Sizing an AI pilot when the AI bubble may burst
The essentials in 30 seconds
- An AI pilot is sized by one question, one loss cap, and one exit rule, and bubble risk shapes the cap.
- As of October 2026, 95% of AI projects fail in production (MIT 2025, cited by the EQOS Asia AI study).
- Insufficient or biased data causes 42% of failures, so the pilot tests data quality first.
- A cap of 5% of the annual AI budget is one workable starting point.
- A Reading from EQOS Asia costs US$358 and returns in 24 to 48 hours.
Sizing an AI pilot is the first spending decision that a business leader controls fully. As of October 2026, the EQOS Asia AI study shows that most AI projects stop before production, and AI valuations have risen ×4 in 18 months. A well-sized pilot buys evidence at a price your company can carry in any market. This page extends the decision guide on AI spending under bubble risk.
On this page
Compare the EQOS Asia offers and request a 30-minute pre-diagnostic.
What question should an AI pilot answer when the AI bubble may burst?
A pilot answers one question: does this use case create measurable value with our data, our people, and our workflow? Everything else, from vendor demos to market excitement, is secondary. Write the question in one sentence and the number that would answer it. If the sentence needs 2 numbers, the pilot is 2 pilots, and each deserves its own cap.
The bubble question changes the pilot in one way: it makes the cost of learning visible. A pilot that costs little and answers clearly keeps its value whether the market rises or falls. A pilot that depends on a supplier’s promotional price loses value the day prices reset. Test the use case at the price you expect to pay in year 2.
How do you cap the loss of an AI pilot?
Start from the loss you can carry, then work backwards to the scope. One workable rule caps the pilot at 5% of the annual AI budget. Take a company with a US$1,000,000 AI budget: the cap is US$50,000, and a pilot that fails entirely leaves 95% of the budget available for the next idea. These figures are an illustration, so set your own.
Count every cost inside the cap: licences, compute, data preparation, staff time, and outside help. Staff time is the cost that people forget, and it often exceeds licence fees. Review spending against the cap every 2 weeks, and alert the sponsor when spending reaches 80% of it, so decisions about the last 20% are made with open eyes.
How narrow should the scope of an AI pilot be?
Make the scope narrow enough for one team to own it and real enough for users to depend on the output. Pick one process, one team, and one dataset. A pilot on invoice matching in one finance team teaches more than a pilot on a vague goal such as improving productivity. Narrow scope also makes the result easy to compare with the old way of working.
Set a duration of about 90 days. Shorter windows miss seasonal patterns and slow adoption, and longer windows invite scope creep. Fix a start date, a mid-point review at day 45, and an end date. Dates keep the pilot tied to a decision, and they let the budget owner plan the next release of funds before the market moves.
Which data and team size does an AI pilot need?
Insufficient or biased data causes 42% of AI project failures in the study (EQOS Asia AI study, April 2026), so audit the data before the first model runs. Check completeness, labelling, bias, and edge cases on a sample. The study recommends this step before any deployment, because AI amplifies noise when the signal is weak.
A small team is enough: one business owner, one data specialist, one engineer, and a part-time reviewer who checks outputs. 4 people can run a focused pilot and keep coordination to one daily message. The study also records that 47% of executives took a major decision based on a hallucination, so name the reviewer from the start.
Which metrics prove that an AI pilot works?
Choose 3 metrics: a quality metric, a business metric, and a trust metric. Quality is the error rate against a reference set. The business metric is time saved, cost cut, or revenue added per unit. Trust is the share of outputs that reviewers accept as they are. The study notes that hallucination rates range from 0.7% to 30% depending on model and domain, so measure yours on real cases.
Write the targets before the pilot starts, and record the baseline of the current process. A baseline turns every result into a clear difference. Keep the metric list short, because every extra metric gives a team another way to describe a weak result as a success.
When should an AI pilot end, scale, or stop?
Decide at the end date with 3 outcomes in mind. Scale when all 3 metrics reach their targets. Extend once, for up to 30 days and with a fixed extra cap, when 2 metrics reach the target and the third is close. Stop when quality or trust stays below target, and record what the team learned. Each outcome is a good decision when the evidence supports it.
Scaling follows your budget gates rather than the pilot’s momentum. The study gives a viability reference: a score above 0.5 marks a successful deployment, and 4 corrections lift the market score from 0.337 to 0.401 (EQOS Asia AI study, April 2026). Apply the same corrections to your pilot, from data audit to retraining, before you commit production funds.
How can EQOS Asia help you size an AI pilot?
EQOS Asia scores the pilot decision on 9 dimensions and returns a reproducible result. A Reading costs US$358 and arrives in 24 to 48 hours, and a Decision at US$910 integrates your own documents in 48 to 72 hours (EQOS Asia pricing page, October 2026). The Λ simulator offers a free first estimate in about 10 minutes.
Bring the use case, the budget, and the cap you have in mind. A 30-minute pre-diagnostic tests whether the pilot is sized to the lesson and shows which of the 8 tiers fits. You leave with a score, a ranked set of options, and a pilot plan your board can approve in one meeting.
See also in this guide
Frequently asked questions
How big should an AI pilot be?
Size it to one question and a loss cap your company can carry. A cap of 5% of the annual AI budget is one workable starting point, with one process, one team, one dataset, and about 90 days. Adapt every figure to your own situation.
What should an AI pilot measure?
Measure 3 things: quality as an error rate against a reference set, business value as time or cost per unit, and trust as the share of outputs that reviewers accept. Record the baseline of the current process before the pilot starts.
Why test data quality before running the pilot?
The EQOS Asia AI study, April 2026, lists insufficient or biased data as the cause of 42% of AI project failures. A data audit before the first model runs shows whether the signal is strong enough to support the use case.
What happens at the end of an AI pilot?
The team scales when all 3 metrics reach their targets, extends once for up to 30 days when 2 metrics do, and stops when quality or trust stays below target. Each outcome records what the team learned for the next use case.
Test your knowledge
1. How many questions should an AI pilot answer?
- One
- Five
- As many as possible
Show the answer
One.
2. What does the page suggest as a starting cap for a pilot?
- 50% of the annual AI budget
- 5% of the annual AI budget
- The whole AI budget
Show the answer
5% of the annual AI budget.
3. What share of AI project failures does the study link to insufficient or biased data?
- 4%
- 95%
- 42%
Show the answer
42%.
4. What is the suggested duration of a pilot?
- About 90 days
- About 3 years
- About 1 week
Show the answer
About 90 days.
5. When does the team scale the pilot?
- When the supplier offers a discount
- When the budget is spent
- When all 3 metrics reach their targets
Show the answer
When all 3 metrics reach their targets.
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
Sources: EQOS Asia AI impact study (eqos-asia.com/case-studies/ia-2026.html), EQOS Asia pricing page (eqos-asia.com/pricing.html), EQOS Asia Λ simulator (eqos-asia.com/simulator.html) and contact page (eqos-asia.com/contact.html), consulted on 22 October 2026. Page updated on 22 October 2026.