The Adoption Architecture: Why the Companies That Got It Right Kept Humans in the Loop

Executive Summary

• The companies with the highest AI adoption rates didn’t win with better messaging — they won with better architecture. Citi (70% adoption, 182,000 employees), IKEA (EUR 1.3B new revenue from reskilled workers), and Colgate (3,000-5,000 employee-built AI tools) share one design decision: workers remained meaningfully in the loop during and after deployment. None of them automated workers out of consequential decisions.
• The fear driving resistance is rational, not irrational. 65% of employees aren’t worried about machines replacing them — they’re worried about skilled colleagues with AI outcompeting them (Mercer, n=4,500, 2026). The intervention that works is not reassurance. It is ensuring every worker remains a necessary part of the output.
• Human-in-the-loop deployment is not a compliance concession — it is an adoption strategy. When workers must review, approve, or override AI outputs as part of their job, they get continuous low-stakes practice with AI. That practice produces trust. Trust produces the 2.6x usage consistency that separates organizations capturing AI value from those running expensive pilots that plateau.
• The alternative — full automation, fast rollout, no human oversight — produces performative adoption. AI usage climbs (up 13% in recent surveys), confidence falls (down 18%), and workers use tools on low-stakes tasks while protecting the work that matters. The organization bought licenses. It did not become productive.
• The workflow redesign question and the anxiety question are the same question. Companies that answer “which tasks stay human?” with specificity eliminate the fear. Companies that leave that question unanswered create it.

The Misdiagnosis

Most AI adoption programs treat the adoption problem and the anxiety problem as separate workstreams. Communications handles the fear. IT handles the rollout. Change management handles the training. The result is a well-messaged, well-rolled-out program that stalls at 30-40% active usage — roughly where industry averages sit — because the underlying design question was never answered.

The design question is: after this AI deployment, what is the human’s job?

If that question has no specific answer, workers answer it themselves. Their answer is usually pessimistic. AI job loss fear jumped from 28% to 40% in two years (Mercer, n=4,500, 2026). 73% of employees are staying in their current roles out of fear rather than loyalty — what Mercer calls “job hugging.” These workers are present, compliant on metrics, and quietly protecting the work that makes them irreplaceable. They are not using AI on anything consequential.

The misdiagnosis is treating this as a communications problem. The correct diagnosis is that it is an architectural problem. Workers are behaving rationally given an ambiguous deployment. The fix is removing the ambiguity by design.

What the Successful Cases Actually Did

Colgate: Training Before Access

Colgate deployed AI to 14,000 non-plant workers with a single sequencing decision that most companies reverse: mandatory training before tool access. The CEO participated visibly. The signal was not “AI is coming and we support you” — the signal was “we are investing in your capability before we change your work.”

The result was not compliance. It was ownership: 3,000-5,000 workers built their own AI assistants from the bottom up, approximately 10% of which were deployed across entire business lines. Workers who understood the tool before it arrived at their desk did not experience it as a threat to their judgment. They experienced it as an extension of it.

The sequence matters more than the content of the training. Training after deployment communicates: here is a tool that is already changing your work, now catch up. Training before deployment communicates: here is a capability we are giving you, now decide how to use it. These produce different psychological relationships to the same technology.

Citi: Peer Architecture at Scale

Citi achieved 70% adoption across 182,000 employees in approximately two years. No mandate. No compliance requirement. The mechanism was 4,000 volunteer AI Accelerators — employees who self-selected into a visible, recognized role as peer demonstrators.

The design insight is that peer credibility outperforms executive credibility by approximately 20% on adoption metrics. A colleague saying “this made my job better and here is how” addresses the real anxiety — not the abstract fear of AI, but the specific fear of being left behind while peers advance. When adoption spreads peer-to-peer rather than top-down, it simultaneously deploys the tool and answers the competitive anxiety. The worker watching a colleague demonstrate AI is learning that the path forward includes them, not around them.

