and link to the original. --- ## Guide for bookkeepers: Claude as an assistant configured with a CLAUDE.md, what to delegate and what to always review URL: https://agentsfinance.io/entries/guide-claude-for-bookkeepers-claude-md-rules-2026/ Source date: 2026-07-31 (July 31, 2026) — Publication date; the page was last updated on 15 October 2026. Type: Guide / manual Who does it: Michael Robichaud (Cowork Consultant) (Individual) — A consultant who configures Claude for bookkeepers and offers implementation services; the guide links to his commercial offer. Sector: Bookkeeping and small firms (with US examples, such as 1099 forms) Tools: Claude (workspace with memory, files and a CLAUDE.md file), Bank and ledger exports, Aging reports (AR/AP) Source language: English ERP: general Process: Accounting, Treasury Level of evidence: First-hand experience — Someone explains how they do it or did it, without verifiable figures. It is documented practice, not proof that it will work in your case. Main source: Cowork Consultant — Claude for Bookkeepers: A Practical 2026 Guide (Michael Robichaud) — https://coworkconsultant.com/blog/claude-cowork-for-bookkeepers/ Verified: 2026-10-17 ### Summary Michael Robichaud, a consultant who rolls out Claude at accounting firms, explains how to set up the rules file (CLAUDE.md), which tasks the assistant can prepare (reconciliation, categorization, client reminders, close) and which always need human review. It is a consultant’s guide with an associated commercial service. ### What it does It explains five workflows the assistant prepares as drafts: bank and card reconciliation, transaction categorization, collection and payment reminders based on an aging report, a monthly close checklist with notes, and preparing contractors for 1099 forms. It includes a CLAUDE.md template with sections for chart of accounts, categorization rules by vendor, clients and their particularities, close checklist, message tone and tasks that are never done without review. It devotes a section to the confidentiality of client data. ### Reported result or impact The guide does not quantify savings. It states that the assistant works better as a configured drafter than as autonomous accounting software, that the most important factor is a well-made CLAUDE.md and that a person must review every figure before it reaches the client. ### What has really been tested The author says he has spent months configuring Claude for bookkeepers and describes what he has “seen hold up”, but provides no named cases or measurements. The guide is also a pitch for his implementation service. We have not tested the recommendations. ### How to replicate it Adapt the template’s structure to your regulations and your chart of accounts (the 1099 form examples are from the US and do not apply in Spain). Set it so that a new or doubtful vendor is flagged for review and not guessed. Read your AI provider’s terms and the confidentiality clauses with your clients before uploading data. 1. Start with a single workflow, usually bank reconciliation, which is the most repetitive and the easiest to check against the statement. 2. Write a thin CLAUDE.md for that workflow, with your chart of accounts and your matching rules, and nothing more. 3. Run it on last month’s real statement and compare the draft with what you know to be correct. 4. Run a pilot with a single client before extending it to the rest of your portfolio. 5. Add a “Never without my review” section covering contractor classification, payroll and anything that affects a tax return. ### What to keep in mind - The most reusable part is the separation between what is delegated as a first draft (matching transactions, suggesting categories for usual vendors, drafting reminders) and what always requires review (new or ambiguous vendors, payroll, tax returns, entries with estimates). - The guide contains calls to hire the author’s services. We have summarized only the practical content. --- ## Prompt for an AI assistant to prepare a draft accounting reconciliation without posting anything URL: https://agentsfinance.io/entries/prompt-draft-reconciliation-agent-no-posting/ Source date: 2026-09-27 (September 27, 2026) — Date of the main source. The prompt text is this directory’s own wording, inspired by the three sources, and is not text from any of them. Type: Useful prompt Tools: Any AI assistant with file access (for example, Claude, ChatGPT or Codex), Exports from your ERP or bank (CSV or Excel) ERP: general Process: Accounting, Treasury Level of evidence: Untested proposal — Material written by this directory based on published practices. We have not tested it with real data: try it with sample data first. Main source: Igdrasil — Can AI do your bookkeeping? Lessons from Claude Code and Codex (Philip Eriksson) — https://igdrasil.se/en/resources/guides/can-ai-do-your-bookkeeping/ Additional source: LinkedIn — Tyler Whitlock, CPA, MBA: Reconciling Accounts with Claude Cowork (walkthrough) — https://www.linkedin.com/posts/tylerwhitlock50_tb-activity-7444829615598088192-BbQp Additional source: OnlyCFO — How to Build AI Agents for the Accounting Close — https://www.onlycfo.io/p/building-ai-agents-in-accounting Verified: 2026-10-17 ### Summary Instructions ready to adapt for an AI assistant that reconciles an account against the trial balance: it uses only the files provided, posts no entries, checks for duplicates, separates what ties out from what is doubtful and leaves a list of questions. It is our own wording based on practices published by accountants, not tested with real data. ### What it does It turns the practices described by three people in the field into a single set of instructions: give the assistant only the files it needs, ask it to prepare and not to post, check for duplicates as a fixed step, distinguish what has evidence from what is doubtful and require it to reread its own figures. One of the sources reports that an agent treated rows that were only linked to drafts as posted; that is why the prompt asks for a final figure tie-out instead of a summary. ### Reported result or impact There are no figures: it is a wording proposal. The sources it draws on report time savings (from two or three days to one hour of review in a close, or from about 2 hours to about 5 minutes in a reconciliation), but these are not results of this prompt. ### What has really been tested Nothing with real data. It is written from what the three sources report, without copying their texts, and we have not tested it with a specific AI assistant. Try it on a close that has already been reviewed and with anonymized data, and compare the result with a person’s before using it in a live close. ### How to replicate it Fill in the variables, attach only the essential files and use anonymized data at first. Ask for a single account, review each list the way you would review a junior accountant’s work and save in your rules any correction that repeats, so you do not have to explain it again. If your assistant can connect to the ERP, start in read-only mode. ### Prompt (copyable) License: Original text of this directory. You may use and adapt it freely; it is not text from the cited sources. Warning: do not paste confidential or personal data; always review the output. ```text Act as an accountant preparing a reconciliation for someone else to review. You do NOT post anything. Goal: reconcile account {{ACCOUNT}} for the period {{PERIOD}} against the balance in the trial balance. Input data (use ONLY these files; if something is missing, say so and stop): {{FILES}} Company rules: {{RULES}} Steps: 1. Check that the files are complete and that the totals tie out to the opening balance. List any inconsistency before continuing. 2. Match transactions using the simplest possible rule (amount, date within a tolerance of {{TOLERANCE_DAYS}} days, and reference). Do not match by similarity if there is more than one option. 3. Look for duplicates (same amount, date and description) and transactions that look imported twice. Do not delete them: flag them. 4. Split the result into four lists: matched with evidence, matched with doubt, unmatched and possible duplicates. 5. Group differences below {{THRESHOLD}} separately, without explaining them as final. 6. Do not invent causes. If you can only suggest a hypothesis, write “HYPOTHESIS” and the question that would need answering. Output: - A table with: reference, amount, status, evidence (file and row) and comment. - The balance per the files, the trial balance figure and the difference. - A numbered list of questions for the reviewer. - A proposal of adjusting entries ONLY as text (account, debit, credit, reason). Do not post or export them. Before finishing, reread your own figures and confirm that the lists add up to the total difference. If they do not add up, say so. ``` Variables: - {{ACCOUNT}}: Account number and name (for example, 572 Bank current account) or a description of the item to reconcile. - {{PERIOD}}: Month or fiscal year the reconciliation refers to. - {{FILES}}: List of the attached files and what each contains (bank statement, ledger, trial balance). - {{RULES}}: Your own criteria: usual accounts per supplier, sign conventions, what to do with fees or recurring items. - {{TOLERANCE_DAYS}}: Days of difference allowed between the bank date and the booking date (for example, 3). - {{THRESHOLD}}: Amount below which a difference is considered immaterial, by your own criterion. ### Limits - It does not replace review by a person or your ERP’s controls. An assistant can get amounts, signs and dates wrong. - If you paste tables with client names, tax IDs or IBANs into an AI service, that data leaves your environment. Anonymize it or use a business version with the safeguards your company has validated. - Files can contain text with hidden instructions aimed at the assistant. Do not give it write permissions on your systems while you test. --- ## Open-source MCP server to connect