7 Governance Failures Quietly Undermining AI Investment ROI
AI investment ROI in mid-market financial services is rarely defensible to audit standard. The gap between approved benefits and provable benefits is structural.
- AI investment ROI in mid-market firms is rarely defensible to audit standard. The gap between approved benefits and provable benefits is structural — and growing as deployment expands.
- The seven failures below describe where ROI leaks. Each is a governance gap rather than a technology failure. Each is recoverable — but recovery requires CFO-led architectural intervention, not more AI deployment.
- The board-defensible position is not “we delivered the AI programme.” It is “we can evidence the return on the investment and would defend the evidence under audit.” Few mid-market firms are at that position today.
The conversation about AI ROI in UK mid-market boards has shifted in the last twelve months. Six months ago, the question was “what should we invest?” Three months ago, the question was “are we investing enough?” Today, the question is “what did we actually get for the investment we made — and can we defend the answer?”
The honest answer in most firms is uncomfortable. The investment has been made. The use cases are live. The benefits case showed substantial returns. The actual returns are partial, unevenly attributed, and not defensible to audit-standard. The CFO knows it. The CIO knows it. The board is starting to ask.
This is not an AI quality problem. The models work. The deployments are technically competent. The governance architecture that should have surrounded the investment was not built — and without it, the returns that were promised at approval cannot be evidenced at review.
The seven failures below describe how AI investment ROI quietly leaks through governance gaps. Each one is recoverable. Each one is materially expensive. Each one is invisible until the CFO is asked to defend the return — and discovers that the architecture which would make the return defensible was never commissioned. The starting point is naming where you currently sit.
AI spend is dispersed across multiple cost lines with no single accountable line item
The CFO is asked: what is the AI investment? The honest answer is “it depends what you count”. Some sits in IT capital. Some in marketing operations. Some in customer service tooling. Some in consultancy retainers. Some in vendor contracts that include AI as a feature. Some in employee training. The aggregate cannot be assembled without a special exercise.
Without a single accountable line, the investment cannot be governed as a category. ROI cannot be measured against a defined denominator. Spend cannot be controlled against a target. Reallocation cannot be evaluated, because the comparable lines are not comparable.
This is the foundational governance failure. Every other failure below is downstream of it. The CFO cannot answer board questions about AI without first being able to point to what is being spent on AI — and to what is being returned for that specific spend. The architectural fix is to create an AI cost category in the management accounts, with sub-categories by use case, owned jointly by CFO and CIO and reviewed at the same cadence as any other category of capital investment.
There is no baseline operational metric against which AI return is measured
The AI use case goes live. Six months later, the team reports success. Cycle time has improved. Customer satisfaction is up. Throughput has increased. The team is confident the AI is delivering.
The CFO asks: what was the cycle time before AI? What was customer satisfaction before? What was throughput before? In most mid-market firms, the answer is “we have rough estimates” or “we measured something similar”. The baseline is not crisp.
Without a crisp baseline, the return is not measurable to audit standard. The improvement may be real. It may also be the result of seasonal effects, parallel process changes, sample composition, or measurement methodology. Without baseline discipline, the AI cannot be credited or discounted with confidence. The architectural fix is to establish operational baselines before AI deployment — the baseline metric, the measurement methodology, the comparison cohort, and the attribution rule documented in advance. The post-deployment measurement uses the same protocol. Without it, the most common outcome is that AI looks better than it is — and the next investment case is built on the inflated reading.
Benefits are attributed to AI without attribution discipline
The marketing campaign converted better. AI was part of the campaign. The improvement was attributed to AI. The benefits case is updated to claim the win.
This is the attribution problem. AI is rarely the only thing that changed. The team that deployed AI also redesigned workflows. The campaign that used AI also had new creative. The customer service team that adopted AI also rolled out new training. Each improvement may be real. The portion attributable to AI specifically is much smaller than the portion claimed.
Most mid-market firms do not have the attribution architecture to separate the AI effect from the surrounding changes. Holdout groups are rarely run. Counterfactual baselines are rarely modelled. Attribution is done by inspection: “the AI launched, the number went up, the AI must have driven it”. The cost shows up at year-end review when the CFO has been promised certain returns and the returns appear in aggregate but cannot be precisely attributed. The next AI investment case relies on the same un-attributed evidence. The architectural fix is attribution architecture — holdout cohorts, counterfactual modelling, pre-registered hypotheses. These are not new techniques. They are simply not built in.
The board approved an investment case. Twelve months later, the CFO cannot defend the returns. Neither was an AI failure. Both were governance failures.
The AI investment was sized once at approval, never re-sized as scope evolved
The original case was approved at a defined scope. Three use cases. Two integrations. One vendor. A budget of £X. A timeline of Y months. Returns of Z over three years.
Six months later, the scope has expanded materially. Two more use cases were added — they seemed adjacent and low-cost. A new vendor was brought in for a specialised capability. The original budget has been topped up via change requests. The cumulative spend is materially above the approved case.
The board sees each change request individually. None crosses the threshold for re-approval. Cumulatively, the AI investment has grown significantly. The ROI calculation has not been refreshed. The original ROI was for the original scope. The actual ROI must now be assessed against the actual investment, but the two are no longer comparable. The board cannot defend the AI portfolio because the portfolio they approved is not the portfolio that exists. The architectural fix is to schedule formal portfolio reviews at programme milestones — not just at board cycles — that re-baseline investment and re-baseline expected return.
