Aldrich & Steele, a fictional mid-market public accounting firm with about 300 staff. Board pressure to "do something with AI" rising for four quarters. Zero pilots shipped. Managing partners cautious, no clear owner. Below is the artifact shape a real engagement produces: bet inventory, allocation table by reach, kill list, and the language the Managing Partner uses in the next board meeting.
Mid-market public accounting firm. Three offices, partner-led service lines in audit, tax, and a small advisory practice. Calendar fiscal year, busy season runs January through April 15. The board has asked the Managing Partner three quarters in a row to articulate the firm's AI strategy. The partner group is cautious by training, not by accident. The Tax Partner has read every vendor pitch and pulled the trigger on none. The IT Director has been quietly authoring a security memo about ChatGPT. Senior associates and managers are already using AI on personal accounts to draft client memos and reconcile workpapers, and nobody has named that fact in a partner meeting. Seven candidate bets surface in week one. None has a documented horizon, named owner, or kill criterion. This is the firm the Three-Horizon review is built for.
What the board wants: a defensible allocation, a named owner per bet, and language the Managing Partner can quote without overpromising. Not a chatbot. Not a vendor demo. An allocation thesis.
What we agreed to in week one: interview the partner group, the IT Director, two managers, and four senior associates. Inventory every AI initiative anyone in the firm has touched, sanctioned or shadow. Score each by reach (Core, Adjacent, Transform), payback profile, and named owner. Ship a board-ready memo by end of week four.
Four on Core, two on Adjacent, one on Transform. The classification surfaces what the firm is already doing in shadow, what's worth funding, and what should never have been on the list.
Status: Shadow use, two senior associates
Catches anomalies on draft 1040s and entity returns, drafts client-facing memos explaining variances year over year. Already in informal use by senior associates on personal AI accounts during last busy season. Real productivity gains, zero firm sanction, real data-handling exposure. The clearest Core win in the portfolio if the firm names an owner and a sanctioned tooling path.
Status: Vendor demo, no pilot
Auto-ties the trial balance to source documents, flags unsupported balances, drafts the lead-sheet narrative. Demoed by two vendors. No pilot scheduled. The audit partners are skeptical that an AI tool can survive peer review. The bet is real, the resistance is real.
Status: Vendor pilot, one client
Processes AP for one mid-size client as a managed-service add-on. Six-month pilot, no production owner, no scaling plan. Sitting on the boundary between a clean Core ship and a stalled pilot.
Status: Concept, partner discussion only
Pulls from prior-year engagement letters, flags scope changes, drafts the new letter for partner review. Discussed in two partner meetings. Nobody has been asked to build it. Two hours of partner time per engagement, multiplied by hundreds of engagements per year.
Status: Concept, Tax Partner sponsor
Researches edge cases across IRS publications, state guidance, and prior-year firm memos. Routes the research question to specialist agents (federal income, state and local, international, entity selection), returns a synthesized memo a partner can edit. New capability, not a faster version of an existing one.
Status: Concept, no funded team
Scores the firm's existing client base on fit for higher-margin advisory work (CFO services, M&A support, strategic tax planning). Surfaces the 40 clients most likely to expand into advisory inside 18 months. Pulls from billing, engagement notes, and industry data. Unlocks pipeline the advisory practice cannot manually surface.
Status: Concept, Managing Partner interest only
Replaces the annual audit for mid-market clients with an ongoing assurance subscription. Agents reconcile and review client books monthly, partners sign quarterly. Completely new product line, new pricing model, new partner-level risk posture. The Managing Partner has mentioned it in two board meetings. Nobody has been asked to size it. Owned at the Managing Partner level with board sponsorship, not delegated.
Aldrich & Steele's current implicit allocation is roughly 0% Core, 0% Adjacent, 0% Transform. Nothing has shipped. The shadow tax-return work is the only active production usage and it's not on the firm's books. The review's recommended mix anchors on the Bain headcount-output thesis: AI compounds when it lets the same headcount produce more output, not when it lets the firm cut headcount. For a partner-led service firm where the asset is senior judgment, Core dominates. Adjacent extends the practice. Transform is a single founder-level bet.
| Reach | Allocation | Named owner | What ships |
|---|---|---|---|
| Core Defend the practice |
70% | Managing Partner with COO support |
Tax-return prep copilot (sanctioned and tooled), audit-workpaper assistant (one-engagement pilot in Q3), engagement-letter drafter (built in-house, ships by Q4). AP bot folded into managed services. |
| Adjacent Build new capabilities |
20% | Tax Partner with IT Director support |
Multi-agent tax research desk gets a funded build with the Anthropic manager-IC principle as the design constraint: one orchestrator agent manages specialist agents, partner remains final reviewer. Advisory-fit scoring parked at concept with quarterly review. |
| Transform New operating model |
10% | Managing Partner with board sponsorship |
Agent-led continuous-audit offering gets a funded discovery sprint to size regulatory, peer-review, and pricing implications. No product build until discovery returns a defensible thesis. |
Core dominates because the practice asset is senior judgment, and Core bets compound senior-associate hours back to the partner group. Adjacent funds one capability with a named handoff plan. Transform carries one funded discovery sprint, no product build yet. The kill list (next section) removes three candidates and reframes a fourth.
