The AI budget nobody owns

The invoice usually arrives before the architecture. A marketing team buys a writing tool, engineering adds a coding assistant, support subscribes to a chatbot, finance experiments with document extraction. Each purchase is small, each is justified by a local pain, and none of them is a decision anyone would describe as strategic.

Twelve months later the company is running six AI vendors, three of which overlap, two of which hold customer data under contracts nobody has read, and one of which is being paid for seats that were never activated.

This is the most common shape of AI adoption in a company of 30 to 200 people, and it is not a spending problem. It is an ownership problem.

Why it happens

Three forces push in the same direction.

The buyer is not the owner. The person who feels the pain — a marketer with too much copy to write, a support lead drowning in tickets — is the right person to adopt a tool and the wrong person to assess lock-in, data residency or whether the tool will still exist in two years.

The vendors sell to that person. AI products are priced for a credit card, not a procurement committee. Free tiers, seat-based plans and monthly cancellation make adoption frictionless, which is exactly why it spreads without review.

Nobody is accountable for the total. In a company with a CFO, spend aggregates somewhere. In a company without a technical executive, it aggregates nowhere: it lives on departmental cards and in personal expense claims.

A team adopts a tool
Local pain, local budget, one click. The right call for that team, invisible to everyone else.
A second team adopts an overlapping one
Nobody compares, because comparing is somebody’s job and the job is vacant.
Data starts leaving the building
Customer records, contracts and code go to vendors under terms no one has reviewed.
The renewal arrives
The first moment the company looks at the whole picture is the moment it is asked to pay for it.

The usual sequence. None of the four steps is unreasonable on its own.

Cover of the ABN Asia film on the AI budget nobody owns

What it actually costs

The visible cost is the subscription line. The expensive part is everything around it.

Duplication is the easy one to fix and the easiest to miss: two tools that do the same job, bought by two teams, each convinced it is the only one.

Integration debt is worse. Every tool that touches a workflow has to be connected to the systems of record, and each connection is a small piece of software nobody owns. When one vendor changes an API, something breaks in a place no one is watching.

Data exposure is the one that ends up in front of a board. Sending customer data to a model provider is a processing decision with legal consequences; it should not be made by whoever clicked "accept" on a free trial.

Lock-in is the quiet one. Prompts, fine-tuned behaviour, evaluation sets and workflows accumulate inside one vendor's interface. Moving later costs more than the subscription ever did.

AI vendors is a typical count for a 100-person company after a year of bottom-up adoption

people accountable for the total spend in most of those companies

inventory is the first deliverable: what is running, who owns it, what data it sees

An illustrative case

Take a 40-person services firm. Three AI tools in use: a writing assistant at 20 seats, a support bot at 10 seats, and a document extractor billed by usage. Nobody has cancelled the seats belonging to the two people who left, the support bot duplicates a feature already included in the helpdesk plan, and the extractor is sending client contracts to a vendor whose data terms were never reviewed.

None of those facts is dramatic. Together they describe a company paying twice for one capability, exposing data it did not intend to expose, and holding no lever to change any of it — because there is no owner to change it.

What an owner does

The work is unglamorous and takes about a month.

Inventory. Every tool, every seat, every contract, every data flow. Most companies discover 30–50% more spend than they expected.
Consolidate. Pick one tool per job. The saving is real, but the point is the shorter list.
Set a purchase rule. Anything that touches customer data or costs more than a defined threshold goes through one review.
Write the exit. For each vendor: how would we leave, and what would it cost? A contract without an exit is a decision you have already made.

Inventory, then consolidate
One tool per job, one owner per tool, one list everyone can see.
Attach the data question
What leaves the building, under which terms, and who signed for it.
Define the purchase rule
A threshold and a data test, so the next adoption is a decision rather than an accident.
Keep the exit clause alive
Review at renewal: still used, still compliant, still cheaper than the alternative.

Four steps, roughly a month of part-time work, and the budget becomes a number someone can answer for.

Why this is a fractional role

A full-time CTO would own this, and in a larger company that is exactly what happens. Below that threshold, the work is real but the volume is not: an inventory once, a rule to maintain, a renewal to review twice a year, and a decision when something new appears.

That is a cadence, not a headcount. It is also the difference between a company that can answer "what are we spending on AI and why" and one that finds out when the card is declined.

The budget is not the problem. The absence of an owner is.

Source: ABN Asia analysis. Second in a series on fractional executive leadership.

Contact: [email protected]