AI in Tax Practice (2026 Edition): What Actually Works at a Boutique Firm

#The frame

Most “AI in accounting” content falls into two camps:

Camp 1: AI will replace accountants. Usually written by tech publications. Wrong about the specifics (regulatory + liability constraints prevent end-to-end AI tax prep), wrong about the timeline (10-year framing for changes that are more 2-3 year), wrong about the impact (boutique firms benefit more than they get displaced).

Camp 2: AI is dangerous + unreliable + we’re not touching it. Usually written by traditional firms protecting status quo. Wrong about the technology (current models are far better than 2023 baselines), wrong about competitive dynamics (firms that don’t adopt will lose to firms that do), wrong about client expectations (clients increasingly expect AI-enabled response times).

The honest middle: AI is genuinely useful for specific tax-firm workflows, useless for others, and dangerous for a third category. The trick is knowing which is which.

This article is what we’ve actually deployed at ETS — what works, what doesn’t, what we’re testing, what we won’t touch.

#What we use AI for (and how)

#Use case 1: Document parsing + extraction

When a client uploads a W-2, 1099, K-1, or tax return PDF, AI extracts the structured data (employer name, wages, federal withholding, state withholding, box codes) into a normalized format. Used to take 5-10 minutes of manual data entry per document. Now takes ~30 seconds with human review.

For an engagement with 20-50 source documents, this is hours of reclaimed time per client. Multiplied across 30 active clients per partner, this is a real capacity unlock.

Tool stack: GPT-5 / Claude / specialized tax-data extraction services. Output reviewed by a human before it enters our books or returns.

#Use case 2: Bank-feed categorization (in Kick)

Our bookkeeping platform Kick has built-in AI categorization. It learns from the categorization patterns specific to each client over time. For a client with 500 transactions/month, AI gets ~85% of categorizations right out of the box and ~95% after a few months of learning.

The 5-15% that AI gets wrong is where the human bookkeeper adds value — and that’s exactly where AI was always going to struggle (edge cases, ambiguous transactions, anything requiring business context AI can’t have).

#Use case 3: Tax research

For tax-law research (recent provisions, edge cases, IRS rulings, state-conformity questions), AI-assisted research tools (specialized tax-AI like Blue J, Spellbook, plus general models) dramatically accelerate the work. A research question that used to take 1-2 hours of CCH / RIA Checkpoint navigation now takes 5-15 minutes.

Critical caveat: AI citations are validated by humans before they appear in client work. Hallucinated case law and made-up code sections are a real risk. The pattern: AI finds the relevant authority faster; humans verify it’s actually applicable.

#Use case 4: First-draft client communications

For routine client emails (engagement letters, status updates, document requests, meeting recaps), AI generates clean first drafts in seconds. Human editor adjusts for voice, accuracy, and context.

Saves ~70% of the time spent on email drafting. The voice still feels like ours because we edit for tone before sending. No client-facing communication goes out without human review.

#Use case 5: Operational documentation

This is the unsexy use case nobody talks about. Internal SOPs, runbooks, training docs, process documentation — AI is incredible at producing first drafts of operational content based on conversational descriptions of how things work.

Multiplied across a small firm, having genuinely good operational documentation that didn’t take partners 40 hours to write is a real productivity unlock.

#Use case 6: Code + automation

Our internal systems (Basecamp automations, Kick integrations, custom dashboards, this website) involve real software development. AI-assisted coding (Claude Code, Cursor, GitHub Copilot) lets us build internal tools at ~3-5× our previous pace.

The ETS website you’re reading right now was built primarily through AI-assisted development against tight specifications. Human review at every step; AI accelerating execution.

#What we use AI for vs what we don’t

The clearest way to see the line we draw: the same capabilities that are safe on the left become dangerous on the right.

A note on each “don’t”:

Final tax positions. Every tax position on every client return is reviewed by our partners (CEO for tax substance, COO for client-facing communication). AI can suggest research; AI doesn’t decide positions.

Form 8867 due-diligence affirmation questions. These are EA-personal-attestation questions under Circular 230 §10.34. Only the licensed EA can answer them. AI is not allowed near these fields. This is the strictest part of our tax-prep workflow.

Client-facing real-time communication. Our 1-3 business day response SLA on Basecamp is met by humans, not AI. Clients book us to talk to us. If we needed AI to respond to them, they’d be buying a different product.

IRS representation + audit defense. Anything involving direct IRS interaction (POA, audits, appeals, Tax Court) is human-only. Not because AI isn’t capable; because the regulatory + relationship dynamics require human presence.

Anything where AI hallucination is a tax-position risk. The most dangerous AI failure mode is confidently wrong output that gets cited as authority. Code section numbers that don’t exist. Cases with the wrong holdings. Regulations that say the opposite of what AI claims they say. We verify every AI-sourced citation before it touches client work.

#The hybrid workflow

What it actually looks like for a typical Tax Analysis engagement. Total partner-hours per Tax Analysis: ~16 hours per engagement, down from ~24-28 pre-AI.

That 30-40% time reduction is what lets us cap at 3 new clients per month without expanding the team. The cap is the same; the throughput per partner is meaningfully higher. Clients still get the partners on every call — they just get them faster than before.

#What we’re testing (not yet deployed)

A few capabilities we’re piloting but haven’t put in production:

1. Client portal voice assistant. Basecamp portal Q&A bot that can answer client questions about their own documents + their engagement status. Currently testing. Not deployed because the trust/liability dynamics aren’t clear yet.

2. Predictive estimated-tax modeling. ML-based quarterly estimated tax calculator that learns from each client’s actual income patterns. Currently in alpha testing.

3. Automated bookkeeping reconciliation. End-of-month reconciliation that flags discrepancies, suggests adjustments, and produces audit-ready output. Promising; not production-ready.

4. Automated state-tax conformity tracking. A system that tracks which state has conformed to which federal tax provision (OBBBA in particular has 50-state conformity confusion). Tracking this manually is annoying; automating it would save weekly research hours.

#What’s clear by 2026

A few patterns we now consider settled:

1. AI-assisted tax research is a real productivity unlock — but humans verify citations. Hallucinated authorities are the #1 failure mode.

2. AI-assisted bookkeeping categorization works and gets better with client-specific training data. This was the 2023-2024 promise that’s now reality.

3. AI doesn’t replace partner judgment. It shifts what partners spend time on. Less time on data entry, more time on strategy + client relationship.

4. Boutique firms benefit more than mid-size firms. Larger firms have more associates whose work AI partially replicates; smaller firms have less associate work to begin with and use AI to expand partner throughput.

5. Liability + regulatory frameworks lag the technology. Circular 230, state-bar rules, malpractice considerations all assume human judgment at every step. AI deployments have to work within these constraints, not around them.

#What we tell clients

Three things, when AI comes up in conversation:

1. We use it. Aggressively. Where it adds value. We don’t hide that.

2. It doesn’t replace the human work you’re paying for. Every position is human-reviewed. Every form is human-signed. Every conversation is human-led.

3. It lets us deliver more substance per dollar than firms that aren’t using it. Bigger Tax Analyses. Deeper research. Faster turnaround. Same partner-client model.

If you’re a prospect comparing ETS against another firm, ask both firms what they’re using AI for and how. The answers will tell you a lot about how each firm thinks about the next decade.

#What I’d tell other firm builders

If you’re running or starting a tax / accounting firm and thinking about AI adoption:

Don’t try to replace human work end-to-end. The regulatory + liability environment makes this dangerous. Don’t be the test case.

Do automate the parts of the workflow that don’t require judgment. Document extraction. First-draft generation. Research acceleration. Bookkeeping categorization. These are where the wins are.

Verify every AI output before it touches client work. Especially citations + numerical claims. AI confidence is not correlated with AI correctness.

Invest in operational documentation. AI is shockingly good at producing it from conversational input. Every SOP you have written down is leverage for future hires + future automation.

Stay close to the regulatory developments. AICPA, state CPA societies, NAEA — they’re all forming positions on AI in practice. Don’t get caught with a deployment that creates compliance issues.

The firms that figure out the human-AI hybrid well over the next 3-5 years will be the boutique firms of the next decade. The firms that either reject AI entirely or try to fully automate around it will both struggle.

---

This article is part of “From the Operator Seat” — Ramon’s lane on firm philosophy + operations. The articles in this category are for prospects who want to see how the firm actually works, and for other tax-firm builders thinking about the same questions.

If you’re a prospect comparing tax firms and wondering how AI deployment affects what you’d get from ETS, the Discovery call is the right next step. Honest answer: we use it heavily where it adds value; every partner-led conversation is still partner-led.