How to Run an Innovation Lab: Build a School, Not a Showroom

Capgemini studied 340 organizations that had opened innovation labs and put the failure rate at up to 90% — the labs existed, held workshops, produced demos, and changed nothing about how the parent company made money. The study is from 2017, and I keep citing it anyway, because the failure mode it names hasn’t aged a day: companies build labs to look innovative, then discover there’s no mechanism for anything the lab learns to travel back into the organization.

I’ve been on both sides of this — inside a lab that got quietly reabsorbed, and later helping run innovation programs that survived. The difference wasn’t talent, budget, or beanbag density. It was that the failed lab was a showroom — a place executives brought clients to demonstrate the company’s relationship with the future — and the surviving programs were schools: places where a rotating set of people deliberately practiced a learnable craft (taking an idea from hunch to evidence, cheaply), inside a structure designed so the learning had somewhere to go.

That’s the whole thesis, so let me say it once, plainly: an innovation lab is a pedagogical institution or it is theater. The output that compounds isn’t the prototypes — most of those die, and should. It’s the people who cycle through and come out knowing, in their hands and not just from a book, how to de-risk an idea. Everything below — the structure, the cadence, the governance, the case studies — is in service of that.

First, walk the graveyard

Before designing anything, spend an hour with the closures, because they cluster into three named failure modes.

The mid-2010s lab boom produced a remarkable shutdown wave: the New York Times’ R&D Ventures (2013), Microsoft’s Silicon Valley research lab (2014), Turner’s Media Camp (2014), Disney’s research outpost (2016), Adecco’s Ignite Lab (2016) — a list VentureBeat compiled under the title “It’s time to ditch your innovation lab.” Nordstrom’s celebrated innovation lab — the one whose one-week sunglasses-app video every design-thinking workshop showed — was downsized in 2015 after four years, its staff reassigned, when retail margins tightened. Coca-Cola’s Founders platform, which paired seasoned entrepreneurs with company resources, shut down in 2016 after roughly three years. Even the most decorated public-sector example, Denmark’s MindLab — the world’s first government innovation lab, studied and copied globally for sixteen years — was closed in 2018 when its funding ministries pivoted to a digital-transformation task force.

Three patterns, one per story. Nordstrom: a lab funded as a discretionary showcase gets cut the moment the core business sneezes, because it never wired its output into anyone’s P&L. Coca-Cola Founders: importing entrepreneurs doesn’t transfer entrepreneurship — the capability lived in the visiting founders and left with them, which is precisely the pedagogy failure; nobody inside learned the craft. MindLab: sixteen years of admired work, still killed — because a lab whose sponsors change is a lab whose mandate must be re-earned continuously, and MindLab’s final director said as much on the way out. Prestige is not a moat. Absorption is.

The Xerox PARC story belongs here too, but as a corrective to the usual telling. PARC invented the GUI, the mouse, Ethernet, the laser printer, and object-oriented programming — and Xerox commercialized essentially one of them (the laser printer, which alone repaid the lab many times over). The popular lesson is “big companies can’t innovate.” The real lesson, as Tendayi Viki argued on PARC’s forty-seventh anniversary, is that invention without a transfer mechanism and a business model is a donation to your competitors. Xerox didn’t lack ideas; it lacked a bridge between the lab’s output and the company’s ability to sell anything that wasn’t a copier. Apple walked across the gap Xerox never built.

The structural evidence: separate, but connected at the top

There’s an actual quantitative answer to “where should the lab sit,” and it’s older and stronger than most people realize. O’Reilly and Tushman’s 2004 Harvard Business Review study followed 35 attempts at breakthrough innovation across structures. Projects run inside the existing functional organization succeeded about a quarter of the time. Projects run as unsupported cross-functional or skunk-ish teams with no senior linkage succeeded roughly never. Projects run in an ambidextrous structure — a separate unit with its own processes and culture, but integrated with the parent at the executive level, sharing a senior sponsor who owns both the core and the new — succeeded more than 90% of the time.

Read that distribution again, because it kills the two most common lab designs simultaneously. Embedding “innovation” inside business units fails because the core’s metrics, cadence, and antibodies grind exploratory work to dust. But full separation — the cool loft downtown, badge access, different laptops — fails just as reliably, because separation without executive integration is how you build a PARC: brilliant, beloved, and unabsorbable. The design that works is uncomfortable on purpose: the lab is protected from the core’s processes and accountable to the core’s leadership, through a sponsor senior enough to force the handoff conversation when something works.

Lockheed’s Skunk Works — the term everyone borrows and almost nobody reads about — was exactly this shape. Kelly Johnson’s team built the XP-80 jet fighter in 143 days in 1943 with a couple dozen engineers, and his famous 14 rules are mostly rules about connection under protection: the program manager gets near-total authority (protection) but reports practically to the top of the company and stays close to the customer (integration), with a small team, minimal reporting, and mutual trust between lab and buyer. People quote Skunk Works to justify secrecy and autonomy. Johnson’s actual system was autonomy plus a direct line to power and to the customer — the two wires most modern labs never install.

