Grow the Practice with Governed AI

The license, the hierarchy, and the season: how a CPA firm actually works

The principle

Strip the mystique and a CPA firm is a licensed hierarchy with a calendar. It sells assurance and advice under a license, produces them through a preparer-to-reviewer-to-partner ladder, and the signature at the top carries the liability for everything beneath it. Three facts govern the economics. Capacity is brutally seasonal: the firm that is comfortable in June is drowning in February, and the year is won or lost in a few compressed months. Realization is fragile: it erodes one small write-down at a time, usually discovered at billing, after every decision that caused it has already been made. And the constraint everyone feels first is people: in the AICPA's 2024 top-issues survey (N=667), finding qualified staff ranked as the top issue for every firm category except sole practitioners, with retention second at the largest firms (AICPA, 2024; self-reported rankings, not a measured shortfall, and not an AI finding). Growth plans stall exactly where licensed attention runs out, because reviewers and partners who can sign do not scale by wishing.

The trap

The trap is chasing throughput with tools while leaving the operating model alone. A research subscription here, a client portal there, and next season the memos arrive faster but review is still the bottleneck, the document chase still eats the first six weeks, and realization still surprises everyone at billing. The licenses were real. The season never moved, because nobody redesigned the workflows the tools were supposed to serve.

The checklist

• Name the constraint by service line: winning work, producing it, reviewing it, or signing it. They are different problems with different fixes.
• Follow one in-season week of partner and manager time and mark the hours only a license could have spent.
• Count what the firm rebuilds every year: request lists, memos, schedules, research already done in a prior file.
• Put honest numbers on last season: write-downs by client, overtime against plan, files that started incomplete.
• Choose workflows to improve, not tools to buy. A tool dropped on an unchanged workflow returns almost nothing.

Where judgment beats the tool

An hours analysis shows where licensed time goes. It cannot say which of those hours are the product. Some partner hours spent slowly, on a hard call or a client sitting across the desk, are exactly what the client is paying for, and automating around them would cheapen the firm. Deciding which senior hours are the value and which are habit is a call only the partners can make.

Four outcomes that matter, and one that does not

The principle

Four outcomes justify this whole program. Grow without proportional hiring, so the practice can take more work than the hiring market will staff. Expand capacity without thinning the review hierarchy, so growth never quietly spends what the signature means. Turn institutional knowledge, the prior-year files, research memos, and review notes the firm already paid for, into controlled, reusable leverage, so the best work compounds instead of retiring with its authors. And adopt AI while protecting client confidentiality, independence, and trust, because in this profession trust is the license to operate. Notice what is not on the list: adopt AI. Adoption is a means. The moment it becomes the goal, the program starts optimizing for usage instead of for the firm.

The trap

The trap is measuring the means. Licenses issued, seats active, prompts per preparer per week: activity metrics reward the appearance of change while the operating outcomes sit unmeasured. A firm can hit every adoption target it sets and end the next season with the same incomplete files, the same review bottleneck, and the same write-downs discovered at billing.

The checklist

• Write each of the four outcomes as an operating sentence with a named owner, not a slogan on a slide.
• Tie every initiative to exactly one outcome. An initiative that maps to none is a hobby.
• Baseline the outcome metrics before the first pilot, and before the season, or the after will have no before.
• Retire activity metrics from partner reporting. Keep them for operations, where they belong.

Where judgment beats the tool

Outcomes conflict at the margin: pushed far enough, capacity presses on review depth, and reuse presses on the fresh skepticism each engagement deserves. A dashboard will not arbitrate that tension. Where the firm sets each trade, engagement by engagement, is a leadership decision, and it is the one clients experience directly.

Who has to say yes: the six chairs in the room

The principle

Nothing durable happens in a partnership without consensus, and this program touches every chair at the table. The managing partner asks whether it grows the practice without endangering what the practice is built on. The COO or firm administrator asks whether the season gets more reliable or just busier. Partners-in-charge ask what happens to the quality of work that carries their name and their license. The CIO asks whether this becomes one governed platform or a subscription in every cubicle. The quality and risk partner asks what it does to confidentiality, independence, and the standards the firm answers to. The CFO asks what it costs, what it returns, and who will stand behind the number. Six different questions, and the program has to hold a real answer to all of them.