AI assistants to Odoo: read, aggregations and, if enabled, write URL: https://agentsfinance.io/entries/resource-community-mcp-server-odoo-2026/ Source date: 2026-10-10 (October 10, 2026) — Date of the latest published release (v0.9.0). The repository has existed since March 2025. Type: Community resource Who does it: ivnvxd and project contributors on GitHub (Individual) — An independent community project; it is not an official Odoo product. Tools: MCP server in Python (launched with uvx), Odoo 16 or later (MCP module for production; test mode over XML-RPC), Claude Desktop, Claude Code, Cursor or another MCP client Source language: English ERP: Odoo Process: Accounting, Reporting Level of evidence: Project documentation — Described by the project’s own documentation or README. It says what the project claims to do and how it is used, not that it works in your case: we have not tested it. Main source: GitHub — ivnvxd/mcp-server-odoo (README and releases) — https://github.com/ivnvxd/mcp-server-odoo Verified: 2026-10-17 ### Summary A community project on GitHub (MPL-2.0 license, more than 400 stars, latest release published on 10 October 2026) that lets Claude and other assistants query and modify data in an Odoo instance: invoices, customers, products and aggregations. It has a test mode that needs no module and a module for production. ### What it does It exposes Odoo as tools an assistant can call: search and read records, create, update and delete (respecting per-model permissions), count, inspect fields and compute groupings and sums on the server without pulling all the rows. There is an option, which must be expressly enabled, to call business actions such as posting an invoice or confirming an order. According to the README, test mode works with any Odoo that has XML-RPC, and for production it asks you to install an MCP module that adds access controls. ### Reported result or impact It publishes no savings figures. What it offers is a documented and maintained bridge so that an assistant can query Odoo data in natural language, for example to ask about invoices or totals per customer without exporting to Excel. ### What has really been tested This is the project’s own description in its README: continuous integration, test coverage and numerous stars indicate use and maintenance, but they do not prove that it works on your installation. We have not installed it or tested it with an Odoo instance, and we have not reviewed the code. ### How to replicate it Start on a test copy of Odoo, never in production. Install uv, create an API key for a user with minimal, read-only permissions and add the server to your client’s MCP configuration with the Odoo URL, the key and the database name. Ask read-only questions (pending invoices, totals per customer) and compare the answers with Odoo’s reports. Do not enable write or workflow actions until you have defined what the assistant may do and who reviews. ### What to keep in mind - An assistant with write permissions on the ERP can create or delete records by mistake. The level of risk is set by the permissions of the user you give it, not by the assistant. - What the assistant queries travels to its AI provider: review its terms and data-protection rules (GDPR) before using it with real customer data. - It is a community project: check the license (MPL-2.0), the maintenance pace and compatibility with your Odoo version. --- ## Gartner: AI in finance takes 9 to 10 months to deliver value in tasks such as accounts payable and receivable URL: https://agentsfinance.io/entries/gartner-cfo-survey-finance-ai-september-2026/ Source date: 2026-09-24 (September 24, 2026) Type: News ERP: general Process: Accounting, Reporting, Controlling and management accounting Level of evidence: Opinion / analysis — A study, survey or analysis by a third party. Useful as context; it does not prove that a specific tool works. Main source: Gartner — Gartner Says CFOs Must Take a More Disciplined Approach to Finance AI Investment (24 September 2026) — https://www.gartner.com/en/newsroom/press-releases/2026-09-24-gartner-says-cfos-must-take-a-more-disciplined-approach Verified: 2026-10-17 ### Summary A Gartner survey of 160 finance leaders, conducted from January to April 2026, indicates that data extraction, accounts payable and receivable automation and report generation usually deliver results in about 9 to 10 months. More complex cases, such as forecasting or insight generation, take longer. ### What it does It is not a tool but a market study. Gartner asks finance leaders how long it takes to get the expected value from each AI use case. It concludes that investment is concentrated in quick-return and productivity uses, such as reporting and process automation, and recommends not neglecting those that improve decision-making, risk management and revenue even if they mature more slowly. ### Reported result or impact Data stated by Gartner: 160 finance leaders surveyed between January and April 2026. Returns