The badge and recognition system Citi used — visible without salary impact — gave Accelerators identity around AI proficiency rather than around job title. This is a subtle but important design decision: it signals that AI capability is a new dimension of professional standing, not a replacement for existing standing.

IKEA: The Economic Argument for Reskilling

IKEA’s case is cited frequently as a reskilling success. It is more usefully understood as an economic argument made credible by specificity. When AI handled routine customer service queries, IKEA did not tell 8,500 call center workers that their jobs were safe. It told them their jobs were changing — to interior design consultants, complaint resolution specialists, and high-value service roles that AI cannot perform. EUR 1.3B in new revenue was generated by those workers in their new roles. Voluntary turnover dropped 20%. 70% reported higher work satisfaction.

The message that worked was not “we won’t let AI replace you.” It was “here is the specific new role, here is what it pays, here is the training that gets you there, and here is the revenue it will generate for the company.” Vague reassurance fails because workers recognize it as vague. Specific economic arguments about new roles succeed because they answer the actual question: what is my job after this?

The Professional Services Firm: Fixing the Incentive Structure

The least-cited case in this literature is arguably the most instructive for mid-market companies. A 2,200-person professional services firm (HBR, November 2025) deployed AI and watched individual productivity jump 30-40% while organizational productivity remained flat. The reason was not resistance. It was rational self-protection: workers billing by the hour were hiding AI use because efficiency threatened their billable hours. The adoption program had not touched the compensation structure.

The fix was architectural: restructured compensation to 80% base plus 40% performance incentive, expanded job grades from 6 to 14, added AI proficiency to competency models. Organizational productivity rose 22%. The company reduced prices 10%, generating a 20% sales increase. Labor cost rose only 5%. Profitability increased 3%.

The anxiety in this firm was not about job loss. It was about income loss from a compensation model that punished efficiency. The correct diagnosis changed the intervention entirely. Companies running adoption programs without auditing their incentive structures are solving the wrong problem.

The Human-in-the-Loop Design Pattern

Each of these cases embeds a common architectural decision that is rarely described as such: workers remain meaningfully in consequential decisions. Colgate workers built the tools themselves. Citi Accelerators demonstrated and guided adoption. IKEA workers moved into higher-judgment roles. The professional services firm restructured recognition around AI-enhanced performance rather than AI-replaced performance.

This is the human-in-the-loop pattern applied as an adoption strategy rather than a governance requirement.

In agentic AI deployment, human-in-the-loop (HITL) means a human must approve an AI action before it executes. In production, this is often framed as a safety control — a compliance cost to be reduced over time as the system proves reliable. That framing misses what HITL does to the human.

Workers who review and approve AI outputs as part of their job are doing something specific: they are exercising judgment on AI-generated work. This does two things simultaneously. First, it creates continuous practice — the repetition that produces the trust that produces the 2.6x usage consistency. Second, it answers the design question. The worker’s job, post-deployment, is to be the judgment layer above the AI output. That is a specific, skilled, non-automatable role. The anxiety has a direct answer embedded in the architecture.

Companies that deploy AI with strong HITL requirements for compliance reasons — financial services, healthcare, legal — accidentally get better adoption outcomes. Their workers know exactly what their post-AI job is: reviewing, approving, and overriding the system. The governance requirement did the change management work.

Companies that skip HITL to move faster leave the design question unanswered. Workers answer it themselves, pessimistically, and protect their work accordingly.

The Workflow Redesign Connection

The implication for workflow identification is direct. The question is not “which tasks can AI automate?” That question produces a list of threatened jobs. The question is “which tasks require human judgment to be trustworthy, accurate, or accountable?” That question produces a job architecture that includes AI and humans in defined roles.

Draup’s task decomposition framework is the most operationally useful tool in the corpus for this. Every task in a role falls into one of four categories:

The architectural decision that determines adoption outcomes is how many tasks land in the second category — AI-Primary with Human-Review. This is the HITL zone. It is where workers develop AI fluency through practice, where consequential errors get caught, and where the worker’s post-deployment role is most legible.

