Finance Already Owns the Scarce Skill in AI Agent Design

A purchase order records an approved amount of 10000. The vendor invoice for the same order prints $10,000. To a person these are the same number. To an automated comparison they are two different strings, and a correct invoice is rejected for reasons unconnected to the transaction.

Reverse the failure and it turns serious. An expense amount held as text in one system and as a number in another reads 54 and 54.00 as different values. A duplicate claim passes, nothing is flagged, and the workflow reports success, because from its own point of view nothing went wrong. Neither problem involves artificial intelligence, and both are settled before any model is chosen.

The timing is awkward. Finance leaders have committed to agents faster than they have built the means to see what those agents are doing, and the two most recent waves of Deloitte’s CFO Signals survey capture both halves of that gap in the same population.

Gartner, meanwhile, projects that more than 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. That is a forecast rather than an outcome, but the reasoning repays attention. None of the three causes is a limitation of the technology. All three follow from design decisions the implementing organization made, or skipped.

Why Agent Projects Fail Quietly

Most software either works or does not, and testing settles the question. An agent operating on financial data behaves differently. It handles the clean cases correctly, writes a satisfying row to a table, and reports success. The defects live in the cases the design never contemplated, which are rare by definition. That is precisely why they were missed.

A person doing the same work experiences ambiguity as friction. Being handed three purchase orders when you expected one is jarring, and the reviewer stops. An automated flow experiences nothing at all. It takes the first row that matches and proceeds with complete confidence, identically on the first invoice and the five hundredth.

The Decisions That Actually Matter

The Miles Masterclass Copilot Studio: AI Agents for CFO Teams, built by Uttam Pai, CPA, is instructive here because its three workflows, overdue receivables, three-way invoice matching, and expense claims, stay simple enough that the design decisions remain visible. Three of them carry most of the risk.

The first is identity. Each workflow depends on a field that identifies a record uniquely: the purchase order number appearing on invoice, order, and goods receipt alike; the invoice number for a receivables update; the email address identifying a claimant and therefore the correct approving manager. Key a payables lookup on vendor name instead, and a supplier with four open orders returns four matches. The process continues regardless, against the wrong record, and the error surfaces in a reconciliation weeks later, if at all.

The second is permitted values. The receipt extraction prompt constrains expense category to exactly four options and supplies the mapping from real documents onto them, because the following step uses that category to retrieve a spending cap from a policy table. A model that invents reasonable wording of its own breaks the lookup. Read that specification without the AI framing and it is a chart of accounts: a closed value set, an explicit mapping, and a downstream process depending on it.

The third is where authority sits.

Underneath all three sits the log. Every item processed writes a row carrying a specific reason rather than a generic failure flag. The difference looks cosmetic and is not. A reviewer working a queue of items marked Failed has to reopen each one before learning anything, while a reviewer working named exceptions can triage the queue at a glance and route most of it without opening a single document.

That distinction compounds, because exception data is also detection data. The ACFE’s Occupational Fraud 2026 report, covering 2,402 cases, records a median loss of $104,000 and a median twelve months before detection, with 43 percent of cases surfacing through tips rather than any control. That runway is long largely because monitoring is sampled rather than continuous.

One boundary holds the whole design together. Research by Bradford Levy in the Journal of Accounting Research in 2026 found that non-reasoning language models perform poorly on numerical reasoning tasks relevant to accounting and should be applied judiciously. In these workflows the model never computes anything. It reads named fields off a document and returns them in a fixed shape, while every comparison and threshold test runs deterministically in the workflow itself.

The Skill the Profession Already Has

What remains is ownership, the consideration least likely to be settled first. A team that builds three agents in a quarter has created three pieces of financial infrastructure, each needing an owner, a change process, and a reviewer for the exception queue it generates. A flagged invoice nobody examines is worse than no flag, because it produces the documentation of a control without the control.

None of this is artificial intelligence work. It is the work of documenting a process properly, which the profession has always known how to do and has often deferred because the manual version ran adequately on institutional memory. An agent will not. It executes the design exactly, at volume, at three in the morning, with no instinct that anything looks unusual.

Read the Gartner forecast again with that in mind. Escalating costs, unclear business value, inadequate risk controls. Those are specification failures wearing technical clothing, and they describe a gap CFOs, controllers, and auditors are better equipped to close than anyone else in the organization. The profession’s contribution to this technology is not adoption. It is precision.