from data extraction, accounts payable and receivable automation and report creation generally arrive in 9 to 10 months. Gartner adds that low AI literacy is now the main barrier, ahead of talent, partly because agentic coding tools have reduced the technical difficulty. The figures per use case are in charts and in a paid report that we could not consult. ### How to replicate it Use it as a reference for setting expectations, not as a guarantee. For each AI pilot, set a realistic time-to-value (9 to 10 months for document, collections and payments automation according to this study, longer for forecasting), reserve budget for hands-on team training and review midway which initiatives to cut. If you run an advisory firm or a small business, start with the quick-return cases and leave the analytical ones for a second phase. ### Nuances - It is a press release from a consultancy that sells studies; the public summary does not include the complete figures per use case. - The sample is 160 finance leaders with no public detail on country, sector or size; it cannot simply be extrapolated to small businesses in Spain. - It does not discuss any specific ERP, which is why this entry appears as general context. --- ## Odoo 20 adds an AI accounting assistant and agents that automate processes URL: https://agentsfinance.io/entries/odoo-20-ai-accounting-assistant-2026/ Source date: 2026-09-24 (September 24, 2026) Type: News ERP: Odoo Process: Accounting, Treasury, Reporting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Odoo — Meet Odoo 20 (24 September 2026) — https://www.odoo.com/blog/odoo-news-5/meet-odoo-20-2439 Additional source: Odoo Experience 2026 — Your accounting on autopilot with Odoo 20 (talk) — https://www.odoo.com/event/odoo-experience-2026-9099/track/your-accounting-on-autopilot-with-odoo-20-10030 Additional source: Odoo 19 — Document digitization (documentation) — https://www.odoo.com/documentation/19.0/applications/finance/accounting/vendor_bills/invoice_digitization.html Verified: 2026-10-17 ### Summary With Odoo 20, presented on 24 September 2026, the system adds an accounting assistant that answers questions about reports, audits reconciliations and lets you pay bills from Odoo, plus agents that create automations from natural-language instructions. It is a vendor announcement; the entry includes no customer results. ### What it does The accounting agent answers questions such as how much you are owed or what your cash flow is, drawing on report data. If you suspect a reconciliation error, Odoo runs a cycle-by-cycle audit and returns a summary with recommendations. Bills can also be paid from Odoo, one by one or in bulk. Outside accounting, an agent lets you create automations by describing them in natural language and, if you prefer an external AI tool, you can connect the database with access rights intact. The Odoo Experience talk adds that Odoo 20 automates deferrals, loans and fixed assets and uses AI predictions on accounts and analytics to speed up reconciliations. ### Reported result or impact The blog gives no figures. The Odoo Experience talk promises that automations will save “countless hours” every month, a marketing claim without measurement. It also does not clarify availability by edition or for the Spanish localization. ### How to replicate it Before migrating, review the full Odoo 20 release notes and confirm with your integrator the availability of these features in your edition and with the Spanish localization. If you already use Odoo 19, you can try vendor bill digitization, which uses OCR and AI, runs on pay-as-you-go credits (one credit per document; XML files do not consume any) and offers to enable automatic posting for a vendor after three validated invoices without changes. Test it first with recurring vendors and compare the drafts with what a person would post. ### Nuances - It is a launch blog post: it describes what the product does in general terms and does not go into limits, AI costs or configuration. - The digitization documentation cited is for version 19; we do not confirm that Odoo 20 keeps exactly the same credit model. --- ## Holded launches its MCP server to connect Claude, ChatGPT and other assistants to your account URL: https://agentsfinance.io/entries/holded-mcp-server-claude-chatgpt-2026/ Source date: 2026-06 (June 2026) — Holded’s documentation says “launched in June 2026” without a specific day. Type: News Source language: Spanish ERP: Holded Process: Accounting, Treasury, Reporting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Holded — MCP server | Holded API (page in Spanish) — https://www.holded.com/es/desarrolladores/mcp Additional source: Holded — MCP tools reference (page in Spanish) — https://www.holded.com/es/desarrolladores/mcp/herramientas Additional source: Holded Academy — Connect Holded with AI (help article in Spanish) — https://help.holded.com/es/articles/15646468-conecta-holded-con-ia Verified: 2026-10-17 ### Summary In June 2026 Holded