It depends less on the agent than on your books. A read-only agent answering questions from a clean ledger can be useful within weeks. An agent that posts journals or prepares payments takes longer, because it must be tested on your own closed periods, wired into approval routes and signed off by the controller before it touches anything.
What decides the timeline?
Four things, in roughly this order of weight.
Access to the systems of record. If the agent needs the ledger, the sub-ledgers, the bank feeds and the procurement system, each connection needs an owner, a service identity and a security review. A single ledger with a modern interface is quick. Five entities on three ledgers is not.
The state of your data. An agent learns your chart of accounts from your posted history. If cost centres were remapped and nobody cleaned up the old ones, the agent will copy the mess. Cleaning comes first.
The depth you want. Reading and drafting are faster to approve than executing. Anything that moves money needs the most evidence and the most sign-offs.
Your own approval cycle. Internal audit, the controller and sometimes the audit committee will want to see the design. Their calendars often set the pace.
For the general version of this question across functions, see how long it takes to deploy an AI agent. Regulated banks face extra steps, covered in how long it takes to deploy an AI agent in a bank.
How should the work be phased?
| Phase | What happens | Who signs it off | Typical depth at the end |
|---|---|---|---|
| Scope | Pick one process, name the owner, list sources | Process owner | None |
| Connect | Read access to the ledger and feeds | Systems owner, security | 1 |
| Replay | Run against two or three closed months | Controller | 1 |
| Shadow | Run alongside the team for a live month | Controller | 2 |
| Grant | Routine items within limits | Controller, internal audit | 3 |
| Extend | Next capability or entity | Process owner | Earned per agent |
The replay and shadow phases are where time goes, and they are where you should spend it. Skipping them saves a month and costs you the controller's trust.
Where should a finance team start?
Start with reconciliation. It reads from systems the agent can trust, produces work a person reviews anyway, and gives fast evidence of whether the agent understands your books. We argue the point in reconciliation is where finance agents earn trust.
Avoid starting with anything that depends on spreadsheets nobody owns. If your accruals live in a workbook with hard-coded links, fix the workbook or keep it out of scope. Our guide on letting an AI agent work in your finance spreadsheets covers how to bring them in safely when you are ready.
A worked example
Take a manufacturer with two entities on one ledger, aiming for an agent that matches bank lines and drafts reconciliations.
For illustration: scoping and connecting take three weeks, mostly waiting for a service identity. Replay on January and February takes two weeks and shows the agent mis-reads one bank's narrative format. A parser fix and a second replay take another week. March runs in shadow. In April, matching goes to depth 3 for items under 5,000 with an exact reference; everything else stays as drafts. Roughly ten weeks, with most of it spent on access and evidence rather than on the agent.
Adding supplier payment proposals afterwards is faster, because the access and the review routine exist, but it still needs its own replay.
What stays with a person throughout?
The controller decides when each phase is passed. Releasing payments, posting above materiality, reopening periods and signing the accounts stay with people at every phase. The finance agent depth matrix shows where each capability usually tops out.
Which controls keep a fast deployment safe?
- One process at a time. Each new capability starts again at read-only.
- Written pass marks before testing. So the deadline does not move the bar.
- Expiry built in. Certification lapses and must be renewed, which keeps the evidence current.
- A record from day one. Every verdict in replay and shadow is filed, giving auditors a history from before go-live.
The topic guide AI agents for finance: books you can audit shows how the pieces fit.
Questions people ask
Can we go faster by skipping the shadow month?
You can, but you lose the only test on live, messy data before the agent acts. If time is short, shorten the replay rather than the shadow run.
Does year-end change the plan?
Avoid granting new depth in the weeks before year-end. Auditors and the controller will have little time to review, and a new capability is better proven in a quieter month.
Who should own the deployment plan?
The finance process owner, with IT as a supplier. A plan owned by IT tends to measure connections made; a plan owned by finance measures whether the books are still right.