The trap

The trap is the champion-led initiative that answers one chair. It moves fast on borrowed enthusiasm, then dies in a partner meeting the day the quality partner asks the independence question nobody prepared for, and the silence costs the champion a year of credibility. In a consensus organization, the unanswered chair is a veto waiting for its moment.

The checklist

• Map the six chairs to named people, including the ones who hold the role without the title.
• Write down each chair's question and the evidence that would satisfy it, before the program is proposed.
• Brief the skeptics privately before the partner meeting, not at it.
• Give the quality and risk partner a genuine design seat. Controls added at the end read as concessions; controls designed in read as competence.

Where judgment beats the tool

An org chart names the titles. It does not reveal whose no actually ends a program in your partnership, or whose quiet yes moves the undecided. Reading the real decision structure of your own firm is judgment, and no tool has ever held it.

Workpaper & Research Leverage: first drafts that reach the reviewer faster

The principle

Workpaper & Research Leverage runs from the prior-year file to a draft on a licensed reviewer's desk: retrieval of last year's workpapers and memos, technical-research support from approved sources, and first drafts of the schedules, narratives, and memos that make up an engagement file. It is where most preparation hours go, and most of those hours are assembly: rolling forward, hunting for what the firm already found last year, formatting what a reviewer will reformat anyway. The drafting layer is where the causal evidence is strongest: in a preregistered experiment, access to a general AI assistant cut time on occupation-specific professional writing tasks by about 40 percent while raising judged quality about 18 percent (Noy and Zhang, Science, 2023; N≈453, general writing tasks, not tax or audit conclusions, and evaluator-scored rather than realization-measured). The adjacent pattern matters for a profession that hires its staff young: across the 2023 causal studies, AI assistance narrowed the gap between less and more experienced workers on supported tasks (Noy and Zhang; Dell'Acqua et al.; Brynjolfsson, Li and Raymond). That is a staff-leverage story, and it is an inference from task-level studies, not a measured onboarding ROI; it raises the stakes on review rather than lowering them.

The workflow, stage by stage

• Prior-year retrieval: last year's workpapers, memos, and carryforward items surface with provenance, instead of living in whoever prepared them, wherever they now work.
• Technical-research support: guidance arrives synthesized from the firm's approved research sources with citations attached, for a licensed professional to verify, never to take on faith.
• First drafts: schedules, narratives, and memo drafts start from firm templates and the prior-year file, so preparers edit upward instead of assembling from a blank workpaper.
• Reviewer handoff: drafts reach the reviewer with sources visible and open questions flagged, so review notes address judgment instead of formatting.
• Write-back: what review corrected gets filed with the engagement, so next season's first draft starts from this season's best.

The trap

The trap is the draft that reads finished. A fluent memo with a confident conclusion invites a saturated reviewer in March to skim what they would once have rebuilt, and a conclusion no licensed professional has actually owned slips through on good formatting. The failure is not the draft. It is the moment fluency gets mistaken for review.

The checklist

• Restrict research support to the firm's approved sources, with citations a human verifies before anything relies on them.
• Curate the template and prior-year corpus with an owner and review dates, so drafts start from the firm's best, not the file share's oldest.
• Label every AI-assisted draft as unreviewed until a licensed reviewer owns it. No unlabeled drafts, ever.
• Keep conclusions human: the tooling assembles, cites, and formats; the licensed professional concludes and signs.
• Measure preparer-to-reviewer turnaround and review notes per engagement together. Either one alone will lie to you.

Where judgment beats the tool

Research support can find the guidance. It cannot decide whether it applies to this client's facts, whether what is in the file is enough, or whether the position is one the firm will put its name behind. Those calls are the license, and they do not delegate.

Client Onboarding & PBC: end the document chase

The principle

Client Onboarding & PBC is the least glamorous workflow in the firm and one of the most expensive: request lists assembled by hand, documents trickling into inboxes and portals, staff reconciling what arrived against what was asked, seniors chasing what did not, and engagements starting weeks late with files still incomplete. The chase consumes exactly the weeks when capacity is scarcest, and it sets the tone of the client relationship for the year. Rebuilt on governed rails, the same workflow runs in the background: complete request lists at acceptance, intake tracked automatically, communication drafted for human review, and licensed attention entering only where judgment lives.

