Key takeaways
- Busy season is a capacity problem. The gains come from the work surrounding the return, not from the return itself.
- Tax research is the worst starting point. Confidently wrong code references cost more to catch than they save.
- Client financial data means the Safeguards Rule applies, and a personal account is not a defensible place for it.
- Set it up in the off season. Nobody adopts a new tool in March.
The compression problem in an accounting firm is not that the work is difficult. It is that fourteen weeks of the year contain a disproportionate share of the annual workload, the people who can do it are finite, and the constraint shows up as overtime rather than as unfinished work. Anything that moves hours out of that window has outsized value, and anything that adds friction during it will be rejected on contact.
That shapes what a firm should do with AI, and more importantly when.
The gains are around the return, not inside it
Partners tend to look for AI in preparation and review, because that is where the hours are. It is the wrong place to look first. Preparation is already handled by tax software that encodes the rules properly, review requires professional judgement, and both carry consequences that make an unverified suggestion expensive.
The recoverable time sits in the surrounding layer, which is larger than most firms realise until they measure it. Chasing missing documents. Explaining to a client why their number moved. Reading through the disorganised pile a client sends and working out what is actually in it. Writing the letter. Answering the same internal question for the fourth time that week. None of this is billable in a way anyone enjoys, and all of it consumes senior time during exactly the weeks when senior time is scarcest.
| Task | Fit | Why |
|---|---|---|
| Summarising the document bundle a client sends | Strong | Everything is in front of you, so checking it takes minutes |
| Drafting the missing information request | Strong | Repetitive format, reviewed before sending, sent constantly |
| Explaining a variance to a client in plain English | Strong | You supply the numbers, it supplies the wording |
| First drafts of engagement and management letters | Strong | Templated work with firm-specific adjustments |
| Turning workpaper notes into a review memo | Good | All substance comes from the preparer |
| Answering internal how-do-we-do-this questions | Good | Reduces interruptions to seniors if fed your own procedures |
| Drafting responses to notices for review | Moderate | Useful skeleton, but every citation and figure needs checking |
| Tax research and authority | Poor | Fabricated or superseded references are formatted convincingly |
| Preparing or signing the return | Unsuitable | Your software already encodes the rules and you sign it |
| Audit conclusions and opinions | Unsuitable | Professional judgement with a documentation standard attached |
The strong rows share one property. The firm supplies the facts and the tool supplies the structure and the wording. When that direction reverses, so that the tool supplies the facts, the verification cost rises faster than the time saved.
A useful test before adopting any AI task in a tax practice: if the output were wrong in a way that looked right, how long would it take someone to notice? Under a minute means it is a good candidate. After the client has filed means it is not.
Why tax research is the trap
It is the most tempting use and the one that reliably disappoints. A general purpose model will produce a confident answer citing a code section, a revenue ruling and a case, formatted exactly as your research platform would format it, and some proportion of those references will be superseded, misapplied or invented. There is no visual signal distinguishing the sound ones from the rest.
The specific failure that catches firms out is not the obviously wrong answer, which a senior spots immediately. It is the answer that is broadly correct in shape but cites the wrong authority, or the answer that was correct two tax years ago. Both survive a quick read.
The workable version is to invert the order. Do the research in your existing platform, reach a position, then use AI to explain that position to a client, to draft the memo documenting it, or to pressure test the reasoning by asking what a reviewer might challenge. All of that adds value with no fabrication risk, because the authority came from you.
Client data and the Safeguards Rule
A tax preparation firm is a financial institution for the purposes of the Gramm-Leach-Bliley Act, which brings it within the FTC Safeguards Rule and the requirement to maintain a written information security plan covering how client information is handled and which service providers touch it. The IRS has reinforced the same expectations for preparers through its own security guidance.
Applied to AI, that produces a straightforward test rather than a philosophical debate. If a preparer pastes a client's 1099s into a personal account, the firm has a service provider it did not select, cannot audit, cannot configure and has not documented. That is a plan failure independent of anything the vendor does or does not do with the data.
