Say no first.
The engagement starts with a paid, bounded, fixed-duration diagnostic — and it can legitimately end in “don’t proceed.” That is the point. Your actual risk isn’t missing an AI trend; it’s writing another check with nothing provable behind it. The diagnostic is priced so that finding out is cheap, and structured so you keep the findings either way.
Bounded scope, fixed structure, a record you keep.
- Scope and metric, agreed first One workflow, a measurable baseline, and executive sign-off on what “working” means — before any work starts.
- A fixed, time-boxed structure The diagnostic has a defined shape and duration. Time-to-first-value is measured in weeks, not quarters.
- A signed findings record What was examined, what was found, against which standard, by whom, on what date. It sits in your audit file whether or not we proceed.
- A go / no-go answer On a “go”: a pilot proposal with phase gates and off-ramps. On a “no”: the reasons, in writing, and the record is still yours. A no costs you the diagnostic and nothing else.
No autonomous anything against your real systems.
No autonomous outbound. No unattended agents over your data. And no rip-and-replace platform pitch — legacy systems are the top scaling barrier (69%, HBR Analytic Services/Appian, Apr 2026, N=385, vendor-sponsored), so the validation gate sits on top of what you already run, not in place of it.
The staffing pyramid bills you for the layer AI just commoditized — and rations the layer AI made scarce.
Big-firm economics are leverage on juniors billed at multiples. The field evidence (Brynjolfsson, Li & Raymond, QJE 2025; N=5,172) shows AI lifts novice output 30–35% while barely moving experts — the junior layer is exactly what got deflated. The scarce input is senior judgment, which the pyramid rations to partner-review hours. WTD inverts the stack: the entire engagement is the scarce layer; generation runs on cheap models behind the gate. You don’t need the army. You need the judgment the army rations — without the baggage train.
| The staffing pyramid | The gate |
|---|---|
| Armies of juniors billed at multiples — the layer AI commoditized | Senior judgment only — the layer AI made scarce; generation runs on cheap, gated models |
| Procurement cycle: a quarter. Minimum bet: six to seven figures | Decision: a signature. Minimum bet: a bounded diagnostic that can legitimately end in “don’t proceed” |
| Deliverable: a deck plus a dependency — land-and-expand by design | Deliverable: a signed record plus a Standards Ledger that lives in your repo |
| Accountability diffused across the pyramid; the partner rotates off; the disclaimer assigns responsibility to management | Residual risk owned by a named human |
| AI use inside the delivery model is opaque to the client | Validation is the product; every gate is documented |
| Extracts capability — your team watches | Deposits capability — your team runs the gates when we leave |
The continuity question, answered before you ask it: the story is not “we have more bodies” — it’s “you keep the system.” Every deliverable is plain text, version-controlled, and designed to be left behind. The documented system survives the consultant.
And the honest scope line: some problems genuinely need an army. If you need a 40-person ERP integration, that’s not us — and the diagnostic will say so.
The exposure is no longer hypothetical: in October 2025, Deloitte Australia agreed to refund the final installment (about AU$97,000) of an AU$439,000 report commissioned by the Australian Department of Employment and Workplace Relations, after the report was found to contain AI-fabricated citations and a fabricated quote from a federal-court judgment; the firm disclosed that generative AI had been used in its production. Unvalidated AI output is now shipping at consulting rates, with accountability diffused.
The pilot: one workflow, $0 until live, off-ramps at every gate.
One workflow, one automated step, one human judge, one signed manifest. Nothing is billed until the workflow is live and producing. Maximum downside is one phase — you can stop at any gate, having paid only for value received. And the exit stays credible throughout: open standards, a portable signed record, your encoded judgment in your version control. Fire me and keep your integrations.
One conversation to pick the workflow and the metric.
Then a time-boxed diagnostic to find out whether it deserves a pilot. If it doesn’t, you’ll be told so — in writing, with the record to keep.