Stage gates for AI budgets under bubble risk

What is a stage gate for an AI budget, and how does it work?

A stage gate is a checkpoint where a project earns its next tranche of money. Before the gate, the team states a goal, a metric, a target, and a date. At the gate, a small committee compares the result with the target and decides to release funds, adjust the scope, or close the project. The decision takes one meeting and leaves a written record.

The format is familiar from product development, and it brings one rule to AI: money follows evidence. A project that shows value in a test unlocks a pilot, and a pilot that shows value unlocks production. Each step rests on data from the step before, so the budget grows only as fast as your knowledge.

Why do stage gates suit AI budgets when an AI bubble may burst?

A bubble means uncertainty about future prices, suppliers, and demand. Stage gates answer uncertainty with options: at each gate you can speed up, slow down, or stop. The EQOS Asia AI study scores the enterprise AI market at 0.337, a fragile zone where a shock can tip the system (EQOS Asia AI study, April 2026). A gated plan absorbs a shock without a rewrite of the whole budget.

The study also records that 22% of failed projects cite ROI not demonstrated, and 31% cite no alignment with the strategy. Gates address both. The metric at each gate forces the team to show a return, and the sponsor at each gate confirms the link with strategy. Both checks take little time compared with the spend they protect.

How many stage gates does an AI budget need?

3 or 4 gates fit most AI projects. Gate 0 approves a scoped test on a small dataset. Gate 1 approves a working pilot with real users. Gate 2 approves a limited production release. Gate 3, where useful, approves the full roll-out. More than 4 gates add meetings without adding information, and fewer than 3 hide the moment where a project should have stopped.

Name each gate by what the team will know after it, and let the money released follow from that. A test answers whether the data can support the use case. A pilot answers whether users adopt the tool. A limited release answers whether the economics hold at volume. These questions give the committee a clear way to read each result.

Which criteria should open each AI budget gate?

4 criteria cover most gates. Data signal: the team audits training-data quality and bias before launch, as the study recommends. Strategy fit: a sponsor confirms the link with a business goal. Explainability: the team can say why the model gives an answer, which the study asks for on critical decisions. Return: the agreed metric beats its target. Each criterion maps to a frequent cause of failure.

Write each criterion as a yes or no question with a number attached. A gate that asks whether the model is good invites debate. A gate that asks whether the error rate fell below 5% on 1,000 test cases invites a decision. Use numbers from your own baseline, since the right target depends on the process and the cost of an error.

How do you split an AI budget across stage gates?

Release money in tranches that grow with evidence. One workable pattern gives 10% to 15% of the total to the scoped test, 25% to 35% to the pilot, and the remainder to production, so most of the money moves after 2 gates have passed. Adapt the shares to your sector, and keep the first tranche small enough to lose without damage to the year’s plan.

Add a reserve of 2% to 5% of the total for retraining, which the EQOS Asia AI study recommends for roles that AI automates (EQOS Asia AI study, April 2026). Hold the reserve outside the tranches so it stays funded whatever the gate result. People who understand the tool adopt it faster, and adoption drives the return metric at the next gate.

When should an AI budget gate stop or pause a project?

Set the stop rule before the gate, when emotions are calm. A project stops when its metric misses the target by a stated margin, when a key risk appears such as a data leak or a public hallucination, or when its sponsor leaves. A pause suits a project that meets the metric yet depends on a supplier decision that is still pending.

The study offers a reference for thresholds: below 0.3 on a dimension is the STOP zone, and a market score above 0.5 marks viability for a successful AI deployment (EQOS Asia AI study, April 2026). Communication stands at 0.305 in the market today, close to that line. Apply the same logic to your project and set your own line.

How often should AI budget gates be reviewed as the market moves?

Hold a gate review when a project reaches its date, and a portfolio review every quarter. The quarterly review rereads the market indicators, checks supplier news, and confirms that the total commitment still fits what the company can carry. If indicators worsen, shorten the next tranche. If they ease, keep the schedule. The rhythm is simple enough for a monthly finance meeting to follow.

EQOS Asia scores gate decisions on 9 dimensions. The Λ simulator gives a first score in about 10 minutes, a Reading at US$358 returns a mathematical answer in 24 to 48 hours (EQOS Asia pricing page, October 2026), and Sentinel re-measures the model every month. A 30-minute pre-diagnostic shows which option fits your portfolio of AI projects.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile