Fund the next phase, not the whole road
The large commitment asks the board to sign a full programme against a projection. By the time real evidence arrives, most of the money is spent, and stopping feels like admitting a mistake. So nobody stops.
The small pilot is cheap, but it answers the wrong question. It proves the technology can work. It rarely proves the business works differently, because it sits beside the work instead of inside it.
There is a third way. An AI programme should earn its budget the way an AI agent earns its autonomy: one step at a time, against evidence agreed before the step began.
A value gate is that step. It is a phase of work with a measure agreed before it starts, a baseline taken at the start, and a decision at the end: open the next phase, change it, or stop.
Budget should be earned like autonomy
In a well-run AI platform, an agent does not get freedom on its first day. A person approves its work. It gets more room only when the record shows it has earned it, and some decisions never leave a person at all.
Money deserves the same discipline. The CFO asks whether the spend is justified. The COO asks whether the work is getting done. At a value gate those become one question: what did the last phase prove?
The logic is the same in both places. Autonomy is extended against evidence, and so is budget. Neither is handed over on a promise.
What a value gate needs
- A measure in business terms. Hours returned to a team, errors caught before they reach a customer, enquiries answered the same day, software spend consolidated. Count it in rand or in hours, never in model scores.
- A baseline taken before work starts. Without a starting number, any later number is a story.
- A threshold written down in advance. Decide what result opens the next phase before anyone sees the result.
- A named owner on each side. Someone in your business and someone on the build team sign the gate together.
- Three honest outcomes. Go, change or stop. A gate that can only say yes is a payment schedule.
Where value gates go wrong
The gate comes too early. AI working inside live operations needs time for people to adopt it. Judge it in the second week and you measure the training, not the value.
The measure is whatever is easiest to count. The number of tasks an agent completed is activity. Value is what changed because of those tasks: the hours, the errors, the response times.
Nobody is allowed to close the gate. If stopping is politically impossible, the gate is decoration, and the programme is the large commitment again with extra paperwork.
The phase is too small to matter. Every phase should put something into live operations. A phase that only produces a report or a demo gives the gate nothing real to measure.
The gate is confused with a pilot. A pilot tests the technology beside the business. A phase runs inside the business, on real work, with real people relying on it. That is what makes its evidence worth paying for.
What changes for the board
The board stops approving a promise and starts approving a record. If the agreement is written around gates, the money at risk at any point is the current phase, not the whole programme. The conversation at each review moves from "do we still believe this?" to "here is what it did."
It also changes what you ask a supplier before you sign:
- What will the first phase put into live operations?
- Which measure and which baseline decide whether the next phase opens?
- Who in our business signs each gate?
- What happens, in writing, if the result says stop?
- Which decisions stay with our people, whatever the evidence shows?
If the answers are vague, the budget is not being earned. It is being assumed.
Earned, step by step
AI earns trust in a business the same way people do: by doing the work, in the open, against a standard everyone agreed. Trust that is earned in steps can be extended in steps. Ask the same of the budget.
Futuvara is a creative technology agency. We build, implement and run AI platforms for our partners. If you want to work out where your first value gate would sit, start a conversation.