The workflow, stage by stage

• Request-list assembly: the PBC list builds from the engagement type and the prior-year file, complete at acceptance instead of growing by discovery in the busiest weeks.
• Intake tracking: what arrives is logged and matched against what was requested, so completeness is a fact on a status board rather than an inbox archaeology project.
• Client communication: reminders and status notes go out on a promised cadence, each one drafted for review and owned by a person before it sends.
• Escalation: what reaches a senior desk is the judgment call, the missing item that changes the work or the client who has gone quiet, not the chasing itself.
• Season readiness: engagements start when files are complete, and the firm can finally see, in one place, which ones are not.

The trap

The trap is turning clients into ticket numbers. An automated chase that sends a fourth identical reminder to a controller who just lost half her team is operationally correct and relationally ruinous. The old failure was the incomplete file in March. The new one is the client who felt processed by a machine all spring, and remembers it when the engagement letter comes up for renewal.

The checklist

• Build the complete request list at acceptance, from the engagement type and the prior-year file, and version it when scope changes.
• Keep one intake status both sides can see, so where things stand is shared instead of contested.
• A human reads, edits, and owns every client-facing reminder. Clients cannot tell which messages you considered routine.
• Escalate judgment, not chasing: senior time enters for the item that changes the work, never for the third follow-up.
• Measure days from acceptance to a complete file, by client, season over season.

Where judgment beats the tool

The tracker knows what is missing. It cannot know that this client's finance team is mid-crisis, that this relationship needs a partner phone call instead of a fourth reminder, or that the missing schedule is a signal about the client rather than about the mail. Reading the silence is the senior skill, and it does not automate.

Client-to-Partner Insight: the whole book, visible in season

The principle

A partner in season runs a book of engagements on memory, hallway status, and whatever the practice systems cough up on Friday. Client-to-Partner Insight rebuilds that picture from live artifacts instead: status across the book synthesized from files, time entries, and intake reality; every deadline visible with its dependencies; hours against budget by engagement while the engagement is still open; and the signals hiding in routine work, surfaced for a partner to judge, never sent anywhere on their own. The picture stops being a Friday performance and becomes an instrument the season can actually be steered with.

The workflow, stage by stage

• Status across the book: engagement status synthesized from working files, time entries, and intake records, not from memory under deadline.
• Deadline tracking: every filing and deliverable date visible with its dependencies, so nothing is discovered the week it is due.
• Realization signal: hours against budget by engagement while the work is live, so a write-down becomes a decision instead of a discovery.
• Advisory signal: patterns worth a client conversation surface for partner review, and go nowhere without a licensed professional deciding they should.
• Partner read: the partner reviews the picture before it drives any action. Synthesis prepares the story; it does not get to tell it.

The trap

The trap is synthesis that launders the season. A model summarizing engagement status will smooth a slipping file into confident prose, and a partner can read amber for a month on a book that is quietly red. The failure is not the summary. It is a reporting chain where no number can be challenged, because no number can be traced.

The checklist

• Give every number in the book one source of record, and synthesize from artifacts, not from prior summaries.
• Make every figure drillable to the file, time entry, or intake record behind it, in one step.
• Read realization from hours against budget while the engagement is open, when pricing, scope, and staffing can still respond.
• Escalate to a named partner with a date, never to a distribution list.
• Route advisory signals to a partner's judgment, never to a client. What the firm says to a client is a licensed professional's call.

Where judgment beats the tool

Synthesis can flag that a client's numbers moved strangely or that a deadline cluster is about to land. It cannot decide which risk is worth a partner's Saturday, which client gets the proactive call, or whether a pattern in compliance work should become an advisory conversation, at what scope, in which relationship. Those are the calls the book exists to serve, and they belong to the partner whose name is on it.

