What a CFO should ask before deploying a finance agent in production
Published 28 April 2026
The pattern is now common enough to be predictable. A finance function pilots an AI agent. The pilot goes well. The vendor proposes a production deployment in the next cycle. The team is excited, the CFO is convinced, and the conversation in the boardroom is about timing rather than design.
That is the conversation where things go wrong.
This post is the eleven questions I would put on the table before a finance agent leaves pilot and enters production. The questions are not the ones the vendor will surface. They are the ones drawn from the categories of failure I have actually seen, not the categories the vendor warns you about.
What the agent is doing in production, in one sentence
The first question is the one that sounds trivial and almost never has a clean answer.
What is the agent doing in production, in one sentence, written down, signed off by the function head and the head of IT?
Vague answers fail this test. “Helping with the close” is not a deployment scope. “Drafting the journal entries for the standard accruals, posting them to the system with the GL code identified by the agent’s rules, flagging any entry where the variance from the prior period exceeds 15%” is a deployment scope. The difference is the difference between a controlled deployment and an undefined one.
If the function cannot write that sentence, the function is not ready to deploy.
Who reviews the output, and how often
The reviewer model is the second thing to nail down.
Every agent output should have a reviewer attached to it. The question is what level of review applies to which output. Full review of every output is unsustainable and removes most of the benefit. No review is unsafe. The right answer is somewhere in between, with the review intensity calibrated to the risk of the output.
The categories that need full review are easy to name. Anything that ends up in a regulatory filing. Anything that posts to the ledger. Anything that goes to the board. Anything that touches a customer or counterparty.
The categories where sampling is enough are the high-volume, low-risk outputs. The variance commentary on the management pack draft. The supporting analysis. The drafting of internal memos.
The third category, which the function will get wrong, is the apparently low-risk outputs that are not actually low-risk. The internal memo that becomes the basis for a strategic decision. The supporting analysis that gets attached to the audit file. Get the third category wrong and you have an agent operating with less review than it needs.
What the audit trail looks like, on a single output
Pick one output the agent has produced in the pilot. A reconciliation, a journal entry, a piece of variance commentary, a KYC file.
Can you reconstruct, from the audit trail alone, the inputs the agent saw, the data the connectors returned, the steps the agent took, the subagents that were called, the output that was produced, and the human review that followed?
If yes, the audit trail is the foundation of your production deployment. If no, you do not have a production deployment. You have a black box.
This is the question I would put first if I had to put only one on the table. The AI governance framework is the longer version. The NIST AI Risk Management Framework’s Agentic Profile (NIST AI 100-5) maps the same expectation to the Manage function: traceability and recoverability of agent decisions across the lifecycle. (Source: nist.gov/itl/ai-risk-management-framework.) If the agent’s audit trail does not meet a NIST-RMF-readable standard, the audit trail does not meet anyone’s standard worth defending.
What happens when the agent is unsure
The deployment design has to include what the agent does when its confidence is low.
The naive answer is “the agent escalates to a human.” The harder questions are which human, at what speed, with what supporting context, and what the agent does in the meantime. Does the agent stop the workflow and wait? Does the agent produce a flagged output and continue? Does the agent fail back to a manual process? The honest answer depends on the use case, and the function needs the answer documented before the deployment.
The function that has not thought through this question is the function that will discover the answer when the agent encounters an edge case in the second week of production.
What happens when the connector is wrong
The vendor connector is presumed reliable. It will not always be.
A connector returns stale data. The model the connector points to has changed its schema. The vendor has updated the data partner’s contract and the freshness has slipped. The connector returns nothing at all because the upstream service is down.
The function needs an answer to each. The agent should know what to do, the reviewer should know what to do, the workflow should not silently produce output on bad data. The data quality piece is the long version of why this matters. The deployment-time version is: what is the fallback when the data layer fails.
What the cost actually is, on a representative volume
The cost of running an agent at pilot volume is not the cost of running it at production volume.
The subscription scales. The model usage scales. The data fees on the connectors scale. The deployment surface fees scale. Some of those scale linearly, some sublinearly, and some have step functions that the vendor will not surface unless you ask.
