What AI Can Do for Your Audit Prep, and What Your Auditors Will Reject

1,120 of the 2,706 active controller postings we track put the external, annual or year-end audit directly in the job description. 89 of those mention AI. That is 7.9%, which is somehow lower than the AI mention rate across our whole database of 54,357 live postings, and it means the single most predictable, most repetitive, most soul-flattening block of work in corporate accounting is the one part of the job nobody is automating.

I want to be careful here, because some of that hesitation is completely earned. Audit prep produces evidence, and evidence that a model hallucinated is worse than no evidence at all - it is an actual finding waiting to happen. But a lot of the hesitation is just habit, the same way we all kept doing manual tie-outs for a decade after Excel could have done them, and the gap between the work that genuinely needs judgment and the work that is pure fetch-and-format is enormous.

How much of the accounting job is actually audit prep?

Roughly a fifth of everything. 10,786 of our 54,357 active postings mention external audit, annual audit, year-end audit, audit preparation or audit requests, which is 19.8%. 2,383 mention workpapers by name. Only 272 say PBC or prepared by client out loud, which tells you the term is jargon the industry uses internally and strips out of job ads.

The concentration by role is where it gets interesting. 47% of senior accountant postings mention the annual audit. 43% of accounting manager postings. 41% of controller postings. 29% of staff accountant postings. So this is not a niche specialty you can opt out of, it is a defining feature of the mid-career accounting job, and if you are a senior accountant reading this there is a coin-flip chance your next role hands you the audit binder.

Now the AI numbers against those same populations. Senior accountant: 103 of 1,209. Controller: 89 of 1,120. Accounting manager: 69 of 732. Staff accountant: 64 of 1,237. Director of accounting: 39 of 235. Across the whole audit-prep population it is 998 of 10,786, or 9.3%, which is barely above the 8.0% baseline for every posting in the database. Employers are hiring people to survive the audit and saying nothing about the tools.

Where does AI actually earn its keep in audit prep?

In the fetch-and-format layer, which is most of the hours. The auditors send a request list, and somebody on your team spends three weeks turning general ledger detail into schedules that answer questions the auditors already know the answers to. That is the part to automate.

Concretely, the wins I would go after first:

Translating the request list into a work plan. Drop the PBC list in, have the model map every item to the system it lives in, the person who owns it, the report that produces it and last year's version. You are not asking it to do accounting, you are asking it to do project management on a hundred-line list, and it is very good at that.

Drafting the flux and variance explanations. Auditors ask why the number moved. The model can read the GL detail, cluster the drivers and draft the first version of every explanation, and then you rewrite the ones that are wrong. We wrote a full playbook on automating flux and variance analysis if you want the mechanics.

Tie-out and cross-referencing. Checking that the number in the schedule agrees to the trial balance, that the trial balance agrees to the financials, that the footnote agrees to both. Deterministic work with a clear right answer, which is exactly where you want a machine.

Reading the contracts the auditors are going to ask about. Leases, revenue contracts, debt agreements, equity docs. Extraction and summarization into a schedule, with a page cite for every field so a human can verify it in ten seconds instead of forty minutes.

Last year's findings. Feed in the prior year management letter and the current close file, and ask what is likely to come up again. Cheap, fast, and the answer is usually uncomfortable in a useful way.

What will your auditors reject?

Anything you cannot trace back to the system that produced it. This is the part people skip and then get burned on.

The AICPA's SAS No. 142, Audit Evidence is explicit that auditors evaluate information regardless of the source it came from or the procedures used to obtain it, and that when they assess information produced by the entity they look at how automated the process was and how strong the controls around it are. That standard cuts both ways for you. A well-controlled automated process is easier to rely on than a spreadsheet somebody built by hand. An uncontrolled model output that nobody can reproduce is worse than either.

So the hard lines I would draw:

Never let a model be the source of a number. It can move numbers, reconcile numbers, explain numbers and format numbers. It does not originate them. Every figure in a schedule ties to a system report you can regenerate.