released a remote MCP server that lets AI assistants query and, if the user authorizes it, modify data in their account: invoices, contacts, treasury and stock. It connects with OAuth 2.1 and the assistant inherits the permissions of the Holded user who authorizes it. ### What it does It exposes the Holded API as tools that an MCP-compatible assistant can call from a single remote URL (https://mcp.holded.com/mcp, with OAuth 2.1 over Streamable HTTP). The reference mentions 356 tools, one per API operation, grouped by area. The assistant can, for example, list invoices, search contacts, check treasury balances or check stock. When authorizing, the user chooses a profile: analyst (read-only), operational (read and write) or custom. ### Reported result or impact Holded publishes no savings figures. What it does document is the control model: the assistant can never do more than the user who authorizes it, and access is cut by deleting the token in Holded or disconnecting the connector. ### How to replicate it You need a Holded account with API access enabled. In Claude (web or desktop), add a custom connector with the MCP server URL and sign in to Holded; the page says one-click connection from the connector directory is coming soon. Start with the read-only profile, which Holded recommends, and try queries such as the list of invoices pending collection. If you enable writing, first check which areas are enabled, because the server page and the help guide do not describe exactly the same thing for what the operational profile covers, for example in payment recording or accounting entries. ### What to keep in mind - According to the help guide, the read-only profile leaves out sensitive areas by default, such as developer tokens, payroll and bank mandates or remittances. - The help guide sums up the philosophy of the operational profile like this: AI drafts and organizes, and the person issues and collects. Anything that moves money or has tax effects stays read-only unless it is deliberately widened. - Connecting a third-party assistant means that the data it queries passes through that AI provider. Review its terms and data-protection rules (GDPR) before using it with customer data. --- ## Sage Intacct expands collections and payables automation and opens its AI Gateway with MCP URL: https://agentsfinance.io/entries/sage-intacct-ai-ap-automation-mcp-may-2026/ Source date: 2026-05-28 (May 28, 2026) Type: News ERP: Sage (Intacct, 50…) Process: Accounting, Treasury, Reporting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Sage — Sage Intacct expands trusted automation across core finance workflows (press release, 28 May 2026) — https://www.sage.com/en-us/news/press-releases/2026/05/sage-intacct-expands-trusted-automation-across-core-finance-workflows/ Verified: 2026-10-17 ### Summary In May 2026 Sage announced payment reminders, AI-powered three-way matching in accounts payable and an MCP-based AI Gateway to connect external AI tools to Sage Intacct. Some features are generally available and others are in an early-adopter program, with differences by country. ### What it does Three-way matching joins the invoice, the purchase order and the receipt, compares prices, quantities and amounts and flags discrepancies by line before payment; it is generally available worldwide. Payment reminders show customers with open or overdue invoices in a single view and allow sending notices with one click or in bulk (global early-adopter program). Custom approvals in purchasing define rules with several conditions, such as supplier, amount or department. SaaS Intelligence 2.0 adds forecasts and cohort analysis for software companies. The AI Gateway connects AI tools to Intacct through a REST API and MCP, respecting roles and permissions. ### Reported result or impact The press release includes no customer figures. It offers a quote from an implementation partner and a Gartner data point on CFO priorities, neither of which demonstrates the effect of these features. ### How to replicate it If you use Sage Intacct, start with three-way matching: for a month, compare the discrepancies the AI flags with those your team detects. Pay attention to availability by region: according to the release, the AI Gateway is generally available in the United States, the United Kingdom, Canada, Australia and South Africa, so it may not be in your country. To test the AI Gateway, use an API user with minimum permissions and log which queries the external tool makes. ### Nuances - This is a vendor press release, with detail on what is generally available and what is in early adoption. - The emphasis is on control and traceability: Sage insists that automation respects approvals and permissions defined by the company. --- ## SAP unveils Joule assistants for close, collections, treasury and planning URL: https://agentsfinance.io/entries/sap-joule-autonomous-finance-assistants-2026/ Source date: 2026-05-12 (May 12, 2026) Type: News ERP: SAP Process: