What people and companies actually do with AI and agents in finance and ERPs

Asurion calls a Copilot Studio agent from Power Automate whenever Dynamics 365 Finance invoice capture leaves an invoice in error, fixing item and PO mismatches before posting. Microsoft's customer story reports touchless AP rising from under 20% to over 70% in the Americas; the figures are vendor-published and the instructions are not public, so we add our own templates and a guard script.

An SAP architect at a consulting partner describes the three Joule Studio agents he built with customers to identify incoming payments, park or post them and run a three-way match in S/4HANA, and the versions he threw away. First-hand experience with paraphrased customer feedback and one unquantified measurement; no customer is named.

Trescloud, an Ecuadorian Odoo Gold Partner, connected Claude to its own Odoo, spreadsheets, Looker and Gmail so its finance team can ask for budget, MRR, overdue receivables, missing customer withholdings and travel-invoice checks. First-hand account by its CFO with no measured figures; Trescloud sells this set-up as Trescloud IA.

Chintan Prajapati, CEO of Satva Solutions, connected the open-source Hermes Agent to Gmail, his bank, Stripe and a second Zoho Books organisation for his Singapore SaaS, and compares it monthly with his accountant's books. First-hand account with his own unaudited estimates and seven errors he fixed; his firm sells such builds.

Alex Willen, who runs a fund that buys Amazon and Shopify brands, gave Claude Code a QuickBooks API key and now has it categorise what the bank rules miss, book the fees Amazon nets out of each payout, keep a ledger per brand and refresh cash-flow projections. First-hand account in his posts on X and his Substack, with his own figures, not audited.

Baker Tilly, a Sage Intacct reseller, built a proof of concept in which Claude reads receivers and vendor invoices through Sage Intacct's read-only MCP server, separates legitimate GRNI from residual balances and drafts the reclass entries in the journal import format for a person to upload. A consultancy's own account with no figures; we add a tested matching script.

Since Intuit Connect 2026, creating and sending QuickBooks invoices from Claude and ChatGPT is GA; Gemini, Perplexity and Muse stay in beta and Intuit's hosted MCP server is an invite-only pilot. We compare seven AI routes into QuickBooks invoicing, classify the 24 pilot MCP tools and ship a tested review script plus assistant rules. Evidence: Intuit docs; no customer case yet.

SuiteWorld 2026 shipped industry SuiteAgents and 45+ SuiteSkills that also run via Claude or ChatGPT over the AI Connector. We separate those surfaces from the 25 MCP Standard Tools, classify the three write tools and one execute tool, and ship a tested matrix plus Python checker. Brennan Industries builds with Claude and Codex, then deploys SuiteScript so routine work stops burning tokens.

cfo-bench is Tom Eberle's open-source (MIT) benchmark in which Claude Code and Codex agents keep the books of a fictional French company under a plan-then-approve protocol, graded by code. We re-computed its published results per request and per task type, and built a tested approval gate from its main finding: agents treat a generic OK as the answer to their own questions.

Anthropic has not published a manual for running a company without employees. What it showed at Code w/ Claude in May 2026 is routines: Claude Code sessions that start on a schedule, an API call or a GitHub event and run in Anthropic's cloud. This guide explains what they do, what they cost, their limits and how a finance team can use them for close drafts without letting an agent post.

On 29 October 2026 Microsoft added more than 25 finance skills and partner plugins to Copilot, including Finance for Cowork (generally available) and Accounting for Cowork (Frontier preview). This guide, based on Microsoft's announcement and its Learn documentation, explains what each plugin does, which ledgers it can reach and how a finance team can pilot it with its own tested skills and controls.

On 22 October 2026 Sublime Security found invoice-fraud emails that hid an instruction for AI email agents, telling them to forward roughly $88,000 invoices to a plus-addressed accounts payable address. This guide turns that security-vendor report (no agent forwarding or loss was reported) into controls, tested scripts and agent instructions for finance teams that let an agent read the AP inbox.

A complete guide to Anthropic’s open-source finance plugin for Claude Cowork and Claude Code: what each of its eight skills does, how to install it, which connectors work and a ready-to-copy prompt for every function. Built from the Apache-2.0 files and three first-hand tests published by practitioners; the prompts are our own and untested on real books.

Manus now runs Xero's open-source MCP server as a ready-made connector, and its new Cue app gives personal agents a wallet. In the server's code, 26 of 51 tools only read, 7 create drafts and 18 change the books at once; this guide sets up an agent that reads and drafts, needs a named approval for immediate changes and gets a weekly script check.

A user on X says his OpenAI dot reconciled his bank, Stripe, PayPal and QuickBooks records back to January 2025, and an OpenAI employee says two dots and his CPA sorted a tax errand in 20 minutes. This guide separates those unverified claims from what OpenAI, Intuit and Xero document, and gives a set-up to copy: smallest plugin permissions, custom rules, scheduled checks and a payout tie-out script.

Owners on X say Grok Bot fixed years of messy QuickBooks books and enters vendor bills from PDFs; others say QuickBooks blocked the bot’s browser after a few weeks. This guide separates what xAI documents from what users claim and gives a set-up to copy: connector, read-only first, written approval rules, routines and spot checks. Owners’ claims are unverified; the templates are ours.

Two accounting professors show in the Journal of Accountancy how a custom skill in ChatGPT for Excel cleans a messy general ledger export, reconciles it to the trial balance and logs every change, in about 10 minutes for 59,157 simulated rows. A tested method on simulated data, not a client result; we add a skill, formulas and a re-check script.

In October 2026 the largest US banks started blocking AI agents that log in with a customer’s password, while keeping read-only access through open banking providers. This guide sums up what each bank said, where the liability lands and how a finance team can let an agent work from the ledger’s bank feed without holding the bank keys. Based on bank statements, terms and press reports; the set-up is our own proposal.

Infinity Group, a UK Microsoft partner, runs Business Central's built-in Payables Agent on its own supplier invoices and reports a 50% faster month-end and one person handling 250 to 350 invoices a month, with every draft reviewed before it becomes an invoice. The partner sells the product; we add Microsoft's MIT-licensed agent instructions, its billing table and what other users report.

Ethical CFO, a Melbourne practice that rescues clients up to five years behind in Xero, has a Claude browser extension fill the Who, What and Why fields of Xero's bank reconciliation screen, never clicking OK, and a second Claude step index the client's inbox to find receipts for blank bank lines. Told by the consultant who built it; no measured accuracy or hours saved.

Anthropic’s new proactive agent, Claude Duet, chases unpaid invoices, prepares payroll and closes the month through the Small Business plugin, whose finance skills are public on GitHub under Apache-2.0. This guide is based on Anthropic’s documentation and the skill files; no tester has published a detailed finance case yet.