What the finance team looks like in 2028

Finance teams are being asked to close faster with fewer people, and the numbers explain why. Too much finance time still goes to gathering data rather than using it, and the close still takes longer than anyone wants. The United States has 340,000 fewer accountants than it did in 2019, and private equity has put $50B into CPA firms in six years, with 11 of the top 30 US firms now backed by it, betting that the work can be done with fewer hands.

AI is the obvious answer, and most teams have tried it. Gartner finds that close to 60 percent of finance teams are piloting or implementing AI. Only 7 percent of CFOs say those investments are having a strong impact. That gap is the most useful number in finance technology right now, because it says the tools are not the problem.

Most finance pilots stall for the same reason: nobody designed the approval step. An agent that matches bank lines is easy to demo. An agent whose matches a controller will sign off on is a different thing. It needs the evidence behind every match, a queue for the ones it is unsure of, a record of who approved what, and a measured error rate the controller can trust.

Teams that skip this end up with a clever tool nobody relies on. The controller checks every line anyway, the hours do not move, and the pilot is quietly shelved. The few who report strong impact tend to have done the unglamorous part first: they decided who approves, what evidence they see, and what number would count as good enough.

There is a data reason too. Many pilots run on a clean sample chosen for the demo, then meet the real ledger and fall apart. Building on last month’s actual books, and testing on a month the agent never saw, is slower to start and much faster to trust.

The work agents take first is the work nobody on a finance team trained for. Matching bank lines to invoices. Chasing remittance details. Preparing recurring and accrual entries. Checking the close against the checklist. Assembling the monthly pack from five systems. Drafting the first version of variance commentary with the number attached.

This is most of the data-gathering time. It is repetitive, it is rule-shaped with many exceptions, and it is exactly where a person’s time is worth least. An agent does it overnight, flags what it could not resolve, and leaves the team a queue of decisions instead of a pile of preparation.

The pattern repeats across every one of these tasks: the agent does the reading and the first pass, and a person handles what the agent could not resolve or should not decide.

People keep judgement, sign-off and relationships. The controller decides whether a variance is a timing difference or a problem. The CFO decides what the board hears. The team decides when a customer’s payment pattern means a conversation rather than a reminder. None of that is preparation, and none of it should be handed to an agent.

The controller’s job changes shape. Less time building the close, more time approving it. Less time finding exceptions, more time resolving the ones that matter. A new responsibility appears too: owning the agents themselves, reading their evals each month and deciding when a rule should change.

Some predict fully autonomous finance by the end of the decade. That misreads how finance works. Every number in a finance function has an owner who answers for it to auditors, lenders, boards and tax authorities. An agent cannot hold that responsibility, so a person must approve what it produces.

The cost of errors is also lopsided. A missed match costs an hour of review. A wrong entry posted without review can cost a restatement. Human-in-the-loop is not a transitional compromise; it is the deployment pattern that lets agents do real work in a function where mistakes are expensive and accountability is personal.

The skills a finance team hires for will shift with the work. Reconciliation speed matters less when an agent matches overnight. Judgement matters more: knowing which exception hides a real problem, which variance needs a board note, and when a rule the agent follows has stopped fitting the business.

Two newer skills appear. One is reading an eval: understanding what a match rate on a held-out month says, and what it does not. The other is writing a good rule: turning "we usually treat these as prepaid" into an instruction precise enough for an agent to follow and a reviewer to check. Neither needs a data science degree. Both need someone who knows the books.

Put together, the finance team of 2028 looks different in shape rather than in title. A smaller preparation layer, much of it automated. A stronger review layer, with controllers and analysts approving agent work against evidence. One or two people who own the agents, the evals and the rules. A close measured in days rather than a week, because the preparation happens overnight.

Getting there does not require a transformation program. It requires one workflow done properly: the one with the most hours and the clearest approval step, built on your data, measured on a month the agent never saw, and run with the controller in charge. The second workflow is easier, because the approval habit already exists.

For firms that deliver finance work to clients, the shape changes the same way at larger scale. An accounting or fractional-CFO firm with agents doing preparation across its client book can serve more clients per reviewer, and the reviewers become the product. Private equity’s bet on CPA firms is, in part, a bet on exactly this shift.