Six operating habits the high performers shared. Workflow first. Escalation default. Hidden costs in the Gantt chart. Gatekeepers triaged early. Executives on a weekly cadence. Failure as precursor.
77%
of implementation pain is non-technical. Change management, data quality, and process redesign — together — dominate the difficulty curve.
71%
median productivity gain when AI runs the routine path and humans handle exceptions. Escalation beats both full autonomy and approval-gating.
61%
of successful deployments had at least one prior failed attempt. Architect for second-attempt learning; the first stumble is the start, not the end.
77% of the pain is not technical
The single most important number in the playbook for project planning. 77% of implementation challenges came from three non-technical sources: change management, data quality, and process redesign. Engineering and licensing got the budget; these three stole the schedule.
Non-tech cost
What it looks like
Goes in the plan as
Change management
Role redefinition, retraining, performance-system rewrites, internal comms.
Engineering line items, NOT "data team will handle."
Process redesign
Mapping the as-is workflow, defining the to-be, validating with end users, rewriting SOPs.
Discovery sprint + iterative redesign sprints.
If those three lines are missing from your project plan, they're hiding in your delays. Put them in the Gantt chart, give them named owners, and run them like engineering work.
"Once the model is integrated, the rest is rollout." The playbook's data says the opposite — the model integration is roughly the easy 23%; the rollout is where 77% of the pain lives.
Redesign the workflow, don't bolt on the tool
The largest separator between high performers and everyone else: how aggressively they redesigned the workflow around what AI made cheap. The McKinsey signal embedded in the playbook is stark — 55% of high performers fundamentally redesigned workflows as part of their AI efforts, versus only 20% of others.
Approach
What it produces
Outcome
Tool bolt-on
"We added AI to step 3 of the existing process."
Captures 10-20% of the available value. Most low performers stop here.
Workflow redesign
"We rebuilt the underwriting pipeline around AI-first triage. The whole shape of the work changed."
Captures the productivity multiples seen in successful cases. 55% of high performers.
"We added Copilot" is not the unit of work. "We redesigned the underwriting pipeline around AI-first triage" is. If the post-AI process diagram looks the same as the pre-AI one with a sparkle icon added, you're bolting on, not redesigning.
For PMs, this is the single highest-leverage decision you make on an AI project. Pick one workflow you'll redesign — not one tool you'll deploy.
Default to escalation-based operating models
How the AI-human boundary is drawn turns out to matter a lot. The playbook compared three operating models and found one that consistently outperformed the others.
Operating model
How it works
Productivity result
Fully autonomous
AI runs the path end-to-end without human checkpoints.
Underperforms — risk surfaces too late.
Approval-based
AI proposes, human approves at every step.
Underperforms — the human becomes the bottleneck.
Escalation-based ★
AI handles routine cases entirely; humans intervene only on exceptions.
71% median productivity gain — the winner.
For most workflows, default to escalation-based. The AI runs the routine path; the human is the exception handler. Reserve approval-gating for safety-critical decisions where the cost of being wrong is genuinely high.
When in your next AI design review someone says "but we should add human approval at each step," push back. Approval-gating sounds safer but in the playbook's data it consistently underperformed.
Triage the gatekeepers — legal, HR, compliance
One of the most counterintuitive findings in the playbook: resistance to AI deployment overwhelmingly came from gatekeeper functions — legal, HR, compliance — and not from end users. End users adopted; gatekeepers vetoed.
Function
Common concern
Engagement strategy
Legal
IP, liability, contract terms, data sharing.
Early review of vendor terms; standing office hours.
HR
Workforce impact, role redefinition, performance criteria.
Co-design the role/process changes from day one.
Compliance / Risk
Audit trails, regulatory exposure, model governance.
Surface eval rubrics + audit logs upfront, not at gate review.
A pilot that the end users love can still die at the legal sign-off. Build parallel adoption tracks for each gatekeeper function, with named owners and standing office-hours cadences. End-user enthusiasm doesn't unblock a compliance veto.
"If the users want it, the rollout will succeed." User adoption is necessary but not sufficient. The veto path runs through legal/HR/compliance, and they're harder to win than users.
Operational executive sponsorship
"Executive sponsor" is one of the most overloaded terms in enterprise software. The playbook found a sharp distinction between strategic and operational sponsorship — and only the operational kind correlated with successful outcomes.
Sponsorship style
What it looks like
Outcome
Strategic approval
Signs the kickoff deck. Reappears at the QBR. Hears about blockers second-hand.
If you can't get a weekly standing 30-min with the executive sponsor, you don't have an executive sponsor — you have a name on a slide. Don't start the project until you have it.
When recruiting an exec sponsor, negotiate the cadence first, not the title. "Will you trade me a standing weekly for the next quarter?" is a more useful question than "will you sponsor this?"
Failure as precursor — 61%
One of the most morale-relevant findings: 61% of successful deployments had at least one prior failed attempt. The companies that ultimately succeeded treated failure as structured learning rather than terminal outcome. The first attempt is more often the start of the engagement than the end of it.