Accounting, Treasury, Controlling and management accounting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: CFO Dive — SAP ramps up push to bring AI agents to finance teams — https://www.cfodive.com/news/sap-ramps-push-autonomous-finance-agentic-ai/819929/ Additional source: SAP — Sapphire 2026 Innovation News Guide — https://www.sap.com/topics/events/sapphire/innovation-news-guide-2026 Additional source: SAP — Connect 2026: the five top announcements — https://www.sap.com/blogs/top-sap-connect-2026-announcements Additional source: SAP — Accounts Receivable Assistant (product page) — https://www.sap.com/use-cases/joule-assistant/accounts-receivable-ai Verified: 2026-10-17 ### Summary In May 2026 SAP announced a family of AI assistants for finance that coordinate specialized agents for the financial close, accounts receivable, billing, tax, treasury and planning. For now it is an announcement: the sources we reviewed give availability dates but no results measured at customers. ### What it does Each Joule assistant takes care of one area of finance and orchestrates several smaller agents. SAP mentions assistants for the financial close, financial planning, billing, governance, tax and compliance, accounts receivable and treasury. The accounts receivable assistant, for example, groups agents that prepare collection calls, send dunning emails, create and resolve disputes and propose offsets between receipts and payments. Joule Studio also lets customers build their own agents. ### Reported result or impact None of the sources we reviewed give measured figures. SAP states, in general terms, that the close assistant “can reduce” the length of the close cycle. According to CFO Dive, SAP itself acknowledges that challenges remain around data readiness, integration into workflows and team training. ### How to replicate it If you already work with SAP S/4HANA Cloud, ask your partner or SAP for the real status of each assistant in your edition and region, because the published timelines (second and third quarter of 2026 for several of them, fourth quarter for the governance one) may have changed. Start with a narrow case, such as prioritizing collections, with human approval for every action and a baseline indicator (days sales outstanding, open disputes) so you can compare afterwards. Before activating anything, review the quality of your customer and invoice master data. ### Context The announcement was made at SAP Sapphire 2026 and is part of SAP’s “Autonomous Enterprise” initiative. According to CFO Dive, SAP talks about more than 50 assistants that in turn coordinate more than 200 specialized agents across every area of the company, not just finance. ### What to keep in mind - This is a roadmap with availability dates per assistant (some in an early-adopter program), not a set of generally available features. - Availability may vary between S/4HANA Cloud Public Edition, Private Edition and regions. - SAP explains that agents act within the user’s permissions and that approval controls remain in place; that is a design promise that has to be verified in each implementation. --- ## Dynamics 365 Finance: the account reconciliation agent suggests how to resolve exceptions URL: https://agentsfinance.io/entries/dynamics-365-finance-account-reconciliation-agent-2026/ Source date: 2026-04 (April 2026) — The release plan covers April to September 2026 and gives no publication date; we use April 2026 as a reference for ordering. The page has been listed as archived since October 2026. Type: News ERP: Microsoft Dynamics 365 Process: Accounting, Treasury Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Microsoft Learn — Enhancements to Account reconciliation agent (2026 release plan, wave 1) — https://learn.microsoft.com/en-us/dynamics365/release-plan/2026wave1/enterprise-resource-planning/dynamics365-finance/enhancements-account-reconciliation-agent Additional source: Microsoft Learn — Account reconciliation (Dynamics 365 Finance) — https://learn.microsoft.com/en-us/dynamics365/finance/general-ledger/account-reconciliation Verified: 2026-10-17 ### Summary The account reconciliation agent in Dynamics 365 Finance compares subledgers with the general ledger, detects differences and proposes what to do with each one, with a rationale. Microsoft has kept it in preview since April 2025 and its 2026 release plan expects general availability in September 2026; we could not confirm that it has already happened. ### What it does It reconciles the subledgers (accounts payable, accounts receivable, tax and bank) with the general ledger. For each exception it shows a suggested action, an analysis and a rationale, and lets you resolve several at once by grouping them. Actions include creating an adjusting entry, matching items, accepting as is (for example, rounding differences) or reversing an entry. The last action can be undone by giving a reason. The 2026 plan adds more precise matching (1-to-1, 1-to-many and many-to-many), fuzzy-logic and tolerance