Lean AI innovation: a human-first framework for corporate deployments

Executive overview

Most AI projects fail not because the technology doesn't work, but because teams skip understanding how real users will interact with it. The bigger risks are human and organisational — not technical.

The antidote is a lean innovation funnel: charter the vision, surface the riskiest assumptions, test them cheaply before building, and use metered funding to stay accountable at each stage.

• Start with customer discovery, not code.
• Identify your "leap of faith" assumptions and test the important-but-uncertain ones first.
• Use structured governance so teams are empowered by evidence, not opinion.

Why traditional AI development wastes time

• Early NLP projects required months of "ground truth" collection — data that proved useless once real users arrived.
• The NatWest virtual agent experiment: launching to 50 real users per day exposed 34 bugs overnight. Traditional testing would have missed most of them.
• Gen AI makes building trivially cheap, which amplifies the temptation to skip discovery entirely.
• Less than 5% of people have used Gen AI; the people building it are a tiny fraction of the intended users.

The four-stage innovation funnel

1. Customer discovery — who is the customer and what problem do they actually have?
2. Solution validation — does a mock-up or Wizard-of-Oz simulation resonate? Would they pay?
3. POC — prove the technical and product approach works.
4. Scale validation — commit full resources: legal, go-to-market, support.

Each stage runs ~6 weeks before a go/no-go decision. Most resource commitment lands only in stage 4.

Identifying and testing leap-of-faith assumptions

• Map assumptions on two axes: how important vs. how certain.
• Test the ones that are high-importance and low-certainty first.
• Ignore the other 8–9 assumptions if you're already confident they'll hold — this cuts time-to-market by 80–90%.
• For code generation tools: developer adoption and job-threat perception mattered far more than whether the tool technically worked.
• An education startup built expensive room-spanning microphones; a cheap handheld pointer turned out to be what teachers actually used.

Low-fidelity validation techniques

• Paper mock-ups of interactions to test flow before any code is written.
• Wizard of Oz: a human operator behind a chatbot interface simulates the AI.
• Staged rollouts: throttle exposure to a small number of real interactions, observe, fix, repeat.
• Use simulations to test whether users would engage before committing engineering effort.

The funnel kill rates and portfolio math

• Plan for a 50% kill rate at each stage: 16 ideas → 8 → 4 → 2 → 1 scaled product.
• Kill rates are not failure — they're prioritisation. All 16 ideas may be viable; the funnel forces focus on the best bets given constrained resources.
• Having a modelled kill rate gives teams permission to recommend stopping without it reflecting badly on their work.
• Humane Pin and Rabbit R1 are cited as products that likely skipped early validation and paid the cost.

Metered funding

• Request resources only when evidence justifies the next stage — not upfront for the full programme.
• Pitch to CFOs: "I'll come back with data before I ask for more money."
• Forces teams not to skip steps to reach a predetermined conclusion.
• After customer discovery: a resonant customer segment is confirmed, but success is not guaranteed.
• After solution validation: willingness to pay is tested; only then do you scale up headcount.

Governance and senior stakeholder management

• Run "growth board" reviews where teams present what they tested, the data, and their recommendation.
• Data-led decisions replace opinion-led ones; senior sponsors must accept evidence that their idea didn't land.
• Cultural shift required: teams need confidence to say "the data says stop" — not find the one positive outlier.
• Finance teams respond well once they understand the process is disciplined and evidence-gated.
• Point to prior corporate failures as proof that skipping experiments costs far more than running them.

Three top insights for corporate AI executives

1. Technology lives inside human and organisational systems — the hardest problems are adoption, not engineering.
2. Map your assumptions, identify the riskiest ones, and test those first. Everything else can wait.
3. Build a business model for innovation: metered funding, explicit kill rates, and a plan for ideas not progressing.