#029 - The Missing Line in Your AI ROI Is Called Integration.
Month nine.
The AI works. That is what nobody planned for.
It answers correctly. It saves the hours the vendor promised. And somewhere in finance, a director is staring at what it cost to get there.
The tool was never the expensive part.
Rewind eleven months.
The business case fit on one slide. Licence on the left. Labour saving on the right. A number at the bottom that made the decision easy.
Every figure was true. Every one of them came from somebody whose job ended before the bill arrived.
The vendor closed at signature. The consultant's scope ended at the pilot. The sponsor's objective for the year read "AI deployed," not "AI still working in twelve months."
Nobody lied. Nobody owned the rest of the bill either.
Then the project started.
Data needed cleaning. Interfaces needed rebuilding. Security asked for a new control. The process owner discovered that the old workflow could not carry the promised value. People needed time to learn a different way of working.
Those costs arrived through other budgets, so the original business case still looked healthy.
The company bought a tool and quietly financed an operating model around it.
The firms will tell you this themselves
The European Central Bank surveyed over 5,000 euro-area firms. More than 70% report using AI. Only 7% report using it intensively.
Then the ECB asked what stops everyone else. The firms did not name the model. They named shortages of AI-related skills (40%), limited usefulness of current AI for their actual business needs (28%), and incompatibility with existing systems (26%).
Skills. Process fit. Integration.
That is not a technology gap. That is the second invoice, itemised by the people who already received it.
A separate IBM survey of 2,000 technology executives puts a number on the blindness: 84% had not fully operationalised AI financial management, and 85% lacked full visibility into real-time AI spend. IBM sells the remedy, so treat that as a symptom report rather than a benchmark.
An instrument built to kill things
Most cost templates exist to justify a decision that has already been taken. This one is built to stop one.
In strategy work I ask for one number before any AI investment is approved: the full cost of the workflow, not the price of the model.
The True AI Cost Card puts seven layers into that number:
- Model: licence, API, hosting, usage and evaluation environments.
- Data: access, rights, quality, pipelines and stewardship.
- Integration: systems, workflow redesign, testing and exception paths.
- Controls: security, privacy, assurance, human oversight and audit.
- Adoption: role changes, training, protected capacity and communication.
- Lifecycle: monitoring, evaluation, model changes, deprecation and exit.
- Run: support, incidents, infrastructure, vendor management and financial control.
Every layer takes five entries: a one-off cost, a run rate, your confidence in both, the owner, and the date the evidence was last checked.
A blank means unknown.
Zero requires an owner, evidence and a date.
Then write the expected value and the full cost over the same period, and ask the finance owner, the process owner and the technology owner to approve the same case.
The verdict most people avoid
Sometimes the full cost kills the business case.
That is the Card working, not failing. You have found out in a planning meeting instead of in month nine, and the budget is still in your hands.
Return it for rework when the economics depend on one cost layer disappearing into somebody else's budget.
Run the Card before your next AI approval. The first blank row shows where another team will inherit the bill.
Your move.
Planning instrument only. Adapt the cost categories and accounting treatment with the responsible finance, procurement, technology and risk specialists.
