← Insights
Briefing · Water & Utilities
Adoption Is a Workflow Problem, Not a Training Problem
How industrial and engineering teams move AI from a few enthusiasts to the whole crew, and why mandates and workshops don't get there.
The bottom line
Real AI adoption is a workflow-transfer problem, not a training-completion problem. Usage spreads when people repeatedly use AI inside a real job, with supervisor support, peer help, workflow fit, and metrics that matter. For engineering and operations teams, adoption has to show up in work orders, troubleshooting notes, shift handoffs, and maintenance planning, not in workshop attendance or license activation.
What actually drives adoption
1
Measure adoption as verified recurring workflow use, not training attendance.
The cleanest evidence comes from training-transfer research: training only matters when it transfers into job behavior. Attendance and license activation are the easy metrics and the wrong ones.
2
Leader modeling matters when it changes the operating system, not when it's just sponsorship language.
The organizations capturing value pair senior involvement with actual workflow redesign. A leader who uses the tool and rewires how the work runs moves the needle; a leader who only endorses it doesn't.
3
Champion networks work when they're peer-support infrastructure, not a fan club.
The credible cases embed the tool into the real process and give people someone to ask. Visible, fast peer users pull the rest along; a roster of enthusiasts with no support structure doesn't.
4
In asset-heavy settings, AI has to earn trust inside bounded operational loops.
The pattern across industrial cases: predictions become recommendations, paired with new ways of working, inside one bounded decision. Trust is earned in the loop, not asserted in a rollout memo.
5
Top-down mandates fail when they force visible activity but leave the problem, workflow, and assurance unresolved.
A mandate produces logins, not adoption. Without a real problem, a fitted workflow, and an assurance path, the activity is theater and the value leaks out.
What it means for a VP of Engineering
✓
Measure use where the work lives.
Track recurring use in work orders, troubleshooting notes, and shift handoffs, not training sign-ups. If it's not showing up in the work, it isn't adoption.
✓
Model it in the operating system, not the memo.
Redesign one workflow around it and use it yourself. Sponsorship language without a workflow change reads as optional.
✓
Build champions as peer support.
Make a few users fast and visible, and give the team someone to ask. That's the infrastructure that spreads use.
✓
Earn trust in one bounded loop first.
Prove it where a decision is measurable and the fail-safe is clear, then widen. Trust compounds; mandates don't.
The wild card to watch
The most-quoted AI failure stats, the "95% of pilots fail" kind, aren't reliable, and they're not the point. Adoption dies when AI stays a personal productivity habit and never becomes a supervised operating routine. Convert it into the routine and it sticks; leave it as a personal trick and it evaporates.
Claim ledger
20/20
Checked
source claims traced to primary or strongest reachable pages
0
Fabricated
no invented figures found
3
Corrected
qualified after source review
9
Demoted
useful signals kept out of the headline
ConfirmedBlume et al. (2010), "Transfer of Training" meta-analysis (89 studies): training matters only when it transfers into job behavior, driven by motivation, work environment, and measured transfer.doi.org
ConfirmedBrynjolfsson, Li & Raymond (2025), QJE: 5,172 support agents, about 15% more resolved issues per hour. A workflow-specific causal benchmark, not industrial.doi.org
ConfirmedMcKinsey, State of AI (2025): about one-third scaling, 39% report any enterprise EBIT impact; high performers pair workflow redesign with senior-leader ownership and role-modeling. Association, not causation.mckinsey.com
ConfirmedBCG, "The Widening AI Value Gap" (2025): 5% future-built, 35% scaling, 60% minimal material value; value tracks reinvestment and people and tech capability.bcg.com
ConfirmedBCG, "AI at Work" (2026): 11,749 respondents; 74% frontline regular use, 66% given limited or no guidance on saved time; strategy beats access. Self-reported.bcg.com
ConfirmedSykes, Venkatesh & Gosain (2009), MIS Quarterly: peer-support constructs improved new-system use across 87 employees. Small but peer-reviewed.doi.org
ConfirmedNIST (2026), critical-infrastructure AI RMF concept note: trustworthy IT/OT/ICS deployment needs lifecycle risk management and actionable trust requirements.nist.gov
ConfirmedRAND (2024), 65 practitioner interviews: wrong problem, workflow mismatch, data and infrastructure gaps, and technology-first hype are the failure mechanisms.rand.org
Corrected"One-off workshops fail": corrected. Targeted training raised LLM adoption from 26% to 41% in one experiment, so the safe claim is that attendance alone is weak evidence, not that workshops never work.arxiv.org
CorrectedAutomated bug-assignment industrial case (auto-assigned 30% of reports at 75% accuracy, about 21% faster resolution): a strong champion-led case, but an under-review preprint, so held as example, not settled proof.arxiv.org
CorrectedGlobal AI trust-and-attitudes survey (48,340 people, 47 countries): corrected to context on trust, literacy, and governance, not evidence of project-failure causality.doi.org