One Core bet is worth a full section because it's already shipping in shadow. Naming it, sanctioning it, and tooling it correctly is the cleanest first move the firm can make. It's also the strongest story to tell the board: not "we're piloting AI," but "we're sanctioning what our senior associates already do, and tying it to the firm's risk posture."
Two senior associates ran roughly 40 returns through a personal ChatGPT account last busy season. The output caught three anomalies the reviewer would have missed and saved an estimated 6 hours per return on first-draft memo writing. The data exposure is real. The IT Director's security memo names this exact scenario.
Stand up a sanctioned tooling path with client-data controls. Name a Tax Partner as the production owner. Build the prompt library and eval set on the 40 returns already run, plus a stratified sample from the firm's prior-year archive. Train the senior associate cohort in week one of next busy season.
If the copilot captures even 8% of senior-associate hours across the tax line and the firm reallocates those hours to advisory work at advisory billing rates, the Core bets pay for themselves inside 6 months. That is the number the Managing Partner takes to the board.
The kill list usually saves more than the engagement costs. Aldrich & Steele's calls below recover partner attention, retire one vendor conversation, and refuse one bet that would have been visible but not Director-altitude.
Surfaced by the marketing director as the obvious starting point. Rationale: not Director-altitude. Doesn't compound. Doesn't touch the work that produces firm revenue. Partner attention is the scarce resource. Spend it on Core bets that move senior-associate hours, not on a widget that answers "what are your hours."
A line item the IT Director had been pricing as a defensible answer to the board. Rationale: a license is not an allocation. Buying seats without naming the workflows they apply to is the accounting-firm version of pilot purgatory. Defer the seat purchase until the Core bets have named tooling.
A real Adjacent bet, but the advisory practice has six FTEs and a 14-month backlog already. Rationale: scoring the pipeline before the practice can serve the pipeline creates exec frustration, not revenue. Park on the Adjacent roster with a quarterly review tied to advisory-practice hiring.
The vendor demos position the tool as an audit replacement. Rationale: that framing will not survive peer review and will not get partner buy-in. Rebuild the pilot scope as a workpaper-prep assistant the manager uses, not an audit conclusion the AI generates. Partner remains the signer. This is the Anthropic manager-IC principle applied to the audit line: agent handles the prep, human owns the judgment.
The artifact ships with the language the Managing Partner can quote, not just charts the board has to interpret. Below: the paragraph Aldrich & Steele's Managing Partner drops into the next board read-out. It neutralizes the "where is our AI strategy" question without overpromising, and it ties the allocation to a number the board can hold the firm to.
"We've structured our AI investment using the McKinsey Three-Horizon model, applied by reach rather than by year. Seventy percent defends the core practice with three production bets shipping this fiscal year, anchored on the tax-return prep copilot and the audit-workpaper assistant. Twenty percent funds one Adjacent capability, a multi-agent tax research desk that extends what our senior associates can take on without expanding headcount. Ten percent funds a single Transform discovery sprint on an agent-led continuous-audit offering, with no product build until discovery returns a defensible thesis. If we capture even 8% of senior-associate hours back to advisory work, the Core bets pay for themselves inside 6 months. We are not buying AI. We are reallocating partner and associate hours toward the work the firm should be selling."
A real Aldrich & Steele engagement runs four weeks. Interviews across the partner group, IT, and the working-level cohort that's already using AI in shadow. Working sessions with the Managing Partner and Tax Partner. A board-ready memo by end of week four, with named owners per bet and a follow-on scope for the Fractional Head of AI Transformation engagement that ships the Core bets.
Kickoff with the Managing Partner. Bet inventory: every AI initiative, sanctioned or shadow. Interview list locked. Board read-out date confirmed.
Interviews: partner group, IT Director, two managers, four senior associates, one client-facing administrator. Initial reach classification (Core, Adjacent, Transform) for each candidate bet. Shadow usage surfaced and named.
Allocation thesis with the Managing Partner and Tax Partner. Kill list with the COO. Failure-mode review per reach category. Owner-naming pass.
Bet inventory, allocation table, kill list, board-defense paragraph, named owners, named follow-on engagement options. Walkthrough with the Managing Partner. Memo handed off.
Deliverables: bet inventory document, allocation table with named owners by reach, kill list with rationale per bet, board-defense paragraph the Managing Partner can quote, and follow-on scope for a Fractional Head of AI Transformation engagement to ship the Core bets.
Three to four weeks. Me plus your Managing Partner plus your COO or IT Director. Board-ready output. The kill list usually pays for the engagement on its own. Typical follow-on is a Fractional Head of AI Transformation engagement to ship the Core bets.