Most companies do not have an AI pilot plan. They have a pilot. Someone picked a tool, a team tried it, and a quarter later nobody can say what it proved. The pilot did not fail. It never had a definition of success that anyone wrote down.
An AI pilot plan is a 90-day commitment to answer one question with evidence: should this workflow scale, change, or stop? It sets the use case, the success threshold, the owner, and the decision date before any tool is switched on. Without those four, a pilot is a demo with a budget.
This post covers what a 90-day AI pilot plan must prove, how to run it in three phases, and what a scale gate looks like. For the week-by-week version, read our 90-day AI implementation roadmap. This one is about the decision at the end.
Why do most AI pilots fail to scale?
The failure numbers are loud and the methods behind them are uneven, so read each one with its limit. MIT’s NANDA initiative reported in 2025 that 95% of enterprise generative AI solutions fail to deliver measurable profit and loss impact. The finding rests on about 150 leader interviews, a survey of 350 employees, and a review of 300 public deployments, as Fortune reported. It is a narrow definition of failure drawn from a small sample, so treat it as a direction, not a base rate.
RAND’s 2024 study of AI project failure says that by some estimates more than 80% of AI projects fail. RAND attributes that figure to outside sources, not to its own research. Its own work was 50 practitioner interviews, and the root causes it found were organizational: leaders who do not state the problem or the success metric, and priorities that shift mid-project.
Executives report the same pattern. An IBM study of 2,000 CEOs, fielded from February to April 2025, found that 25% of AI initiatives had met their ROI expectations and 16% had scaled across the enterprise. McKinsey’s State of AI 2025, a survey of 1,993 participants, found that 39% attribute any EBIT impact to AI, and most of those say it is under 5% of EBIT.
Taken together, the numbers point at the plan, not the model. The pilots that stall are usually the ones where nobody agreed in advance what scaling would require.
What should an AI pilot plan prove in 90 days?
A pilot has to earn three proofs before it earns a budget. If one is missing, scaling multiplies the gap.
- Value. The workflow beats its baseline on a metric the business already tracks, such as hours, error rate, cycle time, or cost per ticket.
- Data. The data the pilot ran on is available, clean enough, and accessible without a manual workaround that will not survive scale.
- Ownership. A named person runs the workflow on a Monday morning, owns the outcome, and has the authority to expand or stop it.
The data proof deserves the most suspicion. Gartner predicts that through 2026, organizations will abandon 60% of AI projects that are unsupported by AI-ready data. The same release says about 63% of organizations either lack the right data management practices for AI or are not sure they have them. Gartner’s release gives two different sample sizes, so use the direction, not the decimals.
Not sure which of the three your organization is weakest on? The free 60-second assessment scores it before you pick a pilot.
Days 1 to 30: choose the pilot and set the bar
The first month produces no AI output. It produces the rules the pilot will be judged by.
- Shortlist two or three workflows where the data exists, the process repeats, and the impact can be counted. Invoice handling, demand forecasting, and support triage qualify. “Improve customer experience” does not.
- Pick one. Our guide to scoping an AI project covers the trade-offs when stakeholders disagree.
- Record the baseline before anything changes. If you cannot measure the before, you cannot prove the after.
- Write the success threshold as a number and a date. Example: cut first-response time by 30% by day 75.
- Name the owner and the person who can say stop.
Days 31 to 60: run a scoped pilot
Run one workflow with one team on one dataset. Scope it tightly enough that a failure is informative rather than catastrophic. If the pilot cannot produce a result inside 30 days, the scope is too wide.
Track three kinds of metrics from day one: performance (accuracy, speed, throughput), adoption (who uses it, how often, where they stop), and business impact (time saved, cost removed, revenue touched). Adoption is the one most pilots skip, and it is where most of them quietly die. Our post on how to measure AI ROI covers the impact side.
Keep a decision log. Write down what worked, what did not, and what you would change. The scale decision in the final phase should rest on that record, not on who argued best in the review meeting.
Days 61 to 90: decide whether to scale, change, or stop
This is the phase most pilots never reach, because nobody scheduled it. Put the decision meeting on the calendar on day one.
Governance belongs in this phase, not after it. McKinsey’s State of AI Trust in 2026, published March 25, 2026 from about 500 organizations, put the average responsible AI maturity score at 2.3. Organizations with explicit ownership of responsible AI averaged 2.6, against 1.8 for those without a clearly accountable function. The report does not tie scores to pilot outcomes, so the link to scaling is our inference. The ownership gap is still a cheap one to close.
