Tax workpaper preparation is one of the better automation candidates in professional services, and one of the easiest to get dangerously wrong. The work is repetitive, high volume and seasonal, which argues for automating it. It is also evidence, which constrains how.
Separate Preparation From Judgement
The single most useful distinction is between assembling a workpaper and deciding what it means.
Assembly is document intake, classification, extraction, tie-out to the trial balance, rolling forward last year's schedules, chasing missing items, and formatting. It is most of the hours and almost none of the risk.
Judgement is the treatment of an item, the position taken, and the conclusion signed. It is a fraction of the hours and effectively all of the risk.
Automation belongs firmly in the first. Firms that get into trouble are the ones that let a confident-sounding system drift across the line, usually because it produced a conclusion and nobody noticed it was a conclusion. The line between rule-following and judgement is the same one we set out in business process automation versus RPA.
What Genuinely Automates Well
- Intake and classification. Sorting a client's document dump into categories, naming and filing consistently. Enormous time saving, low risk, easily checked.
- Extraction from standard forms. Structured documents with predictable layouts extract reliably, and mismatches are detectable against the trial balance.
- Rollforward. Last year's schedule with this year's figures is mechanical, and the diff against prior year is the natural review artefact.
- Tie-outs and completeness checks. Machines are better than tired humans at noticing that a number appears twice with different values in March.
- Chasing the client. Knowing what is missing and asking for it, repeatedly and politely, is pure process.
Where It Must Not Go
Anything producing a position, a characterisation or a conclusion needs a named human who owns it. Not a reviewer who approves an automated conclusion, which in practice becomes rubber-stamping under deadline. A human who forms the view.
The distinction sounds pedantic until a filing is questioned and somebody has to explain how a treatment was arrived at. "The system suggested it and we accepted it" is not an answer that helps anybody.
Automate the assembly of evidence, never the interpretation of it. The moment a workflow starts producing conclusions, somebody has to own each one by name.
Lena Fischer, Solutions Architect, Engineered With AI
The Audit Trail Is The Product
For most automation, the output is the point. Here the record of how the output was produced matters as much as the output.
Every automated step should record what it received, what it did, what confidence it had, and what a human changed afterwards. Without that, you have a faster process you cannot defend, which is worse than the slow one you could.
- Source document reference on every extracted figure, down to the page.
- A visible confidence signal, with anything below threshold routed to a human rather than passed through quietly.
- A complete record of human overrides. These are also your best training signal for what the system keeps getting wrong.
- Versioning, so the workpaper as filed can be reconstructed exactly.
- Named review, not just a timestamp.
The Seasonality Argument
The economics here are unusual. Capacity is needed for a few intense months and idle for the rest, so firms either overstaff all year or burn people out in the peak.
Automating assembly changes that shape more than it changes headcount. The same team handles a heavier season without the quality decline that comes with fatigue, and the review time gets spent on the things that actually need judgement rather than on formatting.
That is the honest benefit. Firms expecting to halve their preparation team are usually disappointed; firms expecting to stop losing good people every April usually are not.
How To Start Without Risking A Season
Run it in parallel for one cycle. Automate intake, classification and rollforward only, keep every existing review step, and compare. You will find out quickly where extraction is unreliable and which document types break it, at zero risk to a filing.
Then extend one step at a time, and only where the check is cheap. Anything whose correctness you cannot check quickly should stay manual regardless of how well it appears to work.
If you describe your document intake, your review steps and where the season actually hurts, we can usually tell you in one conversation which parts are worth automating and which should not be touched.
Wondering what is safe to automate?
Describe your intake, your review steps, and where the season hurts most. We will tell you what automates cleanly and what should stay with a human.