Review Hierarchy & Sign-off: the org chart is the quality system

The principle

Every profession has a quality system. In a CPA firm it is the org chart itself: the preparer does the work, the reviewer questions it, the partner signs it, and the license at the top answers for all of it. That ladder is how one signature can honestly stand behind thousands of pages the partner never personally prepared. AI changes what enters the ladder, drafts arrive faster, better assembled, with sources attached, and it must change nothing about the ladder itself. The studied risk is the reason. The 2023 field experiment that reported gains inside AI's capability frontier also reported professionals' quality falling when they relied on it for tasks outside that frontier (Dell'Acqua et al., 2023; one elite consulting firm, task-level results, not CPA work), and where the frontier sits shifts with the model and the domain. For an accounting firm, the outside-frontier territory is precisely the territory that matters most: whether a technical position actually has authority behind it, whether the evidence gathered is sufficient to support a conclusion. Those are named here as risks the hierarchy exists to catch, not as advice about any engagement.

The trap

The trap is the thinning temptation. Drafts arrive so clean that detail review starts to feel redundant, and someone proposes banking the savings by removing a layer. The hierarchy reads as cost precisely when it is working, because the failures it prevents never happen. Thin it, and the first bad conclusion to slip through will cost more than every hour the thinning saved.

The checklist

• Hold review depth constant regardless of where a draft came from. Provenance changes the reviewer's starting point, never the standard.
• Flag AI-assisted drafts as such, so every reviewer knows what they are holding and reads with eyes open.
• Track review notes per engagement and read the trend honestly: falling because drafts improved is the win; falling because reviewers tired is the warning.
• Keep sign-off a personal act. No tool, template, or default ever stands in for the professional whose name goes on the work.
• Sample-check the cleanest-looking files precisely because they look clean. Fluency is not evidence.

Where judgment beats the tool

A checklist can verify that a file is complete. Only the licensed reviewer can judge whether it is sufficient, and only the signing partner can decide that the firm stands behind it. The hierarchy is not friction on the product. In an accounting firm, the hierarchy is the product.

Where the risk lives: confidentiality, independence, and client trust

The principle

A CPA firm is trusted with the most confidential thing a client has: its finances. That trust, held inside the profession's own obligations, is the asset, and AI touches it in four places. Confidentiality: client financial data entering prompts, corpora, and tools no engagement ever approved. Independence: tool arrangements and data flows on attest work carry obligations that a general-purpose rollout will not notice on its own. Quality: fluent drafts that are wrong in ways a saturated reviewer will not catch, the documented risk the review hierarchy exists to hold (Dell'Acqua et al., 2023; one elite firm, task-level results), and where that frontier sits shifts with the model and the domain, which is why licensed review is a standing control rather than a launch-phase one. And trust itself: some clients have started asking how their firm uses AI on their engagement, and a few now put the question in the engagement letter (a practice observation, not a measured trend; the research library holds no source on how widespread this has become), so the firm needs an answer it is proud of before the question arrives. None of these risks argues against the program. All of them argue for running it inside boundaries, under one rule that never moves: AI assists operations and documentation, and accountable licensed professionals retain every judgment and every sign-off.

The trap

The trap is treating confidentiality as an IT setting instead of an operating boundary. A tenant toggle does not know which engagement a document belongs to, what this client's engagement letter prohibits, that two clients in the same corpus are competitors, or what independence requires on this attest relationship. General policy plus default settings is how a firm ends up technically compliant and actually exposed.

The checklist

• Set data boundaries at the engagement level, mirroring staffing: who is on the engagement defines what the tools can reach.
• Hold attest engagements to the strictest posture in the firm, and make the strictest posture the default question, not the exception.
• Keep an approved-tool list with a sanctioned path for new needs, so the shadow alternative never becomes easier.
• Write the disclosure answer before it is asked, and make it one the managing partner would happily read aloud.
• Stand up an incident path that assumes a miss will eventually happen, and rehearses who tells the client what, and when.

Where judgment beats the tool

A control framework can set the floor. It cannot weigh whether this engagement, this client, this obligation justifies more caution than the standard, or which client deserves a proactive conversation rather than a policy reference. Proportionality is a judgment about relationships, obligations, and consequences, and it belongs to the partners who answer for them.