Keep the prompt and the output. If the auditors ask how a schedule was built and your answer is that someone asked an AI, you have a control problem, not a technology problem. Save the inputs, the prompt, the output and the human review step, in the workpaper, like any other evidence.

Someone signs it. A named human reviews and approves before it leaves the building. Not a rubber stamp - an actual review where they check the tie-outs. If nobody has time to review the output, you did not save any time, you just moved the risk downstream.

Do not put unreleased financials into a tool you have not diligenced. Obvious, and people do it anyway, usually the week before fieldwork when everybody is tired and someone pastes the trial balance into a consumer chatbot. That is the scenario that ends careers.

If your company is public, the control implications get heavier fast, and we went deep on that in letting AI into the close without blowing up your SOX audit.

What about the PBC list itself?

Push back on it earlier than you think you are allowed to. Half the reason audit prep is hell is that the request list arrives as a hundred undifferentiated items with no sense of what matters, and your team burns equal effort on the schedule that supports a material estimate and the one that supports a rounding difference.

Take the list, have a model group it by account, by materiality and by which items are just last year's schedule refreshed, then go to the audit senior with a proposed sequence and a list of items you think are stale carryovers from three years ago. Most of them will agree to drop a few. Nobody ever asks, so nobody ever finds out that asking works.

What I would do

Pick the flux explanations and do only that this quarter. It is the highest-volume, lowest-risk, most obviously improvable piece of the whole exercise, and you can run it in parallel with your normal process so nothing depends on it working. Build the prompt, run it against last quarter, compare it to what your team actually wrote, and see how close it gets. If it gets 70% of the way there, that is real hours back.

Then do the tie-outs, then the contract extraction, and leave the judgment calls alone entirely. The estimate is yours. The reserve is yours. The going concern conversation is yours. That work is why the profession exists and it is worth more, not less, when the mechanical stuff around it collapses.

And a broader thing, because it keeps proving true: the accountants who are getting real value out of this right now are the ones who built something small and imperfect and shipped it into a live close, not the ones who spent nine months evaluating vendors. Build the damn thing. If it is bad you will know inside a week. If you want the software landscape while you are deciding what to buy versus build, the Audit Friendly software directory is there.

Frequently asked questions

Can AI prepare audit workpapers?

It can draft them, and it should not originate the numbers in them. AI is reliable for reformatting system-generated data into schedules, tying figures across documents and drafting variance explanations. Every number still needs to trace back to a report you can regenerate from the accounting system, and a named human needs to review and approve before it goes to the auditors.

Will auditors accept AI-generated schedules?

They will accept schedules whose underlying data and process they can evaluate. Under AICPA SAS No. 142, auditors assess audit evidence regardless of the source and consider how automated the process was and how strong the controls are. A documented, repeatable automated process is often easier to rely on than a manual spreadsheet. An untraceable model output is not.

What is a PBC list?

PBC stands for prepared by client. It is the list of schedules, reconciliations, documents and support your auditors request from your team ahead of and during fieldwork. Only 272 of the 54,357 accounting postings we track use the term explicitly, though nearly 20% describe the underlying work.

How many accounting jobs involve audit preparation?

About one in five. 10,786 of the 54,357 active accounting and finance postings in our database reference external, annual or year-end audit work. It peaks at senior accountant, where 47% of postings mention it, and stays above 40% for accounting managers and controllers.

Is it safe to put financial data into an AI tool?

It depends entirely on the tool and your agreement with the vendor. Enterprise deployments with no-training commitments and proper data handling are a different thing from a consumer chatbot. Unreleased financial information should never go into a tool your company has not formally reviewed, and that review should happen well before the week fieldwork starts.

The interesting part of all this is what happens on the other side. If audit prep stops eating three weeks of every year, that time does not disappear, it goes somewhere - into analysis, into the systems work everyone has been deferring since 2019, into actually thinking about the business. Go take one piece of the request list and automate it before next year-end, and see what you do with the hours.