The short answer
Pick the process that is done many times a week, follows a pattern a new employee could learn in a day, lives in systems you can already get data out of, and has an owner who is fed up with it. Build the smallest version that removes the worst of the re-keying, measure the hours it gives back, and run it for a month with a person checking the output before you widen it.
Do not pick the project that talks to customers first, and do not pick the one whose data is on paper or in people’s heads. Those are the two that fail, and a failed first project is often the last.
Four tests a first project has to pass
A candidate that passes all four is a good first project even if it is small. One that fails any of them will take longer, cost more and prove less, however impressive it sounds.
- Frequency. It happens daily or weekly. A quarterly process cannot be learned from or measured in a sensible time.
- Pattern. Someone can describe how it is done today in a page. If every case is a negotiation, it is not a first project.
- Data. The inputs are digital and reachable: an inbox, a system with an export or an API, a shared drive with a structure. Paper is a scanning project first.
- Consequence. A wrong output is caught by a person before it costs money or reaches a customer, at least for the first three months. Drafts, summaries, classifications and internal routing pass this test; sending, paying and promising do not.
The two projects you should not start with
The customer-facing agent. It is the most visible use of AI and the one that gets pitched first. It is also the one where every weakness in your data, permissions and approval rules is exposed to the public. Do it second, once an internal project has shown you how the AI behaves on your information and your policy exists. The businesses that put an agent on the website first spend the next six months apologising for it.
The project that needs data you do not have. Forecasting from sales history you never captured, a knowledge assistant over documents that are on paper, quality control from cameras not yet installed. Each is a real opportunity and a bad first project, because the AI part cannot start until the data part is done, and the data part is a project of its own with none of the excitement. Do the data work, honestly labelled as data work, and do the AI project the quarter after.
Where first projects usually come from
Ask the team which task they would most like never to do again. The answer is nearly always some form of re-keying: information that arrives in one place (an inbox, a form, a document) and has to be typed into another (the CRM, the job system, the accounting package). Reading the arriving thing, extracting what matters and creating the record in the other system is the first project in most businesses, whatever the industry.
The second most common answer is finding things: the policy, the price, the precedent, the procedure. A knowledge assistant over the documents you already hold is an excellent first project when the documents are digital and reasonably organised, and a poor one when they are not.
Define done before you start
Write two sentences before any tool is chosen. The first says what the project does: “reads incoming supplier invoices, extracts the fields, creates the draft bill in the accounting system and flags anything it is unsure about”. The second says how you will know it worked: “accounts spends the time it used to spend keying on checking, and the flagged rate falls month on month”. If you cannot write the second sentence, you are not ready to start, and no amount of demo will fix that.
Measure honestly. Hours given back, errors caught, days shortened. Not “innovation”, not a percentage somebody made up for the board. A small true number is worth more than a large invented one, because it earns the second project.
Keep a person in it for the first quarter
For the first three months, someone checks the output before it does anything with a consequence. That is not a lack of confidence in the technology; it is how you find out where the technology is wrong about your business, and it is how the team comes to trust it. The checking gets lighter as the record shows it can. Widen the automation in steps, and keep the record, because the day something goes wrong the record is what tells you how.
Three first projects that usually work
Each passes the four tests in most businesses, each pays inside a quarter, and each teaches the organisation something it needs before the customer-facing work: what the AI is reliably right about, where it is unsure, and how much checking a person actually has to do.
- Supplier invoices and statements read, matched and entered as draft bills, with the uncertain fields flagged for accounts to check.
- Meeting notes and call recordings turned into CRM updates and follow-up tasks that a salesperson approves at the end of the day.
- An internal assistant over the staff handbook, procedures and product information that answers with the source shown and says when it does not know.
What a first project should look like in scope and time
Weeks, not months. One process, one owner, one measure, a person checking the output, and a written design short enough to read in a sitting. If the proposal for a first project runs to a programme with phases, it is not a first project; it is a strategy engagement wearing the wrong label, and the two should be priced and run differently. The right first project is small enough that a failure costs you a few weeks and teaches you something, and a success earns the second project on its own numbers.
Where to go from here
The Executive AI Workshop on this site exists to get your leadership team to this decision in half a day: which project first, which two to refuse, and what done looks like. If you want a signal before booking anything, the AI Opportunity Score below ranks the likely candidates from twelve questions.
Published 12 September 2026 · Be AI