Most organizations deploying AI are trying to maximize the first category (AI-Autonomous) for efficiency and minimize the second for speed. The adoption data suggests this is the wrong optimization. The second category is where trust is built. The first category, deployed too fast, is where resistance entrenches.

The practical implication: when mapping workflows for AI deployment, deliberately size the AI-Primary/Human-Review category larger than efficiency alone would suggest. Build in the review step not because the AI needs it, but because the worker does.

Key Data Points

What This Means for Your Organization

The adoption rate your organization reports to the board is almost certainly measuring the wrong thing. License activation and weekly active users tell you how many people opened the tool. They do not tell you whether anyone used it on work that mattered.

The gap between those two numbers — and every organization has a gap — closes through architecture, not messaging. Three decisions determine whether your AI deployment produces real productivity or expensive compliance theater:

First: sequence training before access, not after. The Colgate model is reproducible at any scale. The budget required is not large. The signal it sends is specific: the organization is investing in worker capability, not substituting for it. Workers who understand a tool before it arrives at their desk own it differently than workers who are handed access and told to figure it out.

Second: deliberately preserve human review in the workflow, especially in the first 12 months. Resist the pressure to automate the review step out for efficiency. The review step is where workers develop the fluency that produces sustained adoption. It is also where errors get caught, accountability is maintained, and the worker’s post-deployment role is legible. The efficiency cost of the review step in year one is lower than the adoption cost of skipping it.

Third: audit your incentive structure before your communications plan. The professional services firm case is a warning. If workers are rewarded for time spent or volume produced, AI efficiency directly threatens their compensation. No amount of messaging resolves a compensation structure that punishes the behavior you are asking for. Fix the incentive before the rollout, not after the resistance appears.

If you are mapping workflows for AI deployment and want to stress-test the architecture against the adoption data — rather than discovering the misalignment after the pilot — that is a conversation worth having early: brandon@brandonsneider.com.

Sources

1. Mercer, “Global Talent Trends 2026,” n=4,500 executives, HR leaders, and employees. Large-sample longitudinal workforce study. Credibility: HIGH.

2. BCG / MIT Sloan, “AI at Work 2025: Momentum Builds, but Gaps Remain,” n=10,635 workers across 11 countries, June 2025. Credibility: MEDIUM-HIGH — large sample; BCG advisory conflict noted.

3. Deloitte TrustID and “State of AI Enterprise 2026,” n=3,235 leaders, 24 countries. Credibility: MEDIUM-HIGH — large consulting survey; self-reported.

4. McKinsey, “The State of AI 2025,” n=1,993 respondents, 105 nations, November 2025. Credibility: MEDIUM-HIGH — large consulting survey; self-reported.

5. HBR, “How a Professional Services Firm Used AI to Get Productivity Gains,” November 2025. 2,200-person firm case study, independently reported. Credibility: HIGH — independent journal.

6. Citi AI adoption program, multiple industry reports and conference presentations, 2024-2025. Credibility: MEDIUM-HIGH — company reported; corroborated across multiple sources.

7. IKEA reskilling program, multiple industry analyses, 2024-2025. EUR 1.3B revenue figure corroborated across independent sources. Credibility: MEDIUM-HIGH — company reported with independent corroboration.

8. Colgate-Palmolive AI deployment case, multiple sources including MIT Sloan and HBR coverage, 2025. Credibility: MEDIUM — practitioner account, not independently verified.

9. EY, “Human-Centered AI Transformation,” 2025. Credibility: MEDIUM — consulting survey; advisory conflict noted.

10. Draup, “Work Redesign Framework for AI Deployment,” 2026. Task decomposition methodology across enterprise deployments. Credibility: MEDIUM — vendor-adjacent, but methodology is consistent with independent academic task decomposition research.

11. NBER, Dillon et al., “Shifting Work Patterns with Generative AI,” RCT, 66 firms, n=7,137 knowledge workers, 2025. Zero change in task composition from tool access alone. Credibility: HIGH — randomized controlled trial.

Brandon Sneider | brandon@brandonsneider.com April 2026