By Errin O'Connor — Founder & Chief AI Architect, EPC Group · Microsoft Solutions Partner across all six designations · G2 Leader in BI consulting
In an essay called "Choosing to Stay Human," Ethan Mollick named the risk that the productivity headlines keep stepping over. Agentic systems, he wrote, are designed to just do stuff — which is great for getting stuff done, and bad for learning anything, staying authentic, or avoiding what he calls cognitive surrender: the temptation, when a hard request returns a polished answer, to simply go with it. His team's research behind the warning is the famous experiment with 758 Boston Consulting Group consultants — AI-equipped consultants vastly outperformed their peers on tasks inside the AI's competence, and the study's enduring lesson lives at the jagged edge, where the AI was confidently wrong and the humans who had surrendered to it followed it straight over.
Oxford's Matthias Holweg and Tom Davenport have generalized the same warning to the whole enterprise — repeated AI summarization and rewriting eroding original knowledge and judgment across business processes, a decay they call slopification. And Satya Nadella has supplied the framing that turns the warning into strategy: the modern enterprise runs on two kinds of capital — human capital, meaning judgment, relationships, and pattern recognition, and token capital, the AI capability itself — and the firms that win pair them, because every human-guided workflow generates a better training signal that compounds into tacit knowledge the firm alone owns.
Read those three together and a conclusion falls out that I have not seen a single finance organization act on: your workforce's judgment is a productive asset. Ungoverned AI deployment depreciates it. And unlike every other depreciating asset you own, nobody is booking the write-down.
How the write-down happens — quietly, and in your best people
The mechanism is mundane, which is why it is invisible. The senior analyst who used to build the forecast now reviews the AI's forecast — then, after enough clean quarters, skims it — then approves it. The reviewer's edge was never the arithmetic; it was the pattern library built from years of doing the arithmetic, the instinct that says this number smells wrong before knowing why. Pattern libraries are maintained by use. Stop building forecasts and the library starts aging the day you stop. Multiply that by every summarized document, every drafted contract, every "resolved" reconciliation across the enterprise, and you get Holweg and Davenport's decay at scale — with a twist that should terrify you: the erosion concentrates in exactly the people whose judgment you will need most on the day the AI is confidently wrong. The BCG study's jagged edge is not a lab curiosity. It is a Tuesday in your close process, waiting.
And here is the balance-sheet irony: the enterprise meticulously tracks the cost side of AI — licenses, tokens, capacity — while the offsetting depreciation runs through an asset that never appears on the books. If a machine that produced your forecasts were corroding, maintenance would be budgeted by Thursday. The judgment that checks the machine is corroding, and the budget line does not exist.
The Judgment Ledger
The framework I build with clients treats judgment preservation as a governance discipline with an owner, not a wellness poster. We call it the Judgment Ledger, and it has four entries.
Designated human reps: for every AI-accelerated workflow that is financially material, a defined cadence where humans still do the work unassisted — the forecast built by hand once a quarter, the reconciliation walked manually monthly — not for output, for maintenance of the pattern library, exactly the way pilots keep manual landings current.
Verification with teeth: review steps that require the human to produce something — a stated expectation before seeing the AI's number, a documented challenge — because "reviewed and approved" without friction is surrender with a signature.
Ground-truth traceability: every AI-altered artifact traceable to a verifiable human source, Holweg and Davenport's own prescription, so the game of telephone has a first speaker on record.
A named owner: someone at officer level accountable for the human-capital side of the AI program — measuring where judgment is thinning, adjusting the reps, reporting it to the board alongside the token spend. That seat is precisely what a Virtual Chief AI Officer holds for organizations not yet ready to fill it full-time.
Twenty-nine years of enterprise consulting has shown me every version of tool-induced atrophy — the DBAs who lost the ability to read a query plan after the optimizer got good, the admins who couldn't restore a backup manually the night the console was down. Each time, the organizations that kept a deliberate manual-competence floor looked wasteful right up until the day they looked prophetic. AI is that pattern applied to thinking itself. The stakes went from a bad night to a bad filing.
What I tell clients to do
One. Inventory the financially material workflows where AI now produces and humans now approve. That list is your depreciation schedule.
Two. Institute designated reps on the top five — unassisted human execution on a written cadence, framed and budgeted as asset maintenance, because that is literally what it is.
Three. Redesign approval steps to require produced judgment — an expectation stated before the AI's answer is revealed — and log it. The delta between expectation and output is the cheapest anomaly detector you will ever deploy.
Four. Put human capital on the AI program's board report, next to token capital, with an owner. Nadella gave you the vocabulary. Use both columns.
Reps cadence by workflow class
| Workflow Class | Designated Rep Cadence |
|---|---|
| Financially material forecasts | Unassisted human build 1×/quarter |
| Reconciliations | Manual walk 1×/month on rotating accounts |
| Contract review | Blind clause-extraction spot-checks weekly |
| Executive summaries | Author writes the expectation BEFORE seeing AI output; delta logged |
| All classes | Expectation-vs-output delta feeds the anomaly dashboard (cheapest detector you'll ever run). Owner: vCAIO reports judgment-health next to token spend, quarterly, to the board. |
Where I land
Mollick is right that staying human is now a choice, and choices made passively get made badly. Nadella is right that judgment is the capital that makes the tokens compound — the training signal your firm alone owns. My addition is the CFO's translation: judgment is an asset with a maintenance schedule, and an AI program that books the license cost while ignoring the judgment write-down is overstating its own returns. Boring AI is winning AI — and the most boring line item of all, deliberate human reps, is the one that keeps the winning compounding.
The data behind this (sources and verification)
- Mollick, "Choosing to Stay Human" (One Useful Thing, 2026) — Agentic systems 'just do stuff' — the cognitive-surrender warning; the 758-consultant BCG experiment (Dell'Acqua et al.) and its jagged-frontier lesson.
- Holweg & Davenport (CIO/HBR via Computerworld, June 22, 2026) — Slopification, verification, validation, entropy, model collapse — repeated AI summarization eroding original knowledge at enterprise scale.
- Nadella: human capital vs token capital — 'Every improved workflow generates a better training signal' — judgment as the firm-specific compounding asset.
- Microsoft Research/CMU CHI 2025 study on generative AI and critical thinking (Lee et al.) — Reported reduced critical-thinking effort with higher AI confidence. [VERIFY exact finding phrasing before quoting.]
Third-party figures above are attributed to their named sources as of the Last verified date. EPC Group audit figures are directional findings from client engagements. Items marked [VERIFY] must be confirmed before external quotation.
Frequently asked questions
What is cognitive surrender?
Accepting AI output without engaged verification — Mollick's term for the failure mode agentic convenience invites. When a hard request returns a polished answer and the human simply goes with it, the pattern library that would have caught a wrong answer stops being maintained.
Is AI deskilling proven?
Research from BCG/Harvard/Wharton experiments to Microsoft/CMU survey work points the same direction: less engaged verification as AI confidence rises. Treat it as a managed risk, not a debate — the BCG jagged-edge finding alone is sufficient to require a governance response.
What are designated human reps?
Scheduled unassisted execution of AI-accelerated workflows — asset maintenance for the pattern library, exactly like pilots keeping manual landings current. For a financially material forecast, that means a human still builds it by hand once a quarter, not for output, but to keep the pattern library alive.
Doesn't this sacrifice productivity?
A small, budgeted maintenance cost against the day the AI is confidently wrong in a filing — the ROI case writes itself in one bad quarter avoided. The maintenance overhead is a fraction of one bad close cycle, one restatement, or one discovery proceeding.
Who owns the Judgment Ledger?
A named executive — the vCAIO seat in most mid-market organizations — who reports human capital next to token capital on the AI program's board report. Ownership without a name is a wellness poster; ownership with a name and a board slot is a governance discipline.
Ready to act on this?
Start with the practice most relevant to your estate, or reach out directly for a senior-architect conversation.
📧 contact@epcgroup.net · 📞 (888) 381-9725
Multiple models. One truth.