The lab as a school: the pedagogy that makes it real

Here’s the reframe that separates the labs that compound from the ones that decorate: stop asking “what will the lab produce?” and start asking “what will people who rotate through the lab be able to do afterward that they couldn’t do before?” That’s a curriculum question, and there’s a century of learning science to steal from.

Start with Kolb’s experiential learning cycle — concrete experience, reflective observation, abstract conceptualization, active experimentation — because it’s a clean diagnostic for broken labs. Most labs are stuck oscillating between two of the four stages. Workshop-heavy labs live in conceptualization (frameworks, canvases, sticky notes) and never reach concrete experience with real customers. Hackathon-heavy labs live in experience (build! demo! applause!) and never reach structured reflection, so the same mistakes get rebuilt annually with better UI. A functioning lab forces the full cycle on every project: run the experiment, then sit down and extract what it taught you, then update your model of the customer, then design the next experiment from the updated model. Donald Schön called the professional version of this “reflection-in-action,” and his point cuts: practitioners get better not from experience but from examined experience. The lab ritual that operationalizes it is cheap — a written learning memo after every experiment, arguing what was believed before, what the evidence showed, and what changed. If your lab produces demos but no written record of updated beliefs, it’s a hackathon venue with a nicer name.

Second: deliberate practice, with honest fine print. Ericsson’s classic finding is that expertise comes from practice that is designed, effortful, feedback-rich, and targeted at the edge of current ability — not from mere repetition. (The fine print: a 2019 Royal Society replication found practice explains less of expert performance than the strong version of the theory claimed. The concept survives; the mythology of “10,000 hours guarantees mastery” doesn’t.) What deliberate practice means for a lab is concrete: innovation is a set of trainable sub-skills — interviewing customers without leading them, writing falsifiable assumptions, designing the cheapest test that could kill an idea, reading weak signals in data — and the lab should drill them as skills, with coaching and repetition, not assume they emerge from enthusiasm. The best supporting data point I know is McKinsey’s venture-building research: organizations building their fourth-or-later new business succeed at roughly twice the rate of first-timers and generate about twelve times the fifth-year revenue. Repetition with feedback compounds. That’s not a business statistic; it’s a learning-curve statistic wearing a suit.

Third: intelligent failure, which is the lab’s grading rubric. Amy Edmondson’s Right Kind of Wrong distinguishes basic failures (preventable, in known territory — sloppy), complex failures (system breakdowns), and intelligent failures — in genuinely new territory, in pursuit of an opportunity, informed by a hypothesis, and no bigger than necessary to learn. Only the last kind deserves celebration, and the four criteria are a usable checklist: a lab experiment that fails big, without a written hypothesis, on a question someone in the building already knew the answer to, isn’t innovation — it’s an expensive basic failure with a lanyard. Psychological safety, Edmondson’s better-known concept, is the precondition: people drill honestly and report null results only where doing so is career-safe. The structural trick for making it safe comes from Alphabet’s X, below.

The operating cadence: how a week and a quarter actually run

Pedagogy needs a timetable. The one I’ve seen work stacks three loops.

The week-scale loop is the design sprint — Jake Knapp’s five-day cycle, built at Google in 2010: pick a critical question Monday, sketch competing answers, decide, prototype a facade by Thursday, put it in front of five real customers Friday. Its pedagogical value exceeds its product value: a sprint is a rep — the full Kolb cycle compressed into a week, with real customer contact as the non-negotiable finale. New lab members should run several as apprentices before leading one.

The project-scale loop is build-measure-learn with real stakes, which means every project carries a leap-of-assumptions map — the value/usability/feasibility/viability risks I covered in the discovery post — and attacks the riskiest assumption first. X, Alphabet’s moonshot factory, calls this #MonkeyFirst: if the plan is a monkey reciting Shakespeare on a pedestal, build the talking monkey, not the pedestal — because pedestal-building feels like progress while proving nothing. Astro Teller’s team pairs this with the mechanism everyone quotes and almost nobody copies: X pays bonuses to teams that kill their own projects on the evidence. Roughly 2% of X’s explorations survive, and Teller frames the kill rate as the system working — the cheap deaths are what make the survivors affordable. You don’t need Alphabet’s budget to copy the incentive: publicly celebrate and reward well-argued shutdown memos, and watch how quickly your lab stops running zombie projects.

The quarter-scale loop is the funding gate, and the right technology is Rita McGrath and Ian MacMillan’s discovery-driven planning (HBR, 1995) — the intellectual ancestor Steve Blank credits for the Lean Startup. Instead of funding an annual plan built on fictional projections, you fund to the next checkpoint: document every assumption the venture’s success requires, identify which are testable soonest and cheapest, release only enough money to test them, and re-decide at each gate. This is the governance answer to the Nordstrom problem — a lab funded through assumption-testing milestones can show its board a ledger of risks retired per dollar, which survives a budget review far better than a highlight reel of demos.