Demoted"AI superfans" champion-network examples (media-reported): current and concrete, but not primary-audited. Directional only.wsj.com
DemotedIndustrial embedded-ML operations case (hourly predictions turned into crew recommendations at a large mine): a corporate/vendor case of embedded use, not independent ROI proof.vendor case
DemotedAgentic-AI industrial study (16 practitioners, 12 companies): names a capability-deployment verification gap; a small preprint.arxiv.org
DemotedGartner forecast: 40%-plus of agentic-AI projects canceled by end of 2027 on cost, unclear value, or weak risk controls. A forecast, not observed outcome data.gartner.com
DemotedCross-country adoption study (35 countries, under 3% to 25% adoption, no detectable task-restructuring yet): a preprint.arxiv.org
DemotedVendor-commissioned value survey: nearly 40% of AI time savings lost to rework, only 14% consistently clear positive outcomes. Directional, vendor-commissioned.workday.com
DemotedEnterprise "2x mandate" coding study (802 developers, 196,212 PRs, about 2.09x throughput): the gain came through accumulated use and review-process redesign, not a memo; a non-randomized preprint in a favorable software setting.arxiv.org
DemotedSmall qualitative Copilot study (10 users, 0 of 10 used formal training as their primary learning): peer exchange and trial-and-error dominated. A small preprint.arxiv.org
Demoted"95%"-style universal AI failure rates (MIT NANDA-style): repeated as hard truth despite preliminary, self-reported, or secondary sourcing. Not asserted here.contested
The adoption playbook
01
Pick three recurring workflows before buying more tooling.
Choose jobs that happen weekly or daily, have clear artifacts, and already have review points: work-order triage, engineering change review, troubleshooting notes, reliability-event summaries, shift-handoff prep, spare-parts search, inspection synthesis, or drawing and spec comparison. Exclude safety-critical control until assurance is mature.
02
Define the adoption unit as "role + workflow + frequency + proof."
Example: maintenance planners use the AI work-order assistant at least twice a week, attach the AI-generated summary to the work order, and the supervisor reviews the exception rate every Friday. That is checkable in a way "80% trained" never is.
03
Model the behavior at the leadership cadence.
Leaders and directors bring AI-assisted prep into staff meetings, incident reviews, capital-project risk reviews, and design reviews. Show the source trail, show the correction, show the decision. Leader modeling works when it demonstrates responsible use, not when it broadcasts enthusiasm.
04
Build a champion network around crews and disciplines, not org charts.
Pick trusted translators: a senior technician, a reliability engineer, a planner, a process engineer, a field supervisor, a controls engineer. Give them office hours, early access, escalation paths, and a weekly job: convert one peer workflow, capture friction, and retire one bad use case.
05
Track the adoption dashboard like an operating metric.
Minimum dashboard: weekly active workflows by role, repeat-use rate, supervisor-modeled use, champion touches, accepted and rejected outputs, exception and override rate, rework created, safety and compliance issues, and one physical or business outcome per workflow.
06
Use a 30-60-90 day adoption sprint, then either scale or stop.
Days 1-30: choose workflows, baseline, train in context, launch champions. Days 31-60: require evidence in the real workflow and review exceptions weekly. Days 61-90: scale only workflows with repeated use, manageable rework, and visible operating value. Stop or redesign the rest without shame.
Owner, briefing, proof
Owner
A named process owner per target workflow, accountable for the adoption unit (role + workflow + frequency + proof) and the weekly exception review, not a training coordinator.
Briefing
An adoption-unit decision brief per workflow, so a rollout is funded on verified recurring use tied to an operating decision, not on seats trained or licenses activated.
Proof
The chain from a named workflow to recurring supervised use to accepted outputs and one physical or business outcome. The adoption dashboard, not attendance.
Where to start
Start by defining the adoption unit and baseline for one workflow. If the gap is material, widen to a readiness look at the adoption operating system (dashboard, champion network, leader cadence, and 30-60-90 gates across the target workflows), and build the routine only when the organization wants it run.
Where the evidence stops
Many AI-champion and frontline-adoption figures here are company- or media-reported and self-reported. Leader modeling is associated with value, not proven to cause it. And workshops aren't useless; attendance just isn't enough on its own. Treat the vendor numbers as directional, not settled.
What would change this conclusion
- A rigorous causal study showing leader modeling or champion networks drive durable value on their own.
- Reliable, industrial-specific adoption and failure rates from a verified primary source.
- A controlled industrial study showing which adoption metric best predicts downtime, rework, safety events, or cycle time.
- A replicated study showing workshop attendance or seat activation alone predicts durable use, which would contradict the thesis.
- A large engineering or industrial company publishes audited evidence that champion networks, leader modeling, or adoption dashboards changed operating outcomes.