Tax Season 2026 Postmortem: Where AI Actually Helped Small Accounting Firms (and Where It Quietly Added Work)

Tax season 2026 is the first one most small CPA firms ran with AI in the building. Not in a lab, not in a sandbox, not on a partner’s laptop — in the actual return-prep workflow, in front of actual clients, on actual deadlines. Three weeks after the April 15 push, the partner conversations sound very different from last year. Some firms saved real hours. Some firms quietly absorbed new work nobody wrote on the dashboard. Both stories are true, and the gap between them is the most useful piece of operating data the profession has ever had about AI.

We have spent the last two weeks debriefing partners and seniors at small accounting firms — mostly 5-to-50-person shops, a handful of larger regionals — about what AI did and did not do this season. The headlines you will read in industry press are predictably bullish. The numbers in the actual return-prep WIP reports are more honest. This is the postmortem nobody is publishing yet.

The shape of the season

Compared to 2025, three things changed in 2026 worth naming up front. First, every major engagement-system vendor shipped some flavor of “AI assistant” in the workflow, mostly as a copilot side panel. Second, a small number of firms — the ones we work with most closely — ran custom intake and reconciliation pipelines that were not visible to staff as AI at all. Third, the cheap-frontier-token environment finally broke open: by January, a senior could ask a model to read a long set of K-1s without flinching at the cost.

That is the macro. The micro is where the season’s lessons actually live. We have grouped them into two columns — what got better, and what got quietly worse — because the most common failure mode this season was firms looking only at column A and concluding the season was a win.

What actually got better

These are the workflows where partners reported a clear, measurable improvement and where the WIP report agreed with them. The pattern is consistent: bounded inputs, structured outputs, schema-driven validation, and a citation trail back to the source.

• K-1 aggregation across entities. The single biggest unlock for firms with high-net-worth and family-office clients. Pulling every K-1 a client received, tying each one to the right entity in the firm’s database, and producing a reconciliation against the client’s capital accounts went from a multi-day senior staff exercise to a one-pass review of model output. Firms with custom pipelines reported the largest gains; firms relying on copilots inside their engagement system saw modest improvements but still spent most of the time clicking around the UI.
• 1099 reconciliation against bank summaries. The model reads the 1099-NECs, 1099-INTs, 1099-DIVs, and the client’s bank year-end summary in one pass, flags mismatches, and produces a short list of figures the staffer actually has to chase. Several firms reported cutting the 1099 reconciliation step from a half day to under an hour per complex individual return.
• State-form classification and routing. Multi- state returns are the workflow most firms hate touching. A properly tuned classifier that sorts incoming documents into the right state buckets and pre-populates which state-specific schedules are required removed roughly two hours per multi-state return at firms that built it well. The firms that bought it as a feature reported less consistent gains.
• Document chasing automation. Pure ops-automation, not really AI, but the firms that wrapped a model around their intake checklist sent dramatically fewer “please send the 1098-T” emails. The model knew the engagement, knew what the client had already sent, knew what was missing, and drafted the follow-up in the partner’s voice. Several firms reported that documents arrived a week earlier on average than in 2025.
• First-pass workpaper drafting. The classic repeatable senior-staff exercise: take the client’s documents, drop them into the firm’s workpaper template, fill in the figures with citations. Firms with structured intake pipelines had the cleanest gains here, because the input the workpaper drafter saw was already typed and validated.
• Engagement letter and 7216 consent generation. Boring, repeatable, schema-friendly. Several firms moved the entire engagement-letter process to a templated, AI-assisted generator and reclaimed a meaningful chunk of administrator time before the season even started.

Across all of these, the firms that built or commissioned the workflow themselves reported larger and more durable gains than the firms that turned on a vendor copilot. The reason is the same one we wrote about in the case for building at small accounting firms: a copilot pasted onto a generic engagement system can only improve the workflow as much as the workflow already fits the vendor’s assumptions. The custom path lets the workflow bend to the firm.

What quietly got worse (or stayed the same)

This is the column nobody publishes. It is also the column that decides whether tax season 2027 is actually better than 2026.

• Review time grew. Partners did not trust AI output, so they re-checked everything — often more carefully than they would have re-checked a senior’s work. On returns where the AI did half the prep, the partner touch time often went up, not down. The hours saved at the prep stage moved upstream to the review stage. The total hours moved a little; the cost of those hours, with partners billing at $400-$700, moved more.
• Edge cases burned more time, not less. Multistate returns with K-1 amendments, late-arriving 1099 corrections, IRS notices that arrived mid-prep, foreign accounts — the cases where the model was less reliable were the cases where staff had to undo what the model did before they could redo it correctly. Several firms reported that complex individual returns took longer than 2025 because the AI-prepped first draft had subtle errors that were harder to find than an empty workpaper.
• Junior staff lost reps on the work AI now does. The first-year associates who would have spent 200 hours on 1099 reconciliation and K-1 aggregation are not getting that training cycle. A handful of partners we spoke to are worried, correctly, about what their senior bench looks like in five years if the AI keeps owning the entry-level reps. This is a real workforce question that 2026 surfaced and most firms have not yet answered.
• “Helpful” hallucinations buried in long memos. The pattern that scared partners most this season: an AI-generated client memo or workpaper note that read fluently, included real-looking numbers, and contained one figure that did not actually appear in any source document. Firms without a strict citation trail caught these in review. Firms that trusted the model’s prose did not. The catch rate was much higher at firms that ran an eval suite over the prep pipeline.
• The vendor-copilot tax. Several firms paid per-seat fees for AI features that staff turned off in week two because the suggestions were noisy. The line item stayed in the budget; the value did not. Without telemetry tied to actual workflow outcomes, the only signal the firm had was the vendor renewal email.
• Privacy slippage. The most uncomfortable finding. A non-trivial number of staffers, under deadline pressure, pasted client documents or extracted figures into consumer AI chat tools to “just get an answer.” Firms with a clear AI-use policy and an internal model to point staff at had much less of this. Firms without one discovered the issue during their post-season debrief.