AI vendor pricing is not architected against the business case
AI vendors price on usage — per call, per token, per inference, per seat, per million queries. The cost scales with the deployment, often non-linearly. Volume discounts kick in. Premium features tier upward. The unit economics are usually transparent, but rarely architected against the business case.
What happens is predictable. The deployment scales. Usage scales faster than expected. The CFO sees the vendor bill grow quarter-on-quarter. The vendor relationship was procurement’s. The cost trajectory becomes commercial’s. The architectural connection between unit cost and business value sits nowhere.
In well-architected AI investments, the unit economics are modelled in advance — cost per AI-driven decision, cost per inference, cost per use-case-as-deployed. The model is reviewed quarterly. Trajectory changes are flagged before they become material. In poorly governed AI investments, the vendor cost is a surprise every quarter. The framing “it’s the cost of doing business with AI” is wrong. The cost is the cost of unit economics not being architected at investment stage. The architectural fix is straightforward but rarely done: model the economics, model the trajectory, renegotiate when it diverges.
Recognising the governance gaps is the first step. Naming where your AI investment defensibility sits is the next.
The free Commercial Readiness Assessment positions your organisation across six dimensions of commercial architecture, including AI governance. About ten minutes. No payment. No sales call.
Take the Free Assessment →No defined disinvestment criteria for AI use cases that underperform
Most AI governance frameworks have approval gates. Few have disinvestment gates. The result is that AI use cases get approved on optimistic forecasts and then continue running regardless of whether they deliver.
In commercial portfolio governance — capital expenditure, M&A integration, product launches — disinvestment criteria are standard. A product launch that hasn’t hit defined milestones at twelve months is reviewed for retirement. A capex project that has overrun by a defined margin triggers an executive review.
AI use cases rarely have this discipline. The reasons are partly cultural — AI is treated as strategic investment, not operational deployment — and partly architectural — there is no portfolio governance to apply the discipline against. The cost is significant. AI use cases that aren’t delivering continue to consume resources. The team that built them continues to invest in tuning. The vendor cost continues to accumulate. The opportunity cost — investment that could have gone to AI use cases that work — is unmeasured. The architectural fix is to design disinvestment gates into AI governance: defined performance thresholds, defined review milestones, defined retirement protocols.
Three diagnostic tests for AI investment defensibility
- Can you produce a one-page summary of total AI investment by use case, with attributed return for each, that you would defend in an audit committee? If not, the line-item discipline isn’t built.
- For your three largest AI use cases: what was the operational baseline before deployment, what is the post-deployment measurement, and what is the attribution evidence? If the answers are vague, the ROI is not audit-defensible.
- In the last twelve months, have you disinvested from any AI use case for performance reasons? If no, the governance has only approval gates, not portfolio gates — and underperforming investments are continuing by default.
The CFO governs AI as cost, not as portfolio
The most structural of the seven. Each AI use case is reviewed in isolation. The marketing AI is reviewed against marketing’s targets. The service AI is reviewed against service’s targets. The fraud AI is reviewed against fraud’s targets. Each is a sub-conversation in a sub-functional review.
What is missing is the portfolio view. Across all the AI investments in the business, which are delivering, which are not, where should reallocation happen, what is the marginal return on the next AI use case versus the marginal return on increasing investment in an existing one. This portfolio conversation is standard in capital allocation. It is rarely formalised for AI.
The cost is misallocation. Resources continue to flow to AI use cases that aren’t delivering, while AI use cases that could deliver are starved. The CFO sees aggregate AI spend rising. The CIO sees deployment expanding. Neither sees the portfolio shape. The architectural fix is to bring AI investment into the same portfolio governance as any other capital category. Quarterly portfolio review. Defined reallocation authority. Standardised performance reporting. Most mid-market firms do not do this — and the cost compounds quarter after quarter.
What this means for AI investment defensibility
These seven failures share a structural pattern. None of them are AI failures. All of them are governance failures that destroy the economics of AI investment before the AI itself has had a chance to deliver. The pattern is consistent across mid-market firms in financial services, insurance, and beyond.
The most expensive misconception about AI investment is that it is fundamentally different from other capital investment. It is not. AI investment needs line-item accounting, baseline measurement, attribution discipline, scope governance, vendor economics modelling, disinvestment criteria, and portfolio review. These are unglamorous capabilities. They are also the capabilities that determine whether the investment is defensible at audit, at board, and at the next funding round.
The board-defensible position is not “we have deployed AI across the business”. It is “we can show what we spent, what we got, and how we know the difference”. Most mid-market firms can only credibly answer the first part of that question. The CFO who can answer all three positions the organisation materially ahead of its peers.
This is where governance pays back. Not as a brake on AI investment. As the structural discipline that converts AI investment into defensible returns. The starting point is naming where you currently sit.
Are your AI investments delivering defensible returns?
The free Commercial Readiness Assessment positions your organisation across six dimensions of commercial architecture, including AI governance specifically. You receive a personalised report naming where AI investment is best governed, where it is most exposed, and which of the seven failures above are most likely to be reducing your defensible ROI.
Take the Free Assessment →About 10 minutes · No payment · No contract · No sales call