Writing / / 10 min read

Lab Renaissance

The return of the small, day-to-day-decoupled team — Lab or Center of Excellence — as the organizational-design answer to AI adoption.

What is it?

The lab is coming back. Not the frontier kind that trains models, the organizational kind: a small team pulled out of a company’s daily business and its quarterly revenue, set up as an innovation nucleus and a testing space, and given a mandate the everyday work cannot hold. Whether it gets called a Lab, a Center of Excellence, or an Arbeitsgruppe is noise. The move underneath is always the same. Take a handful of people from different corners of the organization, decouple them from the work, and make them responsible for the one thing no department owns: turning what individuals are figuring out into something the whole company can use.

It is a renaissance because the form is not new, and because the last time we tried it, it mostly failed.

Why the lab went out of fashion

The isolated innovation lab of the mid-2010s is the thing the word still makes most executives flinch at, and it earned the flinch. Set off to the side to protect it from bureaucracy, the lab protected itself right out of relevance — what Kong calls “innovation theater: performative activities that simulate progress but lack tangible impact.” Bright prototypes, good press, and a core business that quietly rejected everything the island produced.

The failure was not the one you would guess. Campos, surveying why innovation efforts collapse, found that “many innovation endeavors fail despite funding availability, talented people, strong technology, and sincere intentions.” Money, talent, technology, good faith — all present, all insufficient. What was missing was alignment between the lab and the business it was meant to change. The lab died of distance.

Hold onto that. The renaissance turns on exactly this axis.

What changed: experimentation became structural

Deterministic software you can specify, procure, and roll out. Generative AI you cannot. Hofmann’s taxonomy of AI strategy puts the difference plainly: contemporary AI carries an “autonomy, learning, and inscrutability that distinguish [it] from previous generations of IT.” You do not deploy a skill; you acquire it by using it, and what it turns out to be good for is discovered rather than specified. Baily and colleagues make the same point from the evidence side, grounding the technology’s value in “field test evidence of productivity improvements from genAI in practical applications… including for writing, computer programming, and responding to call center inquiries.” The gains are real, and they are found by hand, in the work, one use at a time.

So experimentation is no longer a discipline problem to stamp out. It is the structural condition of the technology.

The diagnosis the lab answers

There is more AI experimentation inside a company than any inventory will show, and almost none of it adds up. Mollick states the trap flatly in Making AI Work:

“AI use that boosts individual performance does not naturally translate to improving organizational performance.”

A prompt that saves one analyst her afternoon makes her faster, and nobody else. The individual returns are visible everywhere and the organizational return is nowhere, because nothing carries a discovery from the person who made it to the people who could use it. You end up with better employees inside a company that is getting no better at its business.

The word for what a lab does to that situation is compound. A faster person is not a faster organization. The lab is where a prompt living in one head becomes a tested instrument, and a workflow someone improvised becomes a benchmark the next model has to beat — where individual gains stop leaking and start raising the floor for the next person. Capturing them is a job, and in most companies nobody has it.

WITHOUT A LAB gains leak WITH A LAB gains compound resets to zero each cycle the floor rises each cycle
Same individual effort, two fates. Without a lab the gains reset to zero each cycle; with one, each cycle raises the floor.

The bridge: Leadership, Lab, and Crowd

Mollick’s framework is the cleanest statement of where the lab sits. Three things have to be true at once. Leadership makes experimenting safe and points it somewhere. The crowd — “the employees who figure out how to use AI to help get their own work done” — does the discovering. And the lab turns what the crowd surfaces into shared instruments the organization can hold. Most companies have the first two half-built and the third missing entirely. A lab without a crowd is a silo; a crowd without a lab stays underground and unscaled. The lab is the bridge, and bridging is the whole job — not blue-sky invention, but operationalizing what is already being discovered on the floor.

Leadership makes it safe · points the way Crowd everyone using AI in their own work Lab codifies into shared instruments Tested instruments the whole organization can pick up permission + direction raw discoveries shared instruments evidence it works
Leadership, Lab, and Crowd. Permission and direction flow down, raw discoveries flow up, the lab turns them into instruments the organization can hold — and evidence returns to leadership.

Why a team, and not a mandate

The cheaper answer is to skip the team and tell everyone to use the tools. It does not work, and the number is brutal. Organizations that go this route, Mollick reports, find “the use of official AI chatbots maxes out at 20% or so of workers, and that reported productivity gains are small.” You cannot order people to acquire a skill, and you cannot mandate the sharing of something they gain by keeping.

Nor can you buy your way out. BCG’s AI Radar, surveying corporations that expect to roughly double AI spend as a share of revenue, lands on advice that has nothing to do with procurement: “Build personal fluency. Practice. Practice. Practice.” A lab is where an organization does that practicing on purpose, in one place, so the fluency accrues to the company and not only to whoever happened to practice.

Why now: the consensus turns against the island

The institutions have caught up to the diagnosis, and they caught up at the same moment. The World Economic Forum, in March 2026: “Companies have long talked about AI as a transformational force, but most activity stayed in sandboxes and innovation labs… The real transformation happens when AI embeds into the flow of work.” McKinsey’s QuantumBlack is blunter about the verdict on the old model: “For years, companies operated AI labs — isolated teams running pilots and proofs of concept. That model worked when AI was experimental. It doesn’t work anymore.”

Read quickly, that sounds like the end of the lab. It is the opposite. What died is the island — the lab defined by its distance from the work, the same distance that killed Campos’s well-funded efforts a decade earlier. What returns is the lab rebuilt on the one axis it used to get wrong: close enough to the daily business to feed it, protected enough to still think past this quarter. Some of the institutions are renaming the embedded version a “factory.” The word is theirs; the structure is the small decoupled team, brought back in from the cold.

INSIDE the daily work FAR away TOO CLOSE PROTECTED BAND TOO FAR the new lab lives here operations absorb it, just another delivery team the island that died last decade, a satellite nobody reads
The lab dies of distance. Too close and operations absorb it; too far and it drifts into the island that failed last decade. The renaissance rebuilds it in the protected band — close enough to feed the business, exempt enough to think past the quarter.

How labs dock

Where the lab sits is part of the design, not an afterthought to it. Two variables carry most of the weight.

Docking. Some labs sit directly under leadership in a chief-of-staff, sounding-board function — Google’s Center of Excellence held regular leadership meetings and reported in continuously. Others run largely autonomous and surface only at chosen moments. Neither is correct in the abstract; the right answer depends on what the organization needs the lab to change.

Distance. Too close and operations absorb the unit, and it quietly becomes another delivery team. Too far and it drifts back into the island nobody reads. Labs rarely die of technical causes. They die of position, of mandate — demos with no route into daily work — and of politics, when the pilot review arrives and nobody built the case for permanence before it was needed.

Why it matters

The interesting thing about the lab renaissance is that the answer to the most individualizing technology of the decade is an organizational one, and an old one we already failed at once. AI makes every person more capable in isolation, and that isolation is the whole problem; the response is not a better tool but a deliberately exempt team whose only product is the organization’s ability to learn from itself. You improve the daily business by building something the daily business is not allowed to touch — and you rebuild, on purpose, the structure that died of distance, this time close enough to matter. The question is no longer whether a company should have a lab. It is whether it can hold a team close to the work and exempt from it at once, which is the one thing the last decade proved is hard to do.

Sources