It generated the report in 30 seconds. Then it took your best consultant an hour and fifty minutes to fix the hallucinations, check the stats, and rewrite the bits that were "almost right." Net saving: 10 minutes. Cost: a burnt-out senior employee who now reads everything twice and trusts nothing.
Everyone assumes AI clears the calendar. What actually happens, when a business introduces powerful generative tools without structural change management, is workload creep. A task that used to take five hours can now be generated in five minutes, so the baseline expectation of what a person should produce quietly triples, with no corresponding drop in how much careful thought the work still requires.
The trap has a name: prompt churn.
Instead of thinking clearly about the problem before typing (the audience, the structure, the evidence needed), the employee fires off a vague prompt, reads a plausible-but-off-target draft, tweaks a word, generates again, and repeats the cycle eight to fifteen times before something usable emerges. The planning stage that once forced clarity of thought gets skipped entirely. A task marketed as "five minutes with AI" routinely consumes sixty to ninety.
And the cost doesn't land evenly. A junior consultant who once delivered one carefully considered brief a week can now generate five variations in an afternoon: the machine saved them four hours. It cost the senior partner who has to read, evaluate, and sign off on all five variations a great deal more than that. This is the coordination tax: junior staff save time on drafting; senior staff absorb a new burden of filtering, verifying, and taking reputational responsibility for output they didn't produce. Output volume goes up. Cognitive load goes up. Team morale erodes quietly, underneath a surface of apparent productivity gains.
Any responsible AI workflow keeps a human in the loop: a person who verifies and approves output before it reaches a client, a regulator, or an investor. That's not a design flaw, it's the correct architectural decision. But it introduces a cost most AI adoption guides don't discuss honestly: the human cost of being the loop.
When someone builds a document from scratch, their brain is engaged the whole way through: they own the structure, the accuracy, the argument. When they're handed a 40-page AI-generated document and told to check it before it goes out, their role shifts from builder to auditor. They're no longer creating; they're scanning for errors in work they don't have the same instinctive ownership of. Auditing text for subtle hallucinations and quiet factual drift is measurably more exhausting than writing the original material, and it doesn't feel like the work most professionals signed up to do.
Over successive weeks, human biology reaches a predictable wall. We call it compliance bias: the point where a tired reviewer, worn down by output that's right nine times out of ten, starts to skim, assume, and approve.
That's the exact point where a fabricated case study or a compliance error slips through to a client. Not because the employee is careless. Because an unsustainable cognitive demand was placed on them without anyone measuring it, naming it, or designing the workflow to accommodate it. Verification fatigue compounds silently across weeks: careful review in week one or two, spot-checking by week four, compliance bias fully active by week six, and by week eight the first serious error has already reached a client or a regulator.
The two failure modes here are symmetrical. Banning AI tools outright just drives them underground: the "shadow AI" problem covered in Book 1 of this series. Mandating them company-wide, everywhere, all at once, creates exactly the operational chaos described above. The alternative sits between the two extremes: design AI integration as a series of small, rigorously measured, reversible experiments.
Isolate one bounded task, not "marketing," but "first-draft generation for routine complaint emails." Measure how long it takes manually first; that's your baseline. Run it for 21 days with a maximum of two users, in an enterprise-approved, data-secure environment, with no scope creep. Then calculate Total Task Time, not just the generation step, and compare it to the baseline. If the net saving is under 20% once verification and sign-off are counted, pause and reassess.
That last step matters more than any other in this guide. The most common measurement error in AI adoption is timing the generation step (thirty seconds) and reporting that as the productivity gain. It isn't. The metric that actually matters is the complete workflow: prompt formulation time, plus generation time, plus human verification and editing time, plus approval and sign-off time, compared honestly against how long the task took by hand. Applied honestly, this formula often reveals a genuine 20–40% saving on well-suited, bounded tasks, and a net cost on complex, judgement-heavy work where verification requirements are high. Both results are useful information. Neither should be embarrassing. The experiment is working exactly as intended.
None of this works unless people feel safe reporting that a trial isn't working. Most employees who've sat through the all-hands presentation and watched the software budget get approved are not naturally inclined to walk into their manager's office and say the tool is making things worse. They absorb the friction quietly instead, and the adoption appears to succeed while failing in the background. The fix is a specific, stated commitment from leadership: reporting a failed experiment is a positive contribution, not a professional risk, demonstrated the first time a negative result comes back, by thanking the person, documenting it, and pausing the workflow rather than defending the investment.
The question every business owner needs to answer honestly, before the next tool purchase, isn't how much of this can we automate. It's why are we automating it, and what happens to the people whose work it was? A business that strips out human judgement in pursuit of a cheaper bottom line becomes commoditised the moment a competitor can license the same model at the same price. That's a race to the bottom a small business cannot win against a platform with ten thousand servers.
Think of AI infrastructure as a cognitive bicycle, not a factory replacement. A bicycle doesn't replace human effort. It multiplies it. The rider still decides the route, reads the terrain, and makes the calls a machine can't.
AI is genuinely fast, genuinely consistent at pattern recognition, and genuinely tireless at rote compilation. It is also genuinely unreliable at contextual judgement and genuinely prone to confident hallucination. Your team is the opposite on every dimension. The augmentation model (AI handles structural friction and first-draft compilation, humans provide judgement, voice, and domain authority) isn't a compromise. It's the architecture that plays to the real strengths of both.
That only holds if the human side stays genuinely capable. A verifier who doesn't understand the domain they're checking isn't a human in the loop. They're a rubber stamp with a name. Which is why the book's guidance for protecting junior talent is blunt: keep at least 70% of a junior employee's output originating from their own drafting, even with AI assistance, and track that ratio quarterly, because it drifts toward AI-first without deliberate governance.
The book closes with a 5-minute, self-scored diagnostic, the Human-Centric AI Impact Checklist, across five areas: output creep, fatigue auditing, whether new workflows go through the mini-trial protocol, whether the 70/30 rule is actively protected, and whether your team has a genuinely safe channel to report that a workflow isn't working. Score under 4 out of 10 and you're in what the book calls the Burnout Danger Zone: pause any company-wide AI mandate and run one honest, bounded trial before anything else. Score 8 or above and you've built something durable: technology that powers your people without consuming their cognitive capacity.
If your team has had its own "approve without reading" moment, that's exactly the kind of story shaping what goes into Book 7.