The CFO should ask for the all-in cost on a year-one production volume, with the assumptions written down, and should test those assumptions against the actual usage after the first quarter. The AI in finance cost piece is the framework.
Who owns the agent’s performance going forward
The agent is a deployed system. Deployed systems need an owner.
The vendor owns the model. The connectors are owned by the data partners. The skill is configured by your team. The subagents are vendor-supplied. The deployment surface is operated by the platform. The reviewer is your function.
Who owns the performance of the system across all of those layers? The honest answer is somebody on your team, and the function needs to name the person. If nobody owns the performance, the performance will degrade and nobody will notice.
What the rollback plan is
Production deployments need a rollback.
If the agent is producing the wrong output, what is the path to switching it off? The technical path is usually fast. The operational path is usually not. The team has stopped running the manual workflow. The skill that the manual workflow required is degraded. The institutional knowledge of how to do the work without the agent is rusty within six months.
The function needs a rollback plan that includes the operational path back to the pre-agent workflow. It needs to be tested. It does not need to be invoked, but it needs to exist.
What the user is allowed to override
The reviewer who looks at the agent’s output has to be empowered to override it.
The trap is the agent’s authority. The agent’s output is fluent, evidenced, and confident. The reviewer’s override is the human voice. The function that does not give the reviewer the explicit authority to override the agent without consequence will end up with rubber-stamp reviews and an agent that operates effectively unsupervised.
The CFO should make the override authority explicit, in writing, before the production cutover.
What the regulator will ask
The financial services regulators are paying attention to AI in finance. The FCA’s April 2024 AI Update and December 2025 reaffirmation set a principles-based, Consumer-Duty-anchored expectation. The Central Bank of Ireland’s Regulatory & Supervisory Outlook 2026 named AI and technology risk as a supervisory priority, and S.I. 366/2025 designated the CBI as Ireland’s market surveillance authority under the EU AI Act. The EU AI Act’s high-risk obligations for credit scoring and insurance risk pricing under Annex III §5 are scheduled to apply from 2 August 2026. (Sources: fca.org.uk AI Update; Ireland AI Act framework; EU AI Act Annex III.)
They are not the only audience. The audit firm is. The board is. The audit committee chair is.
The question to put on the table before production: what is the question the regulator will ask about this deployment, and is the answer ready? Most teams have not thought about this. The good ones have a memo. The exceptional ones have rehearsed the conversation.
This is not a hypothetical. The board will ask. The auditor will ask. The regulator might. The function that has not prepared for the question is the function that gets ambushed by it. The Air Canada chatbot case (Moffatt v Air Canada, BCCRT 2024) made the precedent explicit on the consumer side: the company owns the output its AI produced. The same principle is being tested in financial services this year and next. The Klarna AI customer service rollback in 2025, where the company publicly reversed its decision to replace ~700 reps with an OpenAI-powered system, is the cautionary tale on the operational side. Both are recent enough that a CFO can cite them in a board paper without anyone calling the example dated.
What success actually looks like, in twelve months
The last question is the one that defines whether the deployment was worth doing.
What does success look like in twelve months? Not in the form of “the agent is running smoothly.” In the form of specific outcomes. The close is two days faster. The reviewer queue has cleared. The board pack quality has improved. The reviewer turnover has not gone up. The error rate on agent-assisted work is lower than the error rate on manual work.
Write the success measures down before the deployment. Test against them at the year mark. The function that does not define success will discover, twelve months in, that it cannot say whether the deployment worked. The CFO who skips that conversation will be the CFO who cannot answer the board’s question when it lands.
Where this lands
A successful pilot is not a successful production. The shift between the two is where most finance AI deployments lose their definition, and where most of the cost overruns and the incidents come from.
The eleven questions above are not the only questions worth asking. They are the ones that have surfaced the most failure in the deployments I have seen. The CFO who answers them before the deployment is the CFO whose deployment lands. The CFO who answers them after the deployment is the CFO writing the lessons-learned memo for the next quarter’s board meeting.
Choose which one yours is going to be.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader.