Many AI initiatives stall because nobody connected the use case to an outcome that the organisation is committed to delivering. The pilots work. But the business impact doesn't follow, and the pattern repeats with the next use case.

This is a prioritisation and sequencing problem - and fixing it requires honesty about what your organisation can execute, not just what it can build.

Most AI initiatives are approved on enthusiasm, not relevance

AI initiatives typically get evaluated on technical feasibility and team enthusiasm. Can we build it? Does the team want to? The question that rarely comes up at approval stage is whether the initiative connects to a strategic objective.

A team automates a manual process. It works. But the process isn't connected to a KPI the board tracks. Nobody owns the outcome. Nobody fights for it when resources get tight. The initiative dies in staging.

Here's a quick test. Pick any AI initiative in your current portfolio: which KPI does it move, and by how much? If the answer doesn't come quickly, the initiative was approved on interest rather than relevance.

Without prioritisation criteria, seniority decides

When there's no framework linking AI initiatives to outcomes, politics fills the vacuum. The initiative with the most senior sponsor gets funded first. The team with the most compelling business case gets the next slot. The result is a portfolio spread across a dozen use cases - each with enough budget to prototype, none with enough to reach production.

The alternative is to evaluate each initiative on two dimensions: impact (revenue, cost reduction, risk mitigation) and feasibility (data readiness, technological ecosystem, talent capacity, regulatory exposure, change effort). Fund the few that score high on both. Keep a short list of ideas that become viable once specific prerequisites are met.

Feasibility is where most portfolios fall apart. A high-impact initiative that requires well-managed data that doesn't exist yet or a digital operating model that hasn't been built isn't a bet. It's a wish. And what's feasible depends on three conditions most organisations haven't assessed honestly.

Fund the few initiatives that score high on both impact and feasibility.

Why governance, data and talent determine whether AI scales

These three conditions aren't independent - weakness in one limits what the other two can deliver. But most organisations either try to solve all three at once (and make no progress) or assume one will sort itself out (and discover too late that it didn't).

Governance. Not a gate to pass before starting. A design constraint that shapes which bets are viable and how they run. The practical question is proportionality: heavy governance on high-risk applications, lighter controls on internal productivity tools. Most organisations either over-govern everything (and slow the portfolio to a crawl) or under-govern everything (and create risk exposure they don't discover until it's public).

Data and technology foundations. Most AI failures are data pipeline problems, not model problems. Coverage gaps. Stale data. Unclear ownership. Missing lineage. If you're green-lighting AI initiatives without verifying data readiness first, you're writing cheques the data team can't cash.

Talent and organisation. AI doesn't scale through better algorithms. It scales when people know what to do with the outputs - and when the operating model supports new ways of working. This is where things get politically uncomfortable. Upskilling programmes get announced but not resourced. Multidisciplinary teams get mandated but not empowered to make decisions. Adoption plans get written but nobody owns the behavioural change. If you've seen agile transformations fail for organisational reasons, the pattern is familiar: the methodology worked fine. The organisation wasn't set up to absorb it.

How to structure an AI portfolio that scales

The three conditions above don't all need solving at once. But the order matters - and getting it wrong burns through goodwill faster than most organisations expect.

Quick wins first. Start with one to three initiatives where the data is structured and the owners are obvious. Adverse event report triage in pharma. Invoice matching in finance. Target outcomes measurable in a quarter. These buy two things: credibility with leadership and learnings that make larger bets easier to develop. They also surface the practical gaps - in data quality, governance or team readiness - that need closing before anything bigger can work.

Then strategic bets. Shape one to two larger initiatives with board-level relevance: AI-augmented drug development in life sciences, personalised customer engagement at scale in financial services, fleet-wide predictive maintenance across critical infrastructure. These need stronger governance, real change management and executive sponsorship with teeth - not just a name on the charter. Be honest about whether your operating model can support them. A strategic bet that has no framework to run will stall the same way the pilots did.

Then operating capability. This is where AI stops being a project portfolio and becomes part of how the organisation works. Productised data assets. Standard MLOps or LLMOps pipelines. Reusable patterns. Governance that adapts as the portfolio matures. Business ownership of AI-driven outcomes - not IT ownership. The shift from "AI projects" to "AI-enabled operations" is an operating model change, not a technology upgrade.

A strategic bet that has no framework to run will stall the same way the pilots did.

Five questions before you fund the next AI initiative

These apply whether the initiative is a quick win or a strategic bet. If you can't answer each of them clearly, the initiative isn't ready - or it's not a priority.

Which objective does this advance? Name the KPI. Name the expected movement.

• Is the data ready? Who owns it? What are the gaps? What are the regulatory constraints?
• Who has to work differently? What incentives or training close the adoption gap?
• What governance does this need? Which risks matter - bias, privacy, IP leakage, explainability? What controls are mandatory versus optional?
• What happens if it works? How does it get productised - and who runs it in steady state?