Ninety-five percent of accounting firms now use AI in some capacity, yet only about one in five can point to a clear, measurable return from it. That gap between enthusiasm and payoff is the most important number in accounting software right now — because the next wave of tools is not asking you to click buttons faster. It is asking you to describe how your business works in plain English and let software build the workflow for you.
That is the promise of XeroForce, the natural-language AI agent builder Xero introduced in May 2026 and put center stage at Xerocon London that July. Instead of learning another rules engine, you write a prompt: how your month-end close runs, who must sign off, which limits apply. The software turns that description into an always-on agent that works across your accounting data and the other apps you use. Whether you run on Xero or not, the idea matters, because every major platform is racing toward the same destination. This guide explains what XeroForce actually does, what a small business could realistically automate with it, and the five rules to follow before you hand any part of your books to an agent.
What XeroForce Actually Is
XeroForce is a no-code builder for custom AI agents that handle financial workflows. It sits on Xero OS, the company's AI-native financial operating system, and works alongside JAX, Xero's agentic platform. The core interaction is a plain-language prompt. You describe how you work — your process steps, your review requirements, your client sign-off limits, even industry-specific tax rules — and XeroForce assembles an agent that carries out that workflow across Xero and connected third-party apps.
A concrete example from the launch: for a month-end close that involves a long checklist, you build the process one workflow at a time. One step might read, in effect, "on the 5th of every month, get clarification on uncoded transactions." Within seconds the agent is running, and it keeps working in the background over days or weeks — watching for events such as email replies or approaching filing dates — while you do other work.
Three design choices are worth understanding, because they are the difference between a demo and something you could trust with real money:
You define the guardrails in the same prompt. Review steps, approval thresholds, and sign-off limits are part of the workflow description, not a separate admin console. You decide what runs on its own and what waits for a human.
Every workflow keeps a visible trail. Each run shows what executed, when it ran, what data it touched, and what is still waiting on a sign-off. Xero frames this as "Accountable Intelligence" — automate the routine work while keeping human judgment visibly in charge.
Agents operate at scale, not one chat at a time. For accountants and bookkeepers managing hundreds of clients, a XeroForce workflow can run bulk actions across a whole practice, such as reconciling uncoded statement lines everywhere at once, instead of repeating the same conversation per client.
At a follow-up event in the US, Xero added a pre-built month-end agent plus templates for common workflows — reviewing document reconciliation status straight through to manual journals for prepayments and amortization. As of the announcements, XeroForce was in early access with broader availability rolling out, at a moment when Xero reported serving 5 million customers worldwide.
XeroForce did not arrive alone. The same Xerocon announcements included Smart Document Capture that reads source documents into the ledger, automatic bank reconciliation that matches transactions to bank feeds in real time, receipt chasing that flags transactions missing backup, integrations letting accountants query live Xero data from Claude and Microsoft 365 Copilot, and payment features that tailor collection follow-ups to each customer's history and inspect bills for anomalies before money goes out. The pattern is unmistakable: the ledger is becoming the hub that other software orbits.
Why a Build-Your-Own Agent Is Different From the AI You Already Have
Most AI in accounting software today is vendor-defined. The vendor decides which tasks get automated — matching bank lines, suggesting categories, extracting invoice totals — and you get whatever buttons they ship. If your process does not fit their automation, you adapt your process or keep doing it by hand.
Agent builders flip that relationship. The meaningful unit is no longer the feature but the workflow you describe. Compare the old options:
- Rules engines ("if the payee contains 'Shell', code to Fuel") are precise but brittle. They break on every variation the rule author did not foresee, and maintaining hundreds of them becomes its own job.
- Macros and templates replay fixed steps. They cannot ask a clarifying question, wait three days for an answer, then continue.
- A prompted agent carries intent plus constraints. "Chase every invoice over 30 days old, consolidate multiple invoices per customer into one reminder, escalate anything over $10,000 to me before sending" is one description that covers cases a rules list would need dozens of entries to approximate.
That is also why Xero describes the shift as moving from a system of record to a system of action. A system of record stores what happened. A system of action does the next step — drafts the reminder, prepares the journal, flags the anomaly — and shows you what it did so you can approve or correct it.
What a Small Business Could Realistically Automate First
Agent builders reward boring, repeatable work with clear boundaries. Here are the workflows closest to ready, roughly in order of payoff.
Month-end close, one checklist item at a time
Nobody should automate their entire close on day one. Pick the single step that eats the most time with the least judgment — often clearing uncoded transactions or confirming that every bank account reconciled — and build one agent for that step. Once it runs cleanly for two consecutive closes, extend to the next step. Xero's own framing encourages exactly this: build the close process one workflow at a time.
Chasing missing receipts and source documents
Few tasks are more automatable than noticing that a transaction lacks backup and asking the right person for it. An agent can scan for undocumented transactions, match them to the cardholder or purchaser, send the request, follow up once, and escalate what is still missing at month-end. The human work shrinks to answering the questions only a human can answer.
Getting paid faster
Collection follow-up is pattern work with high emotional friction — owners postpone it, and cash suffers. JAX-style payment features analyze each customer's payment history and adapt: consolidate several invoices into one nudge, time the reminder to when that customer usually pays, switch channels for non-responsive accounts. If you describe your own escalation policy in the prompt, the agent applies it consistently instead of whenever you find the courage.