The remedy is not to prohibit the tool, which produces avoidance rather than compliance. It is to put the tool inside the perimeter:
- One firm account on a business tier with a commercial agreement the firm is party to, so the provider is a documented service provider rather than an invisible one.
- Training on firm inputs disabled and the setting recorded, with a screenshot in the same folder as the rest of your security plan evidence.
- Retention set deliberately rather than left at the default, and consistent with how you retain other client correspondence.
- Access managed by an administrator, so a seasonal preparer's access ends when the season does.
- The provider added to your written plan as a service provider, with the same one-paragraph treatment your document portal gets.
Firms that do this can answer a peer review question or a client due diligence questionnaire in a sentence. Firms that have not tend to discover the gap at the least convenient moment.
The seasonal staff detail worth planning for: if you bring on preparers for the season, their AI access needs to start and end with their engagement, on an administrator-controlled account. Firms running on individual sign-ups routinely forget this and leave access open for months after someone has gone.
Set it up between May and November
This is the single most practical piece of advice in this article. A tool introduced in February will be ignored, because nobody learns anything new during compression, and a tool that fails during busy season will be permanently associated with having made a hard period harder.
The off season sequence that works:
- May to June. Debrief the season that just finished while the frustration is fresh. Ask specifically where time went that was not preparation or review. The list of answers is your task list.
- July to August. Configure the account properly, set the spending ceiling, sort out access, and update the security plan. This is the unglamorous part and it takes days, not weeks.
- September to October. Run it on extension work and on the advisory and bookkeeping side, where the pressure is lower and mistakes are cheap. This is where you learn which tasks actually help.
- November to December. Write the task guidance for each role based on what you learned, train people on that rather than on the tool in general, and lock it before the year end rush.
By January the tool is not new, the habits exist, and nobody is learning anything during the weeks they cannot afford to.
Role-specific guidance beats general training
The single biggest determinant of whether an accounting firm gets value from this is whether people know what to use it for on a specific Tuesday. A general demonstration produces a room of people who are impressed and then do nothing.
What works is a short list per role. For an administrator, the client chasing correspondence and the document bundle summaries. For a preparer, the plain-English variance explanations and the workpaper notes cleanup. For a manager, the review memos and the internal procedure questions. For a partner, the client communication drafting and the proposal work.
Three to five specific tasks each, written down, with an example of what good looks like. That produces adoption. Feature tours do not.
What the return actually looks like
Be careful with the numbers you promise internally, because the honest picture is good enough without inflation. Most firms find that the tasks above recover something in the order of a few hours per person per week in the surrounding work, concentrated in administrative and client communication time rather than in chargeable preparation.
During compression, hours moved out of the week matter more than the same hours in September, so the value is not evenly distributed across the year. A firm that measures the benefit in July will understate it substantially.
What it does not do is reduce the number of returns a preparer can complete, or replace review, or change what the firm sells. Partners who expect that will be disappointed by something that is in fact working.
Frequently asked questions
Can we use AI with client tax data?
On a firm-controlled business account, yes, with training on inputs disabled, retention configured, administrator-managed access, and the provider recorded in your written information security plan as a service provider. On personal consumer accounts, no. The Safeguards Rule obligations apply to how client information is handled regardless of which tool it went into.
Should we use AI for tax research?
Not as a source of authority. General purpose models produce convincingly formatted references to code sections and rulings that are superseded or invented, and the plausible ones survive a quick read. Research in your existing platform, then use AI to explain the position, draft the memo or stress test the reasoning.
When should we roll this out?
Between May and November. Configure in summer, practise on extension and advisory work in autumn, write role-specific guidance and train in late autumn. Anything introduced during compression will be ignored, and anything that goes wrong during compression will be blamed permanently.
How much time does it actually save?
Most firms recover a few hours per person per week, concentrated in client correspondence, document review and internal knowledge questions rather than in preparation. The value is weighted heavily toward busy season, so measuring it in a quiet month understates it.
Have it configured before the next season starts
We choose the provider, disable training on your inputs, set retention and a hard spending ceiling, manage access so seasonal staff are handled properly, and write task guidance separately for administrators, preparers, managers and partners. Fixed price, live in 30 days or less.