For portfolio conversations above the project level, a warning about the framework your executives will reach for: McKinsey’s Three Horizons. It’s fine as a vocabulary for balance and terrible as a scheduling tool — Blank’s 2019 HBR piece names the fatal flaw, which is that the horizons were read as time (H3 = distant future), while modern disruptors ship “horizon three” propositions on commodity technology in months. Hold the portfolio-balance idea; drop the assumption that transformative means slow, or your lab will incubate at a leisurely pace precisely the things competitors ship this year.

Stories to steal from, one mechanism each

Every famous lab gets cited for its aura. Steal mechanisms instead — one per story.

Bell Labs: engineered collisions. Mervin Kelly, the architect of Bell Labs’ great age, ran what he called an “institute of creative technology” on a few deliberate principles: put thinkers and doers under one long roof, make corridors so traversable that a physicist couldn’t avoid passing an engineer’s door, and structure the flow from research reservoir to systems engineering to device manufacture so ideas had a pipeline, not just a birthplace. The stealable part isn’t the campus; it’s the refusal to separate inventing from making. If your lab’s builders never sit with the people who’d have to operate the thing, you’re building PARC, not Bell Labs.

Adobe Kickbox: democratize the on-ramp. In 2013 Adobe started handing any employee who asked a red box containing a process guide and a prepaid card with $1,000 — no approval, no committee, no pitch. Around a thousand employees (roughly a tenth of the company at the time) took one; at least 23 ideas advanced to further investment, and Adobe open-sourced the whole kit. The mechanism: the scarcest resource in corporate innovation isn’t money, it’s permission, and Kickbox converts permission from a social negotiation into a physical object. It’s also a pedagogy play hiding in plain sight — the box is a self-paced curriculum in validating an idea, and a thousand people took the course.

UNICEF’s Kosovo lab: the lab as youth academy. UNICEF’s Innovations Lab Kosovo, running since 2010, built UPSHIFT — part workshop, part incubator — that has put nearly two hundred youth-led projects through a design-and-build cycle, with twenty becoming businesses and twenty-eight becoming civil-society organizations, and the model has since scaled to UNICEF offices in a half-dozen countries. It’s the purest existing example of the school thesis: the explicit product is capability in the participants, and the ventures are the exhaust. If a program in Pristina can run this loop on a UN budget, your enterprise lab can.

Haier: the anti-lab. The strongest challenge to the whole lab concept is Haier’s RenDanHeYi model — Zhang Ruimin dissolved the appliance giant’s middle management (about twelve thousand roles) and reorganized ~80,000 people into 4,000-plus self-managing microenterprises, each contracting with the others and facing users directly. No central lab, because the entire company is the lab. You probably can’t (and shouldn’t) copy it wholesale, but it makes the essential point as a limit case: a lab is a workaround for an organization that can’t yet innovate everywhere. The endgame of a great lab is to make itself less necessary by graduating enough alumni that the parent’s default way of working absorbs the craft. Notably, even Alphabet has been moving this direction — since 2025, X increasingly launches projects as independent companies rather than internal graduations, an admission that the transfer-back problem is hard even for the people with the best-funded lab on earth.

Put it to work: the first two quarters

If I were standing up a lab this quarter, the checklist is short and none of it involves interior design.

Before anything else, secure the ambidextrous wiring: one named executive sponsor who owns both a core P&L and the lab, and a written answer to “when something works, which business unit absorbs it, and what do they owe the lab during transfer?” If you can’t get that in writing, don’t start — you’re building the Nordstrom lab, and the shutdown clock is already running.

Staff by rotation, not by hiring. A small permanent spine (two or three people who own the craft, the coaching, and the rituals) plus rotating tours of six to twelve months from the business units. Rotation is the absorption mechanism — every returning alum is a graft of the lab’s methods into a host team. Hiring a permanent staff of external innovators is the Coca-Cola Founders design; the capability walks out the door when they do.

Install the three loops from day one: design sprints as the weekly rep, assumption-mapped projects run monkey-first as the unit of work, discovery-driven gates as the quarterly funding decision. Write learning memos after every experiment; grade failures against Edmondson’s four intelligent-failure criteria; pay real bonuses for evidence-backed kills.

Report learning velocity, not activity. Assumptions tested per quarter, cost per killed idea, cycle time from hypothesis to customer evidence, and — the number that saves you at budget time — alumni now applying the methods in the core. Demos impress; a ledger of retired risk survives a CFO.

And keep the founding question taped above the door, because every drift the graveyard documents — showroom, theater, trophy case — begins when someone stops asking it: who is learning what here, and where does the learning go? A lab that can answer that every quarter is a school with a pipeline. Everything else is beanbags.