rules, bulk actions, MCP integration and the inventory subledger. ### Reported result or impact Microsoft does not publish savings figures in these sources. It does make clear that the goal is to move from guided automation to an increasingly autonomous reconciliation orchestrator, and that future features (fixed assets, projects, intercompany) remain in the plan and may change. ### How to replicate it In Dynamics 365 Finance, open the Account reconciliation workspace, enable reconciliation per module and set a start date after the periods you already reconciled by hand. The automatic process runs every six hours by default and can be adjusted. Start with a single module and a single legal entity, review the agent’s suggestions before submitting bulk actions and check that the entries created carry the justification comment. Verify your Finance version (the documentation mentions improvements in 10.0.46) and availability in your region. ### What to keep in mind - The documentation itself states that part of what the release plan describes had not been released at the time of writing and that dates may change. - The agent proposes and the person confirms: bulk resolution actions ask for explicit confirmation before they are submitted. - The MCP support announced for the agent aims to let it coordinate with other systems, but we have not seen documentation of specific cases. --- ## Sage 50 adds an AI-powered mobile app to check sales, collections and invoices URL: https://agentsfinance.io/entries/sage-50-mobile-app-ai-small-business-spain-2026/ Source date: 2026-03-27 (March 27, 2026) — Date of the post on the partner portal. Sage says the feature has been available since 1 February 2026. Type: News Source language: Spanish ERP: Sage (Intacct, 50…) Process: Reporting, Treasury Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Sage Partner Portal (Spain) — New Sage 50 mobile app: mobility and artificial intelligence for small business management (27 March 2026; page in Spanish) — https://partnerportal.sage.es/2026/03/app-movil-sage-50-inteligencia-artificial/ Verified: 2026-10-17 ### Summary In February 2026 Sage Spain launched a Sage 50 mobile app aimed at managers and owners of small businesses, with voice activity summaries and queries on sales, customer payments and invoices. Sage says it validated it with a two-month pilot with 20 customers, without publishing savings figures. ### What it does The app syncs with Sage 50 and lets you check the state of the business, sales, customer payments and invoice status, and also create quotes from your phone. The AI part generates voice summaries of activity, identifies best-selling products or most active customers and automates tasks such as reminders and meeting scheduling. ### Reported result or impact Sage says the feature was designed after customer interviews and surveys and tested for two months with 20 customers, who confirmed ease of use and value for their daily operations. No metrics on time saved or accuracy are published. ### How to replicate it New Sage 50 customers get the feature when they sign up; existing ones can add it when they renew their license or ask for early activation. If you are an advisory firm (gestoría), ask your Sage 50 clients whether they are interested before renewal and assess which company data becomes accessible from the manager’s phone, so you can define who may use the app. ### What to keep in mind - It is aimed at the manager or owner, not the accounting team: it is a query-and-management layer, not an agent that posts entries by itself. - The source is Sage’s partner portal, that is, a communication from the vendor itself. --- ## Oracle launches Fusion Agentic Applications, with finance agents already available in Fusion ERP URL: https://agentsfinance.io/entries/oracle-fusion-agentic-applications-erp-agents-2026/ Source date: 2026-03-24 (March 24, 2026) Type: News ERP: Oracle Fusion Cloud ERP Process: Accounting, Treasury, Reporting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Oracle — Oracle Introduces Fusion Agentic Applications (press release) — https://www.oracle.com/news/announcement/oracle-introduces-fusion-agentic-applications-2026-03-24/ Additional source: Oracle Fusion Insider — Agentic AI in ERP: four agents you can use today (22 June 2026) — https://blogs.oracle.com/fusioninsider/agentic-ai-in-erp-four-agents-you-can-use-today Verified: 2026-10-17 ### Summary In March 2026 Oracle announced 22 agentic applications for Fusion Cloud, one of them dedicated to collections, and in June it confirmed four finance agents available in release 26B: Ledger, Expenses, Payables and Payments. This is vendor information; these sources include no customer figures. ### What it does Agentic applications are teams of agents with defined roles that work inside Oracle Fusion Cloud and respect its permissions, approvals and audit trail. The finance example in the announcement is the collectors’ workspace, aimed at collecting sooner and