A Portfolio Approach to AI Strategy

The organisations winning with AI right now share an unglamorous trait. They stopped treating each AI initiative as a project with its own business case, and started treating the whole AI investment as a portfolio.

The distinction sounds academic. It is not. A project mindset produces a list of use cases, each fighting for funding, each judged on its own return, each abandoned quickly if the early numbers disappoint. A portfolio mindset produces a small number of tiers with different risk profiles, different measurement timeframes, and different governance — and it gives the CFO a coherent picture of what the enterprise is spending, why, and what it expects in return.

The three-tier portfolio model we use with clients is straightforward. Foundations, productivity, transformation. Each tier is funded differently, measured differently, and governed differently.

The foundations tier is the shared capability the rest of the portfolio depends on. Data domains, the model gateway, the evaluation harness, the governance operating model, the observability stack, the reusable agent scaffolding. This tier does not have a use-case business case; it has a platform business case, indexed to the number and complexity of use cases it enables. It is funded centrally as core infrastructure. Its success is measured by the marginal cost and time-to-production for a new use case, not by the ROI of any specific application.

Under-investing in foundations is the most common strategic mistake we see. It looks like frugality on the current-year budget and produces a slow, expensive, ungovernable AI estate over the next three. Over-investing in foundations before you have use cases to justify them is the second most common mistake, and it produces a beautifully architected platform that no one uses. The right sizing is: foundations sufficient for the next two productivity tier use cases and the next transformation tier bet, refreshed annually.

The productivity tier is where most enterprises should concentrate spend in the first two years. These are the use cases with clear baselines and predictable payback — expert amplification, extraction pipelines, grounded knowledge assistance, engineering productivity tools, contact-centre assist. Each has a documented cost and cycle-time baseline. Each has a target uplift. Each has an owner in the affected business function.

The productivity tier is funded through the business units that own the workflows, at a marginal cost that reflects only their variable consumption of the platform. Success is measured in unit economics: cost per case, hours per report, error rate, cycle time. Governance is lightweight because the risk profile is contained. Executive attention is proportional: quarterly reviews, aggregated dashboards, exception-based escalation.

The value from this tier is real but unspectacular. A well-run productivity portfolio in a large enterprise typically produces 3 to 8 percent operating expense improvement in the addressed functions within 18 months, with a clear path to more as the pattern extends. That is a meaningful number. It is also the number that funds the transformation tier.

The transformation tier is where the interesting bets live. New products enabled by AI. Fundamentally reshaped customer journeys. Categories of work the enterprise did not do before because they were not economically viable. These are the initiatives with unclear baselines (because there is no current-state comparison), long payback horizons, and much higher variance.

The transformation tier needs a different governance shape. Fewer initiatives — three to five active bets is often enough for a large enterprise. Longer horizons — 18 to 36 months to a defensible outcome, not the 90-day sprint reviews the productivity tier can absorb. Executive sponsorship at the top of the organisation, because these bets often cross business unit boundaries and require capacity the natural owners will not free up on their own. And an explicit tolerance for failure: if all five of your transformation bets are on track, you did not pick ambitious enough bets.

The portfolio balance matters. A programme that is all productivity produces steady returns and no upside; the enterprise becomes more efficient and no more competitive. A programme that is all transformation produces a lot of expensive experiments and no reliable cash flow to fund them. The healthy split we see in enterprises that sustain this work over multiple years is roughly 15 to 20 percent of AI spend in foundations, 60 to 70 percent in productivity, and 15 to 25 percent in transformation. The exact numbers vary; the shape is remarkably consistent.

One financial discipline is worth naming explicitly. For every initiative in the portfolio, at every review, three numbers are on the same page: what we planned to invest, what we have invested to date, and what value has been realised (or what leading indicator says the value is on track). The absence of any one of these three is a governance red flag. Most AI programmes we see are missing the third — realised value — because no one designed the measurement before the build.

The portfolio also needs an off-ramp. Every quarter, the review should include a small number of initiatives being sunsetted — either because the productivity target was hit and the work is now business-as-usual, or because the bet did not play out and further investment would be throwing good money after bad. If nothing ever comes off the portfolio, the organisation has confused activity with progress. The best AI portfolios we see are as disciplined about stopping as they are about starting.

There is one more strategic move that separates enterprises that compound from enterprises that flatline. They treat the learning from every initiative — successful or not — as a portfolio asset in its own right. Documented patterns, evaluation sets, reusable components, engagement templates, governance artifacts. That library is the moat, and it grows faster than any single competitor's. Enterprises that treat AI as a series of one-off projects lose that compounding effect entirely; every new team starts from zero. The portfolio approach, done properly, ensures they do not.