The Human Bottleneck in AI
Billions in AI spend. Flat productivity. The problem was never the model.
This week's Playbook Kickoff covered a quiet but telling headline: AI still needs consultants to turn technology into real business results. That one line deserved a deeper look.
Every company today is “doing AI.” New tools. Custom agents. Copilots rolled out across teams. Budgets approved faster than ever.
Ask one simple question — “What has actually improved?” — and the answer is often unclear.
Technology is accelerating. Adoption is widespread. Productivity is flat.
This isn’t new. It has a name: the Productivity Paradox. In the current AI wave, it’s showing up again — in a more expensive form.
1. The last mile
Most organizations have successfully deployed AI. Very few have figured out how to live with it.
The gap between capability and business outcome is where value quietly disappears. Call it the execution gap. Call it the last mile. Either way, it’s a human problem, not a technical one.
Even the companies building AI have run into the same wall. OpenAI and Anthropic have increasingly partnered with consulting firms — McKinsey, BCG, Accenture — not to improve the models, but to help enterprises apply them in real workflows. The challenge is no longer access to AI. It’s knowing what to do with it.
2. Knowledge is easy. Desire is hard.
Any transformation follows a pattern: people need to understand the change, believe in it, then adopt it. AI handles the first part well. It trains at scale. It delivers personalized guidance instantly. It doesn’t get tired.
Then comes the harder part: getting people to care.
AI can explain what to do. It cannot answer why this matters to you in a way that feels real.
When companies try to force that layer through automation, people disengage. Change is emotional. Not informational. A large European bank ran precisely into this — AI surfaced high-value use cases across credit risk and customer interaction, but deciding what mattered and embedding it into daily operations required human consulting support. AI created options. Humans had to create direction.
3. The privacy–effectiveness tradeoff
Useful AI needs data — not surface-level data, but behavioral insight. How people work. Where they struggle. Where they revert to old habits. In theory, this enables real-time coaching and continuous improvement. In practice, it creates tension.
The more the system observes, the more people feel observed. Employees don’t experience it as optimization. They experience it as surveillance. This produces a quiet but compounding loop: more data improves AI guidance; more data erodes employee trust. Without trust, no change sticks.
Build the smartest system in the world. If people don’t trust it, they’ll route around it.
4. The middle manager isn’t going anywhere
There’s a popular narrative that AI will flatten organizations — that management layers disappear as decisions automate. It sounds efficient. It’s also incomplete.
Every transformation turns on one question: “What’s in it for me?” AI cannot answer that credibly. It has no context, no relationships, no trust capital. Middle managers operate exactly in that layer. They translate strategy into meaning, absorb resistance, reframe intent. During high-stakes transitions, people don’t want a dashboard. They want a person.
You can automate communication. You cannot automate credibility.
5. AI solves technical problems. Most companies have behavioral ones.
AI performs well when the problem is structured: automating workflows, optimizing repeatable processes, switching systems. In those environments, it thrives.
Most organizational challenges aren’t structured. They involve unspoken norms, internal politics, conflicting incentives. AI lacks human context — not data context. It cannot read the room, sense resistance, or navigate ambiguity. When companies treat cultural change like a technical upgrade, they get partial adoption.
At Merck, AI pilots delivered results in controlled environments but struggled to scale. The issue wasn’t the models. It was integration, alignment, and execution. Behavior doesn’t change just because tools do.
6. Solow paradox 2.0
In the 1980s, economist Robert Solow observed that computers were visible everywhere except in productivity statistics. Organizations hadn’t yet restructured themselves around the technology. Real gains took years — sometimes decades — to materialize. Not because the technology improved, but because the way people worked changed.
The same pattern is running now. Companies are investing heavily in AI tools, models, and capabilities. They are underinvesting in workflow redesign, behavior change, and adoption systems. Costs go up. Outcomes lag. More than half of CEOs report no significant financial benefit from AI so far.
So what actually works?
The answer isn’t more AI. It’s a different framing: human-led, AI-supported change. In practice, that means:
Start with business problems, not tools. Focus on a few high-impact use cases. Redesign workflows, not just tasks. Assign real ownership — not “innovation teams.” Measure outcomes, not activity.
And accept that the hardest part isn’t technical.
AI is not failing. It is doing exactly what it was built to do.
The real question is whether organizations are restructuring themselves to match it — or expecting technology to fix problems that were never technical to begin with.
As we automate knowledge, we may be overlooking the one thing that actually drives change: desire.
The line 'you can automate communication, you cannot automate credibility' is one of the better summaries of why AI implementations stall that I've seen. In large enterprise transformations, the last mile has always been the hardest part and AI doesn't change that, it just makes the gap more visible when organizations skip it