improving the conversion of promises to pay. In release 26B, Oracle describes four available agents. Ledger answers general ledger questions in natural language and flags variances. Expenses extracts data from receipts, matches it to card charges and triggers reimbursement. Payables standardizes supplier invoice intake. Payments assesses when to pay and which early-payment discounts to take. ### Reported result or impact Oracle describes expected benefits (collecting sooner, less manual effort, better working capital), but the sources we reviewed publish no figures and no customers with measured results. ### How to replicate it If you use Fusion Cloud ERP, check the adoption guide for each agent and your environment’s update schedule in the Oracle Help Center. Oracle’s own blog warns of several prerequisites: some agents depend on technical components managed by Oracle, the conversation with the Ledger agent is English-only in 26B, and Payments needs banking capabilities enabled and dynamic discounting programs already created. Test with real supplier invoices and document your policies before activating the Payables agent. ### Nuances - The March announcement covers 22 applications across finance, HR, supply chain and customer experience; not all of them are finance-related. - Oracle presents as a differentiator that the agents run inside the transactional system rather than as an external layer. That is a vendor architecture argument, not a proven result. - Oracle AI Agent Studio lets customers build their own agents and measure their return through built-in observability tools. --- ## NetSuite introduces Autonomous Close, AI bank reconciliation and planning agents URL: https://agentsfinance.io/entries/netsuite-autonomous-close-ai-bank-reconciliation-2026/ Source date: 2026-02-11 (February 11, 2026) Type: News ERP: Oracle NetSuite Process: Accounting, Treasury, Controlling and management accounting, Reporting Level of evidence: Announcement — A communication from the vendor itself or from a news outlet about a launch. It describes what the product promises, not verified results. Main source: Accounting Today — Oracle NetSuite enhancements feature concert of AI agents — https://www.accountingtoday.com/news/oracle-netsuite-enhancements-feature-concert-of-ai-agents Additional source: NetSuite — NetSuite 2026.1 Features New AI Close and Cash Management, AI Agents for Planning and Reconciliation and More (10 February 2026) — https://www.netsuite.com/portal/resource/articles/financial-management/netsuite-2026-1-features-new-ai-close-and-cash-management-ai-agents-for-planning-and-reconciliation-and-more.shtml Verified: 2026-10-17 ### Summary At SuiteConnect (February 2026), Oracle NetSuite announced a close supported by several agents that watch exceptions and progress, bank reconciliation with generative AI, and planning and reconciliation agents for its EPM module. It is a product announcement relayed by the trade press; it includes no customer results. ### What it does Autonomous Close brings together agents that work in the background: one detects anomalies, another tracks the status of close tasks and their effect on results, another analyzes variances (flux) and locates the cause. It also proposes provisional accruals for expected transactions, such as fees. Bank reconciliation with generative AI extracts transaction details and matches them to the general ledger. In EPM, a planning agent answers in natural language about trends and variances and simulates scenarios, an assistant configures reconciliations and another agent drafts variance explanations. There is also an MCP-based AI connector that lets tools such as Claude query NetSuite data. ### Reported result or impact There are no figures in the source. Evan Goldberg, NetSuite’s co-founder, spoke of reaching a “zero-day” close as an aspiration; it is a stated goal, not a measured result. ### How to replicate it If your company uses NetSuite, ask your partner which parts are already available in your version (the 2026.1 release article details those for EPM) and start with AI bank reconciliation, which is the most contained. Measure beforehand your automatic reconciliation rate and your close hours so you can compare. If you want to connect an external assistant through the MCP connector, create a role with read-only permissions and test with queries such as the overdue invoices from the last month. ### Nuances - The main source is a trade outlet’s account of the presentation; the feature descriptions come from what was said on stage. - The Accounting Today text itself distinguishes process features (close, reconciliation, payments) from analysis features (planning, profitability). - The natural-language bill payment feature relies on an integration with Bill. - Outside AI, but relevant for Spain: according to the 2026.1 release article, NetSuite extends its e-invoicing to Spain with Veri*Factu requirements (digital signature, unique identifiers and a verifiable QR code).