“AI CFO” sounds ambitious. It is also a poor starting brief.
A CFO does not perform one task. The role combines judgment, accountability, negotiation, risk appetite, financing, controls, and context that changes from week to week. Turning that title into an agent creates a vague mandate with a large blast radius.
An SME needs a smaller unit of adoption: one recurring cash decision.
For example:
- Which overdue accounts need action today?
- Which projected cash shortfall needs management attention this week?
- Which customer requires a credit review before the next order is released?
These are narrow enough to define, observe, and improve. They also matter directly to the business.
Finance interest is ahead of finance integration
FIS announced an Agentic Insights Suite for collections, liquidity, risk, and cash-flow forecasting. The company says the suite connects data from ERP, treasury, CRM, credit bureaus, and public sources, then uses that combined view to score risks and recommend actions.[1]
The announcement also cites Deloitte's Finance Trends 2026 survey: 46% of global finance leaders identified working-capital optimisation as a top area for agentic AI, while 14% had fully integrated AI agents into finance.[1]
Those are vendor-cited survey figures, not independently audited outcomes from the announced product. But the gap is useful. Interest is high because cash decisions are repetitive, data-heavy, and time-sensitive. Integration is lower because the information rarely agrees by default.
A customer may have one name in the CRM, another in the ERP, and several legal entities in the receivables ledger. The treasury forecast may use a different cut-off from the weekly management report. Credit terms may exist in contracts, email, or somebody's memory.
A conversational interface does not reconcile those definitions. It only makes the unresolved differences easier to ask about.
Start with recommendation authority
The first finance agent should not move money, approve payments, release credit holds, or modify source records.
Give it recommendation authority instead.
That means the agent can collect approved evidence, apply a defined method, prepare a recommendation, and show why a case needs attention. A finance leader keeps the final decision.
Consider collections prioritisation. A useful agent might prepare a daily list of overdue accounts using:
- invoice amount and age from the ERP;
- payment history from accounts receivable;
- open disputes from the CRM or service desk;
- promised payment dates from approved notes;
- current credit status;
- the account owner's latest verified update.
Its job is not to decide that every late payer should receive the same message. Its job is to prepare a reviewable queue: who needs attention, what evidence supports that recommendation, and which cases do not fit the normal rule.
That last category matters. A strategic customer with an active service dispute is not the same as a customer who repeatedly misses an agreed date without explanation. The invoice age may be identical. The business decision is not.
Run this one-week exercise
Do not begin with a platform selection. Use one week to design the decision first.
Day 1: Choose one cash decision
Write the decision as a question that a named finance owner answers on a regular cadence.
Good: “Which overdue accounts should the collections lead review today?”
Too broad: “How should we improve cash flow?”
Record the current baseline. Useful measures might include time spent preparing the queue, the number of records checked manually, missed follow-ups, or the time between an exception appearing and somebody owning it.
Day 2: List the required evidence
For each recommendation, specify the fields and source systems the reviewer needs.
Do not write “customer data.” Write the actual evidence: invoice number, due date, outstanding amount, currency, dispute status, promised payment date, credit status, account owner, and source timestamp.
Then settle definitions. What counts as overdue? Which date controls? How are part-payments handled? Which record wins when the CRM and ERP disagree?
If the team cannot answer those questions, the problem is not the model. The decision is not ready to be automated.
Day 3: Define the recommendation threshold
State what puts a case into the queue.
A simple starting rule could combine invoice age, value, broken payment promises, and open-dispute status. The agent may use judgment to summarise context, but the entry conditions should remain visible to the reviewer.
Also define what the agent may recommend. “Review for follow-up” is different from “send a legal notice.” Keep the initial action reversible and internal.
Day 4: Name the exception owner
List the conditions that stop the normal route:
- an open dispute;
- a changed bank detail;
- conflicting customer identities;
- a missing document;
- an amount above the agreed threshold;
- a recommendation based on stale data.
Route each exception to a named role, not to “the team.” The agent should package the evidence and state why it could not complete the normal recommendation.
The safe failure path is a useful output. It prevents the agent from guessing just to keep the workflow moving.
Day 5: Review one real cycle
Let the agent prepare the queue in draft mode. The finance owner reviews every recommendation and records three outcomes:
- accepted;
- changed, with a reason;
- rejected, with a reason.
Do not judge the pilot by how polished the summary looks. Judge it by whether the right cases reached the right person with enough evidence to decide.
Measure the decision, not the demo
A successful first cycle should produce a small set of operational measures:
- preparation time before and after;
- percentage of recommendations accepted unchanged;
- number and type of exceptions;
- missing or conflicting data fields;
- time from exception detection to owner assignment;
- any case that would have created the wrong action if a human had not reviewed it.
The misses are not an embarrassment. They reveal where the workflow, data, or instructions need work.
After several cycles, the finance team can decide whether a narrow class of low-risk actions deserves more automation. That decision should follow evidence from the workflow, not confidence in a product demo.
The real asset is the decision design
Models will improve. Interfaces will change. Vendors will add more agents.
The durable asset is your definition of the work: which evidence matters, how the recommendation is formed, who owns exceptions, what success means, and which decisions remain human.
That is how domain experts become AI architects. They do not need to build a model. They need to make their judgment explicit enough that a digital coworker can prepare the work without pretending to own the outcome.
Start with one cash decision this week.
Let the agent gather and prepare. Let the finance leader decide. Then improve the workflow from the evidence.
Sources
[1] https://markets.ft.com/data/announce/detail?dockey=600-202610120800BIZWIRE_USPRX____20261012_BW059792-1 — FIS Equips CFOs With Agentic AI to Forecast Cash Flow and Surface Risk Earlier