The maturity path: baseline, AI-enabled, selectively AI-native

The principle

There are three honest stages, and they belong to workflows, not to firms. Baseline is where most work lives today: knowledge in people, drafting by hand, chasing by inbox, status by meeting. It is not failure; it is a leverage ceiling. AI-enabled is the working middle: the same workflow, with retrieval, drafting, and synthesis assisting inside governed boundaries, and the review hierarchy exactly where it was. Most workflows should live here. Selectively AI-native is the far stage: a workflow redesigned around governed AI because its volume and structure justify it, with licensed judgment and sign-off human by design. The operative word is selectively, and in this profession the calendar adds its own law: the path is walked between seasons. A firm changes its production system in the quiet months and proves it before the deadline wave, or it gambles the season that pays for everything else.

The trap

The trap is the transformation announced in November. A whole-firm program colliding with the busiest quarter guarantees the new system and the old one run at once, in the weeks with no slack to reconcile them, and the season decides the program's reputation before the program was ready. The opposite trap is subtler: staying enabled forever out of comfort, while the document chase and the prior-year rollforward have long since earned their redesign.

The checklist

• Place each significant workflow on the path separately. The firm does not have a maturity level; its workflows do.
• Earn native with evidence from enabled: volumes, error rates, and review-note trends, not enthusiasm.
• Sequence to the calendar: pilot after the deadline, codify in the quiet months, enter the season with a system already proven.
• Keep judgment and sign-off human at every stage, including native. What moves is assembly, never accountability.
• Revisit placements after every season, while the evidence of what strained is fresh.

Where judgment beats the tool

A maturity model can place a workflow on the path. It cannot decide whether the redesign is worth the disruption this year, in this office, in this service line, with this bench. Sequencing is strategy, and strategy is what the partnership is for.

Governance and measurement: rails under the work, numbers you can stand behind

The principle

Governance here is a small set of operating patterns, each of which buys speed as well as safety. Engagement boundaries decide what AI can reach. Access mirrors the engagement hierarchy. Approved sources decide what research and drafts may be built from, and provenance rides along, which is what makes review fast. Licensed review stands before anything client-facing or file-final, with quality criteria written where the reviewer works, and sign-off keeps its name. Retention rules for prompts and outputs align with the firm's workpaper retention obligations, and reuse rules say what carries into next season and what stays with the engagement. That is the whole apparatus, and it runs on the hierarchy the firm already has. Then measurement, because governance without measurement is theater. Six numbers tell you whether the machine is working, and they are the firm's own: days from acceptance to a complete PBC file, preparer-to-reviewer turnaround, review notes per engagement, realization by client and service line, peak-season overtime against plan, and on-time filing rates. Baseline them before the first pilot, count program cost once across every workflow that shares it, expect value to arrive after an adoption lag, and present ranges the partnership can believe. Treat any hours-saved figure that did not come from your own baseline as a hypothesis, and price it as one. And the two numbers a partner group most wants do not exist: the realization impact of AI drafting is unmeasured, and there is no credible mid-market CPA-firm ROI figure in the public record. Your baseline is the only number that will ever be about your firm.

The trap

The trap is vendor math on a seasonal business: hours saved that nobody redeployed, benefits counted in three business cases at once, and a March improvement annualized as if every month were March. A partner group hears one inflated number and discounts the entire program, including the parts that were true. In a firm whose product is rigor, the business case is a work sample.

The checklist

• Build controls into the path of work as defaults, not into a policy document as clauses, and check quarterly for the workflows people route around.
• Price every control by what it buys: speed gained or risk retired. A control that buys neither is friction wearing a badge.
• Baseline every KPI before the pilot that is supposed to move it, and read results across a full season before scaling anything.
• Count shared program cost once, model the adoption lag explicitly, and let each workflow carry only its marginal cost.
• Name where freed hours go: more clients served, advisory conversations, review depth, shorter Saturdays. Unassigned capacity evaporates.

Where judgment beats the tool

The measurement can prove capacity was freed. It cannot decide whether that capacity becomes growth, margin, or a bench that reaches the end of the season intact, and it cannot make the partnership enforce the choice. Value realization is a leadership act. The tools only make it visible.