Use a scale gate. Each row is a yes or no, and the answer drives the action.
| Gate question | If yes | If no |
|---|---|---|
| Did the pilot beat its baseline by the pre-set threshold? | Continue to the next question | Stop, or change the use case and rerun a 30-day pilot |
| Does the data path work without manual workarounds? | Continue | Fix the data layer first. Do not scale on a workaround |
| Is there a named owner with authority to expand or stop? | Continue | Assign one before any budget is released |
| Are access, monitoring, and escalation rules written down? | Scale to the next team | Write them first. Then scale |
Skills are the last check. If the people who will operate the workflow have not been trained, adoption will stall at the second team. Plan a structured training path, not a one-time workshop.
How do platforms and services package the first 90 days?
Disclosure first: Levos is a sister company of Elevates.AI under common ownership, so weigh this example accordingly. We include it because it shows one way a vendor structures a 90-day window around evidence.
Levos sells a 90-Day AI Impact Audit. As listed on its pricing page on 2026-10-20, it is a $25,000 fixed-fee engagement, credited toward year one if the customer signs an annual subscription afterward. It pulls signals across a defined subset of your stack, runs a controlled cohort analysis that adjusts for tenure, role, and tool stack, and returns confidence-scored AI ROI by team with audit logs. The output is a live AI Impact brief and scorecard built from connected data. The 90-day clock starts when data is connected, not at contract signature.
That is a different job from the pilot above. An audit measures the impact of AI your teams already use. A pilot tests one new workflow before you commit to it. The shared idea is the one this post argues for: a fixed window, a stated measure, and a decision at the end. Whatever you buy, ask what the 90 days will prove and who signs off on it.
What mistakes break an AI pilot plan?
Three account for most of the damage.
- Starting with the tool. Buying first and hunting for a problem later reverses the plan. Begin with the outcome.
- Skipping the data check. If the data is fragmented or locked away, no tool compensates. Test the data path in the first 30 days.
- Treating the pilot as the finish line. A successful pilot earns a scale decision. It does not make one.
For the longer view of why pilots stall, see our posts on the 10 percent problem and the missing middle layer between pilot and production.
What should you do this week?
Write down the one workflow you would pilot, the number it must move, the date you will decide, and the person who owns it. If you cannot fill all four in an hour, you have found your first gap. Start with the 60-second readiness assessment and you will know which gap blocks the others before you spend anything.
Then put the day-90 decision meeting in the calendar. A pilot with a decision date is a plan. A pilot without one is a hobby with a budget.
Frequently Asked Questions
What is an AI pilot plan?
An AI pilot plan is a time-boxed commitment, usually 90 days, to test one AI workflow against a stated baseline and decide whether to scale, change, or stop. It names the use case, the success threshold, the owner, and the decision date before any tool is deployed.
How long should an AI pilot run?
Most AI pilots should run about 90 days in total, with roughly 30 days of live piloting inside that window. Anything that cannot produce a result in 30 days of live use is usually scoped too broadly. A fixed window also forces a decision before attention drifts.
What are good AI pilot success criteria?
Good criteria are a number and a date tied to a baseline you measured before the pilot, such as reducing first-response time by 30% by day 75. Add an adoption measure, because a pilot that people stop using has not proved anything about scale.
Why do AI pilots fail to scale?
Studies from MIT, RAND, IBM, and McKinsey point to organizational causes more than technical ones: unclear success metrics, shifting priorities, data that is not ready, and no owner for the next phase. The failure rates differ by study and method, so read each with its limits.
What is a scale gate for an AI pilot?
A scale gate is a short set of yes or no questions answered at the end of the pilot: did it beat the baseline, does the data path work, is there a named owner, and are access and monitoring rules written down. All four yes answers justify scaling.
Be the First to Discover New AI Insights
Follow Elevates.AI on Google to stay updated with the latest AI readiness assessments, governance frameworks, implementation guides, buyer's guides, and enterprise AI best practices.
Follow Elevates.AI on GoogleThe AI Readiness Brief
Every Week, receive practical enterprise AI strategies, implementation frameworks, governance updates, and expert insights—all delivered in a 5-minute read.
| ✓ Enterprise AI Strategy | ✓ AI Readiness Frameworks |
| ✓ Governance & Compliance | ✓ Exclusive Guides & Resources |
Join 500+ AI Professionals
Enter your work email below to receive one high-value email every week. No spam. Unsubscribe anytime.