Reviewing bills before money leaves
Bill protection — automatically inspecting each bill for unusual amounts, changed bank details, or unfamiliar suppliers — is the rare automation that both saves time and reduces fraud risk. Vendor bank-detail changes are exactly how payment-redirection scams succeed, and a tireless checker that flags every altered detail is worth more than the minutes it saves.
The Adoption Numbers Say: Enthusiasm First, Returns Later
Before rebuilding your processes around agents, look at where the profession actually stands.
Intuit's 2026 AI Impact Report, built on responses from more than 34,000 business owners across four countries plus anonymized data from over 5.3 million companies, found that 77 percent of US small and midsize businesses now use AI regularly — up from 48 percent in July 2024. Among those surveyed, 78 percent said AI improved productivity and 43 percent said it increased revenue. Adoption is no longer the question.
The harder question is payoff. A Financial Cents survey of nearly 500 accounting and bookkeeping professionals across North America found that 95 percent of firms use AI in some capacity, while only about one in five could identify a measurable return. As the company's co-founder put it, almost every firm uses AI, but very few are getting measurable value from it.
Both findings can be true at once, and together they carry the lesson: the tool is not the return. The return comes from redesigning a specific workflow — close, collections, document gathering — around what the tool does well, then measuring the before and after. Firms that sprinkle AI over unchanged processes get demos. Firms that pick one workflow and rebuild it get days back.
Five Rules Before You Hand Your Books to an Agent
An agent that touches your ledger inherits every weakness in how that ledger is kept. These five rules apply whether your agent comes from Xero, another platform, or a script you wrote yourself.
1. Clean Books In, Clean Automation Out
An agent does not fix a messy chart of accounts, years of uncleared suspense balances, or bank feeds that have not reconciled since spring. It accelerates them — miscoding faster, with more confidence, across more transactions. Before automating anything, reconcile every account, clear stale items, and make sure each account means exactly one thing. Automation multiplies the quality of its inputs; make the inputs good first.
2. Keep a Human Sign-Off on Anything That Moves Money
Use the guardrail features deliberately: approval thresholds, sign-off limits, escalation paths. A sensible starting policy is that agents may draft, categorize, remind, and flag — but only a human releases payments above a stated amount, posts manual journals to sensitive accounts, or changes vendor banking details. Write those limits into the workflow description itself, review them quarterly, and never widen a threshold just to stop the notifications. The notification is the control working.
3. Insist on an Audit Trail You Can Show Your CPA
Every automated action should answer four questions: what ran, when, what data it touched, and who approved it. That is the standard XeroForce advertises, and you should demand it from any tool you adopt. The audience for this trail is not only you — it is your accountant at year-end, your lender during a review, and potentially a tax examiner. If an agent cannot produce a readable log of its own work, it is a toy, not infrastructure.
4. Start With One Workflow and Measure It
Remember the one-in-five payoff gap. Pick one workflow, record the baseline (days to close, average days sales outstanding, hours spent chasing receipts), run the agent for two full cycles, then compare. If the number moved, expand. If it did not, you have learned something cheap instead of automating your whole operation on a hunch. Resist the urge to build ten agents in the first week; the second agent should be shaped by lessons from the first, not by launch-day optimism.
5. Know What Data Leaves Your Ledger
Agents gain power by connecting your books to other apps — email, chat, document stores, payroll. Each connection is a data flow you are responsible for: client financials moving into a chat tool, bank details passing through a third-party service. Inventory every connection, confirm each one is necessary, and prefer tools that process your data inside their compliance boundary rather than scattering it across vendors. Convenience that you cannot diagram is risk you cannot manage.
Your Ledger Is the Fuel: Why Record Structure Matters More Now
Here is the connection most coverage of agent builders misses. An AI agent reasons over your accounts, your categories, your history, and your attached documents. A well-structured ledger — consistent account names, reconciled balances, every material transaction documented — gives the agent clean material to work with. A tangled one gives it confident-sounding mistakes at machine speed.
That makes boring bookkeeping hygiene the highest-leverage AI preparation there is. Standardize your chart of accounts and stop inventing one-off categories. Attach source documents at entry time instead of during a panicked year-end hunt. Reconcile on a schedule so errors surface while memories are fresh. If you want a tour of how structured records work in practice, the documentation walks through organizing transactions so both humans and software can follow them, and the dashboard shows how clean data turns into reports you can actually review — including reviewing what an agent changed.
There is a second, subtler advantage to keeping your ledger in a transparent, version-controlled format: every automated change becomes visible and reversible. When each agent action lands as an explicit, reviewable entry with a full history behind it, the audit trail from rule 3 comes free with the architecture instead of depending on a vendor's logging page. Plain-text accounting is AI-ready in exactly this sense — structured enough for software to process, transparent enough for a human to verify.
Keep Your Books Agent-Ready From Day One
As build-your-own agents move from early access to everyday tools, the businesses that benefit most will be the ones whose records an agent can actually understand. Beancount.io provides plain-text accounting that gives you complete transparency and control over your financial data — no black boxes, no vendor lock-in. Get started for free and keep books clean enough for any agent, including the ones you build yourself.
