Blog /Accounting

Why Accounting Automation Stalls (and the Tasks AI Should Not Touch)

Anirudh Badam

Anirudh Badam

Co-Founder and CAIO, Adopt AI

20 November 2026

The pilot went well. Someone demoed a reconciliation that finished in minutes, the controller nodded, and the vendor sent a contract. Six months later the team still closes the books the old way, and the tool runs on one account nobody looks at.

That pattern shows up in the survey data. Gartner found 59% of finance leaders using AI in 2025, yet only 7% report a high or very high impact from their AI investment. Most teams adopted something. Few can point to hours they got back. (We collected the numbers in AI in Accounting: 58 Statistics for 2026.)

Automation stalls for a short list of reasons, and most of them get decided before anyone signs. Below are the seven we see most, the mistakes buyers make in their first evaluation, and the work AI should leave with a person.


The short answer

Accounting automation stalls when teams automate low-value work first, skip the baseline, wait for perfect data, ignore how errors compound across steps, underestimate review time, attempt a migration before seeing value, or leave the exceptions with no owner.

AI should prepare the work. A named person should keep sign-off on financial statements, judgment calls, payment release, control accounts and anything you might need to defend to an auditor.


Seven reasons accounting automation stalls

1. The team automated the cheap work first

Gartner's late-2025 data on AI use cases in production put knowledge management on top at 49%, ahead of accounts payable automation (37%) and error and anomaly detection (34%). Knowledge tools are easy to deploy and pleasant to use. They also save minutes, spread across many people, which no one can add up.

Reconciliations, AP intake and flux commentary cost hours, and those hours land on a payroll line. If you want a number the CFO believes, point the first project at work with a visible cost. Our post on the finance AI adoption and ROI gap goes further into this.

2. Nobody timed the process before the tool arrived

You can't prove you saved twelve hours on a reconciliation if nobody recorded the twelve hours. Teams skip the baseline because it feels like busywork, then spend the renewal meeting arguing about anecdotes.

Before a pilot starts, time three things for one workflow across two closes: preparation, review and rework. Rework is the one people forget, and it is the one automation most often makes worse.

3. The team waited for clean data

"Our data isn't ready for AI" came up again and again on calls with finance teams this year. Sometimes it is true: a chart of accounts with three accounts for the same expense, or vendor names spelled four ways.

Waiting for clean data stalls the project too, because the cleanup never finishes. A better test: does the tool tell you what it could not match, and why? A system that reports its unmatched items gives you a cleanup list. A system that forces a match gives you errors you find in March.

4. Errors compound across a long chain

A month-end close runs a dozen or more steps across several systems. Per-step accuracy multiplies. An agent that gets each step right 95% of the time finishes a 10-step workflow without an error about 60% of the time (0.95 to the tenth power). Stretch it to 20 steps and you get about 36%.

That math explains why a demo of one step looks flawless and a full close falls over. Insist on checkpoints between steps, where a check runs or a person reviews before the next step consumes the output.

5. Review time eats the savings

If a person has to re-check every line the tool produced, the work moved to the reviewer's desk. One firm lead we worked with set the bar we now use: net hours saved after model review and error corrections.

Ask vendors how their output gets reviewed. Good answers sound like: the preparer and reviewer are named on every item, exceptions arrive in a queue with the evidence attached, and clean items can be approved in bulk. Weak answers sound like a percentage.

6. The project started with a migration

Some tools need you to move your ledger, rebuild your chart of accounts or replace your close checklist before they do anything. That turns a three-week pilot into a two-quarter implementation, and the budget runs out before value shows up.

The same risk sits behind many "mistakes deploying cloud finance" stories: the team bought a platform, then spent a year making the business fit it. Prefer tools that work inside the systems you already run and write approved results back to them.

7. The exceptions have no owner

IT ran the pilot. The controller attended the kickoff. Then the first batch of exceptions arrived and nobody's job description included clearing them.

Name an owner for exceptions before go-live, and put the reviewer's name on each item. Write the policy down too. A Financial Cents survey this summer found 87% of accounting firms already using AI had no formal written AI policy.


Mistakes buyers make the first time they evaluate AI accounting tools

Judging the tool on the vendor's data. Vendors pick clean samples for demos. Send your own messy month: the duplicate invoice, the partial payment, the vendor who changed bank details.

Accepting a bare accuracy percentage. "97% touchless" means nothing without the denominator. Even in Ardent Partners' AP benchmarks, straight-through processing is the number best-in-class teams are still working to raise, not a solved problem. Vendor figures in the 90s usually count only clean, PO-backed invoices. Ask: what share of our volume has no PO, and what happens to that share? We wrote more about this in The 97% Model.

Skipping the "what happens when it doesn't know" question. The most useful thing an agent can do on a hard item is stop and route it to a person with the evidence. Ask to see the queue of items the tool declined to decide.

Ignoring what you keep if you leave. Ask which parts are yours: the rules, the agent instructions, the decision history. Teams that end up with a dozen AI tools eventually consolidate, and the work they built inside a closed tool goes with the contract.

Buying a point tool per process. A close tool that can't see AP, an AP tool that can't see AR. Each one adds a login and an audit trail you have to stitch together at year-end.


The accounting tasks AI should not touch

AI does well on preparation: pulling documents, matching transactions, drafting entries, drafting variance commentary, chasing missing items. The list below covers work where a person should keep ownership, even when AI prepares the inputs.

TaskWhy it stays with a person
Signing off financial statementsSomeone has to be accountable for the numbers, by law and by job description
Posting to control accountsErrors here distort the subledger tie-out and are hard to trace after the fact
Materiality and estimate judgmentsReserves, impairments, revenue recognition calls and materiality thresholds rest on business context an agent doesn't have
Releasing paymentsMoney that leaves the account rarely comes back, and vendor bank-detail fraud targets this step
Changing vendor bank detailsSame fraud exposure. Require a call-back to a known number
Taking tax positions and signing returnsThe preparer signs, and the penalties attach to the preparer
Reopening a closed periodChanges after close affect reported numbers and audit work already done
Delivering bad news to a client or the boardTrust is built in that conversation, and an email from a tool spends it

A useful test for anything not on this list: if an auditor asked "who decided this, and why?", could you name the person and show the evidence? If the honest answer is "the tool did it," keep a person in the loop on that task.


How to keep an automation rollout moving

  1. Pick one high-volume workflow with a visible cost. A bank or balance sheet reconciliation is a good start. Our month-end close guide lists the usual candidates.
  2. Time it for two closes before you start. Track preparation, review and rework as three separate numbers.
  3. Run the tool on your real data, in parallel with your current process. Compare outputs line by line for at least one close.
  4. Gate every posting at the start. A person approves each entry until the tool earns trust on that item type.
  5. Relax gates by item type, with evidence. Teams often loosen review on routine items over the first one to three months and keep full review on judgment calls.
  6. Measure net hours. Preparation saved minus review added minus rework created.
  7. Keep the audit trail readable. Every item should show who prepared it, who approved it and what evidence they looked at.

How Adopt approaches this

Adopt's AI agents work as preparers inside the systems you already use. They reconcile, match, draft entries and draft commentary, and a named person on your team reviews and approves every item before it counts. We never post to a control account. When an agent can't support a match, it says so and routes the item to a person with the evidence attached, instead of guessing.

Gates start tight and loosen by item type as your team sees the results. The rules and agent instructions your team builds stay yours.

Start free on your own data, or book a demo and bring your messiest reconciliation.

FAQs