Why Most AI Business Cases Fail to Convince Boards. And How to Fix Them

The most common credibility problem in AI business cases is the use of vendor-published ROI data as the primary evidence for projected returns. The fact that Microsoft claims Copilot users save 8.8 hours per week, or that a vendor case study shows 40% productivity improvement, is not evidence that your organisation will achieve comparable results.

Boards that have been through previous technology investment cycles where vendor ROI claims did not materialise in practice are appropriately sceptical of vendor data. A business case that relies primarily on vendor-published figures will be challenged by any finance-literate director who has seen this pattern before.

The fix is to use vendor data as the upper bound of what is possible, not as the central estimate, and to base the investment case on conservative estimates derived from pilot data or comparable internal technology investments. If pilot data does not yet exist, that is itself the argument for a small exploration investment, not for a full programme investment.

AI business cases routinely underestimate costs. The technology licence cost is usually well-specified. The implementation cost is often estimated by the vendor or implementation partner, who has an interest in keeping the number low. What is almost always missing is the change management cost, the ongoing governance cost, the data quality remediation cost, the training refresh cost, and the staff productivity cost during the transition period.

Boards and CFOs who have been surprised by implementation overruns on previous technology programmes will probe these numbers. A business case that does not acknowledge and quantify the non-licence costs will be questioned, and the answers often undermine confidence in the overall financial projections.

The fix is to include a comprehensive cost model that explicitly addresses all cost categories, uses conservative assumptions, and includes a contingency for overruns that is proportionate to the level of implementation complexity.

Many AI business cases describe ambitious business outcomes without specifying who is accountable for achieving them. The business case might project significant productivity savings, but if no specific executive is named as accountable for delivering those savings, and no mechanism is defined for measuring and reporting progress, the projection is a wish, not a commitment.

Boards are increasingly sophisticated about this. A board that asks who is accountable for this and how they will know if it is working expects specific answers: a named individual, a defined measurement methodology, and a reporting cadence. Business cases that cannot answer these questions are incomplete.

The fix is to include in the business case a specific accountability section: who owns the outcome, what metrics will be tracked, at what frequency the board will receive progress reports, and what happens if the programme is not on track.

AI business cases frequently present benefits with confidence and risks with vagueness. A board is entitled to ask about data protection compliance, about the risks of AI-generated errors in the relevant context, about the regulatory position if an AI system produces a discriminatory output, and about the vendor dependency implications.

The absence of a credible risk assessment does not make the board think the risks are low. It makes them think the executive team has not thought carefully about them, which reduces confidence in the overall quality of analysis.

The fix is to include a genuinely honest risk section that identifies specific risks, assesses their likelihood and impact, and describes specific mitigations. A board that sees honest risk analysis is more likely to approve an investment than one that sees only benefit projections, because honest risk analysis signals that the executive team has thought rigorously about the investment.