After first attempt fails…
Low performers do
High performers do
Initial response
Treat as proof "AI doesn't work here."
Treat as discovery of which specific assumption was wrong.
Retro depth
Surface-level postmortem.
Structured retro mapping failure to layer (workflow / model / data / process / change).
Second attempt
Often never tried.
Reframed around the specific finding from the first retro.
Architect the first attempt with the second attempt in mind. Capture failure traces in retros structured enough to reuse. If your retro template doesn't separate "the model was wrong" from "the workflow was wrong" from "the change management was wrong," you'll lose the most useful insight from each attempt.
When pitching the first AI attempt internally, name the second attempt explicitly. "We expect to learn from this one and reshape; here are the three hypotheses we want to test" lands very differently from "this is the AI project."
Quiz — Applied
1. What percentage of AI implementation challenges does the playbook attribute to non-technical sources?
77% of implementation challenges came from change management, data quality, and process redesign — not from technology.
2. By the McKinsey figures cited in the playbook, what fraction of high performers redesigned workflows around AI vs. other companies?
55% of high performers redesigned workflows fundamentally; only 20% of others did. The largest single separator across performance tiers.
3. Among fully autonomous, approval-based, and escalation-based operating models, which produced the highest median productivity gain?
Escalation-based won. Approval-based makes the human the bottleneck; fully autonomous surfaces risk too late.
4. Resistance to AI deployments overwhelmingly came from which group, per the playbook?
A counterintuitive finding. End users adopted; gatekeepers vetoed. Build parallel adoption tracks for each gatekeeper function.
5. What distinguishes "operational" executive sponsorship from "strategic approval" sponsorship in the playbook?
Operational sponsorship is a recurring cadence + blocker list + OKR alignment. Strategic sponsorship is a signature on a slide. Only the first predicted success.
6. What percentage of successful deployments in the dataset had at least one prior failed attempt?
61% of successful deployments followed at least one failed attempt. Failure-as-precursor is the norm, not the exception.
7. Your team's AI rollout is stalled. The end users love the tool but the rollout hasn't progressed in six weeks. Based on the playbook, where do you look first?
A stalled rollout where end users are enthusiastic almost always points at a gatekeeper veto. End-user adoption is necessary but doesn't unblock legal/HR/compliance.
0 / 7 correct
Flashcards — Applied
02 · Applied
77% non-technical pain
tap to reveal →
77% of implementation challenges came from change management, data quality, and process redesign — not technology. Put them in the Gantt chart with named owners.
← tap to flip back
02 · Applied
Change management cost
tap to reveal →
Role redefinition, retraining, performance-system rewrites, internal comms. One of the three non-technical cost lines that eat 77% of project difficulty.
← tap to flip back
02 · Applied
Data quality cost
tap to reveal →
Schema work, dedup, lineage tracing, consent / access review. Treat as engineering line items with named owners, not as "the data team will handle."
← tap to flip back
02 · Applied
Process redesign cost
tap to reveal →
Mapping the as-is workflow, defining to-be, validating with users, rewriting SOPs. Run as discovery + iterative redesign sprints, not as a memo.
← tap to flip back
02 · Applied
Workflow redesign vs bolt-on
tap to reveal →
Bolt-on captures 10-20% of value. Workflow redesign captures the productivity multiples. The largest single separator: 55% of high performers vs 20% of others.
← tap to flip back
02 · Applied
Escalation-based operating model
tap to reveal →
AI handles routine cases end-to-end; humans intervene only on exceptions. 71% median productivity gain — the highest among the three operating-model types studied.
← tap to flip back
02 · Applied
Approval-based operating model
tap to reveal →
AI proposes, human approves at every step. Sounds safe but the human becomes the bottleneck. Underperformed escalation-based in the playbook's data.
← tap to flip back
02 · Applied
Fully autonomous operating model
tap to reveal →
AI runs end-to-end without human checkpoints. Underperformed escalation-based because risk surfaces too late, after damage is done.
← tap to flip back
02 · Applied
Gatekeeper triage
tap to reveal →
Legal, HR, and compliance are the silent veto. Build parallel adoption tracks with named owners and standing office hours — end-user enthusiasm doesn't unblock them.
← tap to flip back
02 · Applied
Operational executive sponsorship
tap to reveal →
Weekly standing cadence with the project lead + explicit blocker removal + OKR alignment. The kind of sponsorship that correlated with success — not signing a kickoff deck.
← tap to flip back
02 · Applied
Failure-as-precursor (61%)
tap to reveal →
61% of successful deployments had at least one prior failed attempt. High performers treat failure as structured learning. Architect the first attempt with the second in mind.
← tap to flip back
02 · Applied03 · Advanced
Production-system framing
tap to reveal →
"This is a production system that uses AI" — not "this is an AI experiment." The reframe that distinguishes the 5% from the 95%, per the playbook.