Opinion/Contrarian: Why AI ROI Fails More in Finance Design Than in Model Design
A practical contrarian take on why AI ROI breaks at the management-accounting layer—and what operators can fix this quarter.
Most AI ROI failures are not model failures. They are management accounting failures.
That sounds uncomfortable, but it explains a pattern many mid-market teams know too well: pilots look promising, dashboards show activity, and leadership still cannot see clear margin improvement. The issue is rarely that AI tools are useless. The issue is that companies track AI like a technology experiment instead of a business investment.
When that happens, teams optimize what is easy to measure (usage, tickets closed, response time) and ignore what actually moves enterprise value (gross margin, throughput per FTE, cash conversion, and risk-adjusted cost).
If you are running AI in operations, finance, or customer workflows, here is the contrarian view: stop asking whether the model is smart enough. Start asking whether your accounting design is strong enough to prove value.
## Why “activity metrics” keep fooling leadership
Most organizations can report AI activity quickly:
- Number of copilots deployed
- Number of automations launched
- Number of employees trained
- Number of prompts or interactions
Those are adoption signals, not ROI signals. They matter, but they do not answer the core executive question: did this create measurable economic benefit after all costs?
This is where teams get stuck. AI programs can look successful in steering committees while still failing the P&L test. A business can spend months improving model quality and still miss the value target because the measurement frame was wrong from day one.
## The four accounting gaps that kill AI ROI
### 1) Benefit statements are vague
“Save time” is not an investable claim. To survive budget review, benefits must map to clear financial levers:
- Revenue growth (higher conversion, better retention, smarter pricing)
- Cost reduction (lower handling cost, reduced rework, fewer escalations)
- Capacity gain (more output without linear headcount growth)
- Working capital improvement (faster cycles, fewer errors, better forecasting)
If a use case cannot map to one of these, treat it as an experiment, not a scaled transformation bet.
### 2) Baselines are weak or missing
Without a clean baseline, every ROI claim is negotiable. Teams often launch AI while other changes are happening at the same time: staffing shifts, policy changes, seasonality, or a parallel systems rollout. Then they attribute all improvements to AI.
That inflates impact and destroys trust with finance.
A stronger setup includes:
- A fixed pre-AI baseline period
- A clear comparison group when possible
- Explicit exclusion rules for confounding effects
No baseline, no credible ROI.
### 3) The full cost stack is hidden
Many cases include only license or API fees. Real operating cost is broader:
- Data engineering and integration effort
- Governance, controls, and review workflows
- Security and compliance overhead
- Change management and training
- Ongoing monitoring and model quality work
If these costs are outside the business case, projected ROI will look fantastic on slides and disappointing in monthly reporting.
### 4) Time horizons are mismatched
Not every AI use case pays back at the same speed. Some automation projects can show gains in one quarter. Others need two or three planning cycles before benefits stabilize.
Applying one payback expectation to all projects creates two bad outcomes: high-potential cases get cut too early, while weak but noisy projects stay alive because they produce short-term activity.
Portfolio governance should classify use cases by payback profile, not by hype level.
## A practical CFO-operator model for AI value
A working model is simple and disciplined:
1. **Define one primary economic KPI per use case.**
Example: cost per resolved service case, quote cycle time, or forecast error impact on inventory.
2. **Define one operational driver KPI.**
Example: first-contact resolution, straight-through processing rate, or exception rate.
3. **Track total cost to run, not just build.**
Include people, controls, model ops, and vendor spend.
4. **Review value monthly, not just at launch.**
AI value drifts when process conditions change; governance must catch it quickly.
5. **Fund by evidence tiers.**
Move from pilot to scale only when economic lift is demonstrated against baseline.
This model is not glamorous, but it prevents “pilot theater” and improves capital allocation.
## Where teams should start this quarter
If your organization already has live AI initiatives, do a 30-day reset rather than starting another pilot.
- Select the top five active use cases by spend.
- Rebuild each case with one economic KPI and a clean baseline.
- Recalculate ROI with full operating costs.
- Freeze expansion for any use case that cannot show evidence-based lift.
- Reallocate budget to one or two use cases with the strongest proven unit economics.
This approach usually does more for enterprise value than launching three new experiments.
## The leadership shift that matters most
AI strategy is often presented as a technology roadmap. In practice, it is a management system decision.
Companies that win with AI do not just deploy better tools. They make better operating choices: tighter definitions of value, stronger cost discipline, and faster reallocation away from weak bets.
They also create shared ownership between finance, operations, and product leaders. When one team owns delivery and another owns measurement, value leaks through handoffs. Shared accountability forces clearer trade-offs, faster decision cycles, and fewer vanity deployments.
That is why the accounting lens matters. It changes who gets funded, what gets scaled, and how quickly teams stop doing work that looks innovative but does not create measurable business benefit.
The contrarian takeaway is straightforward: if AI ROI is disappointing, do not start with prompts. Start with management accounting design.
## Sources
- https://www.nist.gov/artificial-intelligence
- https://www.ifrs.org/issued-standards/list-of-standards/ifrs-15-revenue-from-contracts-with-customers/
- https://www.ey.com/en_gl/services/ai
- https://www.pwc.com/gx/en/services/ai.html
- https://www.microsoft.com/en-us/worklab/work-trend-index
