How to start an AI project: choose one job, write it on one page, and test the system on your own past work before anyone relies on it. A first AI project for a small business should be that small. It is one job that happens many times a week, with someone who can check it, ending in something you own, and it takes a few weeks to several months, depending on the system. Below are the one-page brief with an example filled in, the stages a first project moves through, what moves the date, and what the often-quoted failure numbers count.
What goes on the one-page brief?
Ten lines, written by the owner before any builder starts.
The example is an invented 30-person accounting firm. The brief is short on purpose: if a line will not fit on one page, the job is too big for a first project.
| Line | What to write | Example |
|---|---|---|
| 1. The job | One sentence, the way you would say it to a new hire | “Clients email us tax documents. Dana renames each one, files it in the client’s folder and logs it.” |
| 2. How often | About how many a week, and when the peaks come | About 600 a week, three times that in March |
| 3. Where it starts and ends | The inbox, app or folder it starts in, and where the result has to land | The shared client inbox; the document system and the practice software |
| 4. Who checks it | The person who does it today, by name | Dana |
| 5. Today’s baseline | What the job takes now, counted over one normal week | About a day and a half of Dana’s week; 3 misfiled documents found last month |
| 6. What done means | The bar, set before the test | It files 28 of the last 30 documents the way Dana did, and flags the other 2 instead of guessing |
| 7. What it must never do | The lines it never crosses | Delete a document, email a client, or file one it is unsure of without flagging it |
| 8. The decision it feeds | What a person decides with the result, and who | Dana approves each filing for the first month; a partner reviews the misses |
| 9. What is left out | What this first project does not touch | Reading the numbers inside the documents |
| 10. Where to start | The smallest piece that proves it | W-2s and 1099s only, the two most common forms |
Line 5 is the one owners skip, and without it nobody can say later whether the project paid. Is AI automation worth it? shows how to turn that baseline into a payback number.
What is an AI proof of concept?
A short test that shows an AI can do the hard part of one job on your own examples, before anyone builds the rest.
A pilot is the step after it: the system runs on live work while a person approves every result. In a small company they can be stages of one project: stage 1 below does the work of a proof of concept, and stage 3 the work of a pilot.
What happens at each stage?
Three stages, whoever builds it, each ending in something you hold.
| Stage | What happens | What your side does | What you hold at the end |
|---|---|---|---|
| 1. Build and test on past work | The system is built and run on your last 30 or more cases, and its answers are set beside what your staff did | Pull the cases and write the right outcome next to each | A count of how many it got right, and a list of the misses |
| 2. Handover | The finished system is delivered and written up | Name the person who owns it, and learn how to pause it | The code, the accounts and a one-page guide, in your company’s name |
| 3. Live, with a person approving | It runs on new work, and a person approves each result before it goes anywhere, then moves to spot checks | Approve or correct each result, and note every correction | A record of corrections, which should shrink week by week |
Stage 1 takes longer to prepare than to run. In our own test of four AI models on 80 office jobs, each model’s median answer took under 5 seconds a job; the work was in writing the jobs and their right answers first. Stage 3 is where most of the trust is earned. What happens when AI gets it wrong covers the checks worth keeping after it, and who owns the AI you paid for covers what stage 2 should hand you.
How long does a first AI project take?
It depends most on four things.
These move the date more than anything else:
- How many systems it reads from and writes to, and whether each one has a way in for software (an API) or only a screen.
- How clean the past cases are: whether the right outcome is written down, or lives in someone’s head.
- How many odd cases and unwritten rules the job has. A rule a person applies without thinking still has to be found and written down.
- How fast your side answers questions and checks results. In RAND’s interviews with experienced AI engineers, 14 of 50 said senior leaders often underestimated how long it takes to make an AI model that solves the business problem.
Shorter advice is also around. Asked a nearby question in our tracker (an owner buried in admin, asking where to start), four of six AI assistants said to try one tool on one task for a week or two: Claude, GPT-6 Luna and both Perplexity models. The SBA’s guide for small businesses says much the same: “AI is relatively new, so start small.” That is the right length for testing a ready-made product, and a good step before any build. A built system takes longer because it has to fit the systems and rules a product did not. None of the six used the words pilot or proof of concept.
How do first AI projects usually go?
The numbers quoted most often each count something different, so read the definition before the figure.
| Source | What it counted | Figure |
|---|---|---|
| MIT NANDA, The GenAI Divide (July 2025, preliminary) | Organizations looking at custom or vendor-sold AI tools for a task, and how far they got | 60% evaluated one, 20% reached a pilot, 5% reached production |
| Gartner (July 2024) | A forecast for generative AI projects by the end of 2025, a date that has now passed | At least 30% abandoned after the proof of concept |
| Gartner (February 2025) | A forecast, through 2026, for AI projects without data ready for AI | 60% abandoned |
| RAND (August 2024) | Interviews with 65 data scientists and engineers with five or more years of experience | 84% named a leadership cause as a main reason projects failed |
MIT’s authors call their figures preliminary and directionally accurate, drawn from interviews rather than company reports; the copy we read is hosted by MLQ.ai. The two Gartner rows are forecasts, not counts. The often-repeated “more than 80 percent of AI projects fail” appears in RAND’s report only “by some estimates”, not as its own finding. The reasons the four give overlap: unclear goals, data that is not ready, and tools that do not fit how the work is done. Lines 5, 6 and 8 of the brief are there for those reasons.
What does Airu Labs do on a first project?
We build it.
Airu Labs builds custom AI systems for small and mid-size companies, wired into the software they already use. A first project starts from one job that costs the business time, often written on a brief like the one above. You get the system when it is built, then check it on your own data before relying on it, which is stage 3 above, and you own the code and the accounts it runs on. Choosing which job comes first is covered on AI consulting for small business.
What does a stop look like?
A result below the bar from line 6, with misses that share no single cause.
Decide go, adjust or stop after stage 1, with the brief in front of you:
- Go: it meets the bar from line 6 on the cases that matter.
- Adjust: it misses, but most misses share one cause you can fix, such as a rule nobody wrote down or a bad scan.
- Stop: the misses are spread across every kind of case, or checking its work takes as long as doing the job.
On the accounting firm’s brief the bar is 28 of 30. Say the test files 22 correctly, and 6 of the 8 misses are scanned forms the system could not read. That is an adjust: send scanned forms to Dana and test again. If it had filed 15, with misses on every kind of document, that is a stop, and the firm learned it on the smallest piece of the job.