Finance is the best place to start because the work already has a language of evidence. Approvals, invoices, purchase orders, vendors, budgets, entities, journal entries, close tasks, collections, revenue recognition, accruals, and reimbursement policies all depend on records that must reconcile. The team is not just answering requests. It is protecting the operating truth of the business.
That makes finance a natural first wedge for agentic back office. A model can read an invoice, compare it to a purchase order, find the vendor record, detect missing receiving evidence, classify the spend, identify the budget owner, and prepare the exception packet. A human still decides the judgment call. The system removes the scavenger hunt.
The wrong goal is "make finance faster." Speed by itself is dangerous. A fast bad approval creates cleanup later. A fast policy answer can create inconsistent treatment. A fast close process that hides weak evidence is worse than a slower one. The better goal is fewer unresolved exceptions and less manual reconstruction.
Invoice exceptions are a clean example. Today, a mismatch often becomes a thread: AP asks the requester, the requester asks the vendor, finance checks the PO, someone searches Slack, and the controller eventually gets pulled in. An agentic loop should start by classifying the mismatch. Is the amount different? Is the vendor entity wrong? Is the PO missing? Is the receipt absent? Is the tax treatment unclear? Is there a duplicate invoice risk? Each class should map to a state and next action.
Expense policy is another strong loop. The agent can check category, amount, date, receipt quality, attendee list, travel policy, customer context, approval threshold, and previous exceptions. It can draft the decision and cite the relevant policy. It should not invent policy, shame the employee, or approve ambiguous exceptions without a human. The point is consistency with evidence, not robotic denial.
Close management is where the compounding value shows up. Month-end close is full of dependencies: reconciliations, accrual support, intercompany entries, deferred revenue checks, bank recs, variance explanations, and owner confirmations. Agents can gather support, flag stale tasks, explain why a task is blocked, and draft variance narratives. The controller gets a cleaner review queue instead of another dashboard.
The system-of-record boundary matters. Finance agents should not become side ledgers. Vendor masters, invoices, approvals, budgets, accounting entries, and close status need authoritative homes. The agent can work across systems, but final writes need explicit ownership and logs. Otherwise the company speeds up today's work by creating tomorrow's reconciliation problem.
The approval model should vary by risk. Low-dollar, policy-clean reimbursements might move to post-review sampling. High-dollar vendor payments need pre-approval. Entity, tax, payroll, revenue, or contractual issues require stronger gates. A credit memo may be drafted automatically but issued only after approval. A forecast comment can be prepared, but the finance owner still signs off.
Finance also gives leaders useful metrics. Did invoice exception age fall? Did duplicate vendor creation fall? Did reimbursement rework fall? Did close tasks arrive with better support? Did controllers spend less time chasing context? Did audit evidence improve? These are better indicators than the number of AI-assisted tasks completed.
The political benefit is important. Finance leaders are naturally control-minded. If an agentic loop works in finance without weakening controls, other back-office functions can trust the pattern. Procurement, legal, compliance, and HR can borrow the same language: object, evidence, state, approval, action, audit trail.
The main failure mode is hiding risk under pleasant summaries. Finance work must preserve uncertainty. If a vendor match is probabilistic, say so. If a policy exception is plausible but not clean, route it. If data conflicts across systems, show the conflict instead of resolving it silently. A confident wrong answer is more expensive in finance than an honest incomplete packet.
Start with one loop where the pain is obvious and the action is bounded. Invoice exception triage is better than "AI finance operations." Expense review is better than "automate finance." Close evidence collection is better than "AI month-end close." The narrow loop creates the proof.
A good finance agent should leave the controller thinking: I am reviewing decisions, not rebuilding context. That is the standard.
A finance pilot should also have a narrow no-go list. Do not start with tax treatment, revenue recognition, payroll corrections, bank movement, or anything that changes reported financials without a serious control review. Start where the agent can prepare work and expose problems before the accounting owner acts. That keeps the pilot useful without pretending finance risk is just another workflow setting.
The best early artifact is a packet, not an approval. For invoice exceptions, the packet might show PO, invoice, receipt status, vendor master, requester, budget owner, mismatch type, and suggested next step. For expenses, it might show policy citation, missing fields, prior exceptions, and recommended routing. Finance can then decide which classes deserve more delegation.
A clean pilot also creates a better relationship between finance and the rest of the company. Requesters learn what a complete finance request looks like. Approvers see the evidence before they click. Controllers see which policies generate recurring exceptions. The agent is not the star of the workflow; the improved operating discipline is.
Evidence note: Workday's financial-management product framing reflects how finance work depends on connected records, controls, and planning data using https://www.workday.com/en-us/products/financial-management/financial-management.html.
This is part 2 of 10 in Agentic Back Office.