AI Business Cases

Cloud adoption followed a sequential path with predictable cost categories, which is why TCO and ROI models served it well. AI adoption runs in parallel across the organisation, with initiatives at different maturity stages, on different timelines, and creating value in functions other than the one that funded them. A uniform threshold applied across that landscape misallocates capital in both directions. I set out the full argument in Rethinking Business Cases in the Age of AI: What Boards Need to Know.

IDC research, reported by CIO.com in 2025, found that 88% of AI pilots never reach production, and in my experience the cause is rarely the technology. Cases are approved without predefined scaling triggers, handover processes, or governance designed to evolve from experiment to production, so pilots become curiosities. Securing Buy-In from the Board sets out what to build into the approval itself.

By widening the instrument set. Lagging indicators confirm value after the fact; leading and predictive indicators signal it early enough to govern with, and the Well-Advised principle organises them across the five dimensions Boards consistently care about. The largest returns are typically second- and third-order effects that mature over 12 to 24 months or longer, and Boards that measure only first-order gains risk abandoning initiatives whose real value lies ahead.

No. A healthy AI portfolio holds three different kinds of investment: quick wins that fund momentum, capability builds whose value is the initiatives they enable, and long-horizon initiatives judged on option value and strategic position. Imposing a single hurdle rate collapses that distribution into one category and starves the foundations. I develop the portfolio approach in Finding High-Value AI Opportunities and the finance-function version in The Appreciating Ledger.

They do not, and this is the trap. A business case is a tool of isolation: it evaluates one initiative against narrow criteria and cannot ask whether the pieces reinforce or undermine each other. Organisations that govern AI project by project tend to accumulate incompatible data models, duplicated governance, and pilots that never compound. Avoiding the Business Case Trap explains why the Board’s role is orchestrating systematic capability, not gatekeeping proposals.

Data preparation first. In my experience initial budgets allocate 15 to 20% to it, and organisations end up spending 50 to 65% of project cost cleaning, structuring, and governing data. Then the recurring lines: model monitoring and retraining, governance overhead, specialised talent, and parallel running of old and new systems during transition. The newest omission is oversight labour, the reviewing and correcting AI outputs need before anyone can rely on them; BCG Henderson Institute research in Harvard Business Review (2026) found 14% of workers using AI reporting the strain. The True Investment Profile sets out the cost categories, and The Balancing Item prices the one almost every case misses.

On the internal scorecard, yes. The capability AI investment builds (the cleaned data, governance, prompt libraries, and trained workforce) appreciates through use and is reused by initiatives other than the one that paid for it, which no depreciation schedule records. IAS 38 already sets four tests for an intangible asset, identifiability, control, measurability, and future economic benefit, and modern AI systems meet them; the rules permit the treatment, practice has not caught up. The practical move is to measure it on the management report while the statutory accounts still expense it, and to brief the audit committee early. The Appreciating Ledger makes the case to the CFO.

Start with processes rather than products; the fashionable applications are rarely where the value sits. I score each candidate process on decision complexity, information volume, process variability, cognitive repetition, and knowledge continuity risk, and look hardest at those that score high on several. The shortlist then passes through the Well-Advised principle, which favours initiatives delivering across several pillars over a spike in one, and a feasibility check against the Five Pillars diagnostic. One question comes before all of that: whether the work needs doing at all. Finding High-Value AI Opportunities sets out the method, and Not Everything Needs AI the question that precedes it.