The hours saved at the prep stage moved upstream to the review stage. If the firm did not measure that move, it did not happen.

The pattern that separates the two columns

Step back from the workflows and the line that separates a season-saver from a season-stretcher is consistent across every firm we debriefed.

AI helped where the work was bounded, the inputs were known, the output had a schema, and every figure carried a citation back to the source document, page, and bounding box. AI hurt — or was net-flat — where the work required cross-context judgment, the inputs were ambiguous, the output was prose, and the model was allowed to summarize without proving where each number came from.

That is the same operating principle we wrote about in the AI client intake pipeline piece earlier this year. Tax season 2026 confirmed it under load. Schema and citations are not aesthetic choices for AI workflows in accounting. They are the difference between an audit-defensible pipeline and an expensive way to generate review work.

The metrics that mattered (and the ones that did not)

Most firms entered the season measuring the wrong things. The question “how many returns went through the AI tool” is a vanity metric — it tells you about adoption, not about value. The firms that ended the season with a defensible story about AI ROI measured a smaller, harder set of numbers.

• Cycle time per return type. Not in aggregate. By return type. A 5% improvement in 1040s and a 25% regression in complex partnership returns averages out to a happy 4% improvement that the partner group does not actually feel.
• Reviewer touches per return. The number of times a senior or partner touched a return after first prep. This is the metric that catches the “review time grew” problem before it eats the season’s margin.
• Throughput per senior staff hour. Returns completed per senior hour, separated by complexity. The throughput-per-junior-hour number went up at most firms; the throughput-per-senior-hour number is the one that decides whether the firm grew capacity or just shifted it.
• Citation completeness rate. Percentage of extracted figures with a working citation back to the source. This is the only number that proves the pipeline is audit-defensible, and it is the one firms most often forgot to instrument.
• Partner sign-off speed. Time from “return ready for review” to “return signed.” If AI prep is working, this number falls. If it is not, this is where the dam shows up.
• WIP write-off rate by matter type. Old metric, still the truth. If write-offs went up on AI-prepped returns, the firm did not save anything — it shifted hours into time the client never paid for.

We have written about this measurement discipline in more depth in our six-metric framework for measuring AI ROI. Tax season is the highest-stakes natural experiment a CPA firm runs all year; the firms that treat it as a measurement opportunity, not just a delivery sprint, will be a year ahead of their peers by next April.

What firms should do before tax season 2027

We have one strong opinion and several practical ones. The strong opinion: the firms that wait for vendors to ship the right AI features in time for next season will lose the year to the firms that built one or two custom pipelines this summer. Tax season 2026 was the dry run. Tax season 2027 is the season clients start asking specifically about AI in the engagement conversation.

The practical moves, in rough priority order:

• Run a real postmortem in May, not in November. Pick five clients across the complexity range, pull the time tickets, and walk the partner and senior who touched each return through the AI-touched and non-AI-touched steps. Write down where the time actually went. Do this while the season is still fresh.
• Pick one bottleneck to harden by Q3. Not five. Not the whole platform. One. Most firms we work with should spend Q3 building or commissioning a serious K-1 aggregation pipeline, a serious 1099 reconciliation pipeline, or a serious intake pipeline — not all three. Pick the one that produced the largest review-time spike during 2026, fix it correctly, then come back for the next.
• Write the AI-use policy you wish you had in January. Naming what tools staff are allowed to paste client data into, naming what tools they are not, and providing an internal alternative for the cases where staff legitimately needed AI assistance and went to ChatGPT under deadline pressure. The law-firm AI use policy framework we published earlier this year ports cleanly to small CPA firms with minor edits.
• Reset junior-staff training. Decide intentionally what reps the first-year still gets, what reps the AI now gets, and how the first-year develops the judgment they used to develop by grinding the reps the AI now does. This is the question the profession has not answered yet, and the firms that answer it first will retain the best juniors over the next five years.
• Audit your AI pipeline for citation completeness. Sample fifty returns. For every figure the model extracted, ask whether the citation actually points at the right document, page, and box. The first time you do this you will find more gaps than you expected. That is the point.
• Do not sign another generic intake or copilot SaaS until you have done the postmortem and identified the bottleneck. The default move — renew the vendor and hope next year is better — is the most expensive thing the firm can do in the next ninety days.

The thing nobody is saying out loud

Tax season 2026 was not the year AI replaced CPAs. It was the year AI started absorbing the work that used to make a CPA out of a first-year associate. Both of those statements are compatible. The firms that figure out how to keep training excellent reviewers in a workflow where the AI does the prep will look like the only firms with a senior bench in 2031. The firms that hand the prep work to AI and assume juniors will magically grow into seniors anyway will not.

That is the season’s real lesson, and it is bigger than any single workflow. AI did not replace the junior. It replaced the work that turned a junior into a senior. Putting that piece back into the workflow — intentionally, with partners involved — is the conversation the profession needs to have over the summer, while the season’s data is still warm.

Next step

If your firm is still picking through the WIP reports trying to figure out whether AI helped or hurt this season, the cheapest first move is a 30-minute conversation about which bottleneck to attack first. We will look at your numbers honestly, tell you which workflow is the right Q3 build for your firm, and tell you when the answer is “not yet.” Book the audit here.