Overview

Siddharth Rao spent four years as an AI researcher before joining a large retail bank as their first "Head of AI Innovation." His mandate was to build something new. What he found was that the organisation's existing operating model was perfectly designed to kill anything new. Every experiment required a full business case. Every proof of concept needed legal sign-off before it could touch customer data. Every idea that got traction was immediately absorbed into a BAU (business as usual) team, where the pace of normal operations consumed it. "The organisation wasn't resistant to innovation," he told me. "It was just optimised for something else."

Internal innovation structures are the organisational designs that create protected space for AI experimentation to exist alongside - without being consumed by - normal business operations. Getting this design right is what separates companies that run interesting pilots from companies that reach genuine transformation.

The Organisational Immune System

Every organisation has an immune system: a set of processes, norms, and incentive structures that identify and neutralise things that don't fit the established pattern. Financial controls require ROI justification before investment. Legal review processes require known precedent before approval. Resource allocation processes favour proven activities over experimental ones.

These immune responses are not dysfunctional - they protect the organisation from real risks. But they are calibrated for operational reliability, not for innovation. An immune system that is strong enough to protect a $5 billion business from operational risk is almost certainly strong enough to kill a $500,000 AI experiment that does not yet have a proven ROI.

The structural solution is to create protected zones where the immune response is deliberately modified. Not eliminated - the controls exist for good reasons - but adapted to the different risk profile and timeline of exploratory work.

Innovation Labs and Skunkworks Teams

The most common protected structure is an innovation lab: a team that operates with more autonomy, faster decision cycles, and different success metrics than the rest of the organisation.

The term "skunkworks" comes from Lockheed's advanced projects division, which operated at arm's length from the main organisation to allow it to develop experimental aircraft without being constrained by standard procurement, budgeting, and management processes. The concept applies directly to enterprise AI innovation.

An effective innovation lab has six characteristics.

Physical or organisational separation. The lab team does not sit inside a business unit where their work will be constantly interrupted by BAU priorities. Whether this means a separate physical space, a distinct reporting line, or simply a protected allocation of time, the separation must be real and enforced.

Simplified governance for exploration. Experiments should be able to start quickly. The governance burden that applies to a $2 million production deployment is inappropriate for a $20,000 eight-week experiment. Build a lightweight governance track for exploratory work - faster approvals, simplified legal review for sandboxed environments, pre-approved data access for synthetic or anonymised datasets.

Talent who want to explore. Some people thrive in stable, optimised environments. Others are energised by ambiguity and the freedom to try things that might fail. The latter group should staff the lab. Placing people in a lab who are temperamentally suited to operational roles will produce an expensive team that is uncomfortable with failure and over-invest in polishing things that are still in exploration.

Funding protected from quarterly performance cycles. If the lab's budget is subject to quarterly review against standard ROI metrics, it will be defunded during the first difficult quarter. Innovation investment needs to be ring-fenced at the annual budgeting level, with a multi-year commitment that allows experiments to mature.

Different success metrics. Operational teams are measured on efficiency, reliability, and outcomes. Lab teams should be measured on learning: hypotheses tested, experiments completed, discoveries documented, and capabilities developed. A lab that ran ten experiments, seven of which failed, and produced three insights that informed later production deployments is performing well. Measuring it against the operational team's KPIs would mark it as a failure.

Clear transition pathways to production. A lab that only explores never delivers value. The most important design challenge is creating a structured pathway by which successful experiments graduate from lab to production - with appropriate governance re-engaged as the work matures. Without this pathway, labs become expensive toy departments.

Protecting Experiments from Operational Pressure

Siddharth's experience of good ideas being absorbed by BAU teams is a structural failure: no mechanism existed to protect exploration-phase work from being conscripted into operations before it was ready. The result is half-baked experiments running in production, or experiments abandoned because they could not survive the accountability regime of an operational environment.

Protection requires explicit policy. A rule - enforced by leadership, not just stated - that lab-phase work cannot be pulled into production teams before a defined graduation threshold is reached. The threshold might be: a successful pilot with at least two business units, demonstrated performance on agreed metrics, and a documented business case prepared by the receiving operational team. Until those conditions are met, the work stays in the lab.

This protection also runs the other direction. Lab team members should be protected from being seconded to operational projects during crunch periods. Every time this happens - and every organisation does it - the implicit message to the lab team is that exploration is discretionary and operations are mandatory. Eventually, lab teams stop expecting to be protected, and they start hedging against being absorbed.

Innovation Metrics: Measuring the Right Things

The cardinal error in innovation measurement is using operational metrics. "What is the ROI of the lab?" is the wrong question for the first three years of a new lab. The right questions are: how many experiments have been completed? How many generated actionable insights? How many have graduated to production? How has the lab's ability to run experiments improved over time?

A more sophisticated metric is the experiment cycle time: how long does it take from an approved idea to a decision on whether to proceed? Shorter cycle times mean faster learning. A lab that runs a two-week experiment cycle learns twice as fast as one that runs a four-week cycle - over a year, that compounds significantly.

After the first two or three years, when some experiments have reached production, you can begin adding value metrics: business outcomes from graduated projects, cost avoidance from experiments that revealed failing approaches before significant investment, and talent developed through lab experience who are now leading production AI work. These are the real returns on innovation investment - but they take time to realise.

Culture: The Intangible That Makes Structures Work

Structures create the conditions for innovation. Culture determines whether those conditions are actually used. A lab with perfect structural design but a culture that punishes failure will produce cautious experiments that are safe rather than interesting.

The cultural element that matters most is leadership's demonstrated attitude toward failure. If executives publicly celebrate experiments that failed but generated learning - "we tried X, it didn't work, and here's what we learned that we're using in Y" - that signal is powerful. If the only stories told are about successful outcomes, people read the implicit message: failure is not acceptable here.

Siddharth made a practice of presenting one "learning from failure" story at his quarterly all-hands. Not apologetically. As a feature of a healthy innovation programme. It took eighteen months before people started doing the same thing in their own team meetings, without being prompted. Culture change is slow. Structural change is faster. Both are necessary.

Key Takeaways

• Organisations are optimised for operations, not innovation. The governance, accountability, and incentive structures that protect operational performance will naturally suppress experimental work unless you explicitly create protected zones.
  - Effective labs have six characteristics: separation, simplified governance, exploration-oriented talent, ring-fenced funding, learning-based metrics, and clear graduation pathways. Missing any one of these significantly reduces the lab's chance of generating value.
  - Protection must be explicit and enforced from the top. Labs that can be conscripted for operational projects or absorbed by BAU teams before graduation will not function as innovation structures. Leadership must enforce the boundaries.
  - Measure experiments by learning rate, not ROI. Experiment cycle time, hypotheses tested, and insights documented are the right metrics for an exploration-phase lab. Add value metrics only after production graduates exist.
  - Transition pathways to production are the most critical design element. Labs without graduation pathways generate insights but not value. Define the threshold conditions for production transition before the lab starts, not after experiments succeed.
  - Culture change requires leaders to celebrate failure publicly. Structural protection creates the space; cultural permission determines whether teams actually use it. Leaders must actively model the behaviour they want to see.