Four-week AI pilot kit: one task, one owner, one decision

Gartner reports that by the end of 2025, at least 50% of generative AI projects were abandoned after proof of concept, due to poor data quality, inadequate risk controls, escalating costs or unclear business value. Gartner also points to a pattern behind it: when organizations deploy GenAI everywhere at once, they spread resources across low-impact work, and without success metrics the projects lack measurable business value.

Fortune, reporting on MIT NANDA's "The GenAI Divide: State of AI in Business 2025," wrote that about 5% of AI pilot programs achieve rapid revenue acceleration, while the vast majority stall, delivering little to no measurable impact on P&L. Asked about the startups that do well, lead author Aditya Challapally said, "It's because they pick one pain point, execute well, and partner smartly with companies who use their tools." This kit is built for that kind of focus: one task, one named owner, a baseline, three weekly checkpoints and a decision at week 4.

Week 0: set up

Do this before anyone opens the tool.

1. 01

   Pick one task.

   Choose something your team does often enough to measure in a month, such as drafting quote follow-ups, sorting inbound email or summarizing job notes. Write it in one sentence. If the sentence needs an "and," you have two pilots. Run one.

2. 02

   Name one owner.

   Gartner lists "establishing clear governance and ownership" among the basics of responsible AI. In a small business that means one person, by name, who reviews the work each week and makes the call at week 4. Fortune also reported that giving line managers the job of driving adoption was a key factor in the MIT research, so pick the manager closest to the work.

3. 03

   Record the baseline.

   Gartner advises organizations to "identify and continuously track specific, measurable business outcomes." For one task, that means writing down how things work today:

• How long the task takes each time
• How often it happens per week
• How often it needs fixing or redoing
• What it costs now, in staff time and any tools

Use your own numbers, even rough ones. Without them, week 4 turns into a debate about impressions.

1. 01

   Set a cost cap.

   Gartner warns that "rising costs kill projects even when they're technically successful" and recommends GenAI FinOps practices from day 1. For an SMB, that can be as simple as a dollar limit for the four weeks, covering licenses, usage and any outside help, plus a rule that spend gets checked every week.

2. 02

   Write the keep and kill criteria now.

   Decide what result means keep and what result means kill before any output arrives. Once people have seen a demo they like, it gets hard to judge it fairly.

Weeks 1 to 3: weekly checkpoints

Each week, the owner holds a 30-minute review of what the tool or agent produced. The tool reports its interim work, a person looks at it, and the next week's work changes based on what they find. That lets you steer the work as it goes, instead of handing AI one big goal and waiting for the finished result. (h/t @emollick, who made the case on X for building in checkpoints where the AI reports back so you can guide it.)

Log each review in the template below: what the tool produced, what the owner changed, issues found and spend to date.

1. 01

   Week 1: accuracy on a small sample

   Run the tool on a small batch of real work. A person checks every single output before anything goes to a customer or into a system of record. In the review, look at how often the output was usable as is, what kind of mistakes it made, and whether each one was a small fix or wrong at the root. Adjust the instructions, examples or data access, then carry on.

2. 02

   Week 2: real volume and edge cases

   Move to the normal weekly volume. Keep the human review rule in place, and pay attention to the unusual cases: the odd request, the missing field, the customer who writes three paragraphs instead of one line. Check spend against the cost cap. If spend is on track to pass the cap before week 4, raise it now.

3. 03

   Week 3: compare against the baseline

   Measure the same things you recorded in week 0: time per task, volume, error or rework rate and cost. Then ask the people who use the output what they think. Gartner recommends you "pilot UX with actual users and iterate based on feedback." Whether it helps their working day counts as much as the numbers.

Week 4: keep, kill or extend

The owner makes the decision against the criteria written in week 0. Nobody rewrites them now.

1. 01

   Keep

   if the results meet the keep criteria and spend stayed under the cap. Write down what changes in the process and who maintains it.

2. 02

   Kill

   if the results hit the kill criteria, spend passed the cap or the tool needed so much correction that it saved no time. A kill is a useful result. You learned it on a small, fixed budget, and you have the notes to show why.

3. 03

   Extend

   only once, by two weeks, with a named reason. "Data access arrived late" is a reason. "We'd like to see more" is not.

Record the decision and the reason in one or two sentences. Then pick the next pilot.

The one-page template

Copy this into a doc or a spreadsheet. Fill in every blank before week 1, except the checkpoint log and the decision.

Sources

• Gartner: Why 50% of GenAI Projects Fail: And How to Beat the Odds, Arun Chandrasekaran, January 26, 2026
• Fortune: MIT report: 95% of generative AI pilots at companies are failing, Sheryl Estrada, August 18, 2025
• Ethan Mollick on X: post on building checkpoints into AI work

Reach us at transform@consultiply.ai