Most teams pick their first AI workflow the same way: whatever looked best in the last demo. That is how a capable tool ends up on a task that happens twice a year, or on one where a single wrong answer reaches a client. The first workflow matters more than its size suggests, because it decides whether the team trusts the second one. Choosing it well takes about an hour and a short list.
The short answer
Start with a task that repeats every week, takes real hours, follows rules a new employee could learn from one page, and is cheap to get wrong because a person checks the output before it leaves the business. Score a short list of candidates on those five points and start with the highest score. The first workflow should prove that AI saves time on your own work. It does not need to show everything AI can do.
What decides it
- Frequency. A daily or weekly task pays back quickly and gives enough cases to judge the results. A quarterly task gives four tries a year.
- Hours. Count the time the task takes today, across everyone who touches it. A task that takes 10 minutes a week is not worth a project.
- Clear rules. If the steps fit on one page, a model can follow them. If every case needs judgment nobody can explain, start elsewhere.
- Cost of a mistake. Drafts a person reviews are a safe start. Anything that sends money, deletes records or reaches clients without review is not a first workflow.
- Data access. The inputs should already sit in a tool you use, such as email, a shared drive or a CRM, so the first step is a connection and not a data project.
Anthropic's guide to building agents gives builders the same advice from the other side: find the simplest solution that works, and add complexity only when it is needed.
A worked example
Example, for illustration. A 15-person logistics company lists five recurring tasks and scores each from 1 to 3 on the five points, so the top possible score is 15:
- Replying to "where is my delivery" emails: 13. Daily, hours every week, clear rules, a dispatcher approves each reply. Data access scores lower because the shipment system needs a connection.
- Summarizing customer calls: 12. Frequent and low risk, but each call takes only minutes.
- The weekly management report: 11. Clear and safe, but about 2 hours a week.
- Matching invoices to delivery notes: 9. Real hours, but mistakes touch money and the notes are scanned.
- Onboarding new drivers: 9. Every case is different.
The status emails win. Three dispatchers spend about 1 hour a day each on them, so about 15 hours over a 5-day week. At an assumed internal cost of 150 NIS an hour, that is 2,250 NIS of time a week. That number is an assumption, not a promise, but it gives the first month something to be measured against.
Because the connection is the weak point, the first version can draft replies from the status a dispatcher pastes in. The connection to the shipment system comes once the drafts prove useful.
How to decide, step by step
- List five recurring tasks with the people who do them, in their words. Microsoft's scenario library lists common tasks by department and industry if the list is hard to start.
- Write down the hours each task takes in a normal week, across everyone involved.
- Score each task from 1 to 3 on frequency, hours, clear rules, the cost of a mistake and data access.
- Pick the highest score, and check that a person can approve the output before anything leaves the business.
- Write the one-page rules for that task: inputs, steps, what a good output looks like and what never goes in.
- Agree on one measure for the first month, such as hours spent or the share of drafts sent with light edits.
Mistakes that cost the most
- Starting with the most impressive task instead of the most frequent one.
- Automating a process nobody agrees on. If two people do the task differently, settle the method first.
- Leaving out the approval step. Tools do not always add it for you: in Zapier's AI step, Require approval before running is off by default for each tool. Turn it on for anything that writes or sends.
- Checking only some of the output. In McKinsey's survey from March 2025, only 27 percent of respondents said their organizations review all gen AI content before it is used. In a first workflow, a person reviews everything.
- Choosing without a measure, so nobody can say after a month whether it worked.
When outside help is worth it, and how to choose it
Most teams can run this scoring on their own in an afternoon. Outside help is worth it when the candidates touch several systems, when the team cannot agree which task matters most, or when nobody has the hours to map the work. If you bring in an AI consultant, a few signs of a good fit:
- They start from your workflows and your people, not from a tool demo.
- They leave you a written, scored list and a first step you could run without them.
- Anything that reaches clients or money gets an approval step from the start.
- They set things up in your accounts, so you own the work.
- They say when the right answer is to wait.
The AI consulting page describes that kind of mapping: a short audit of real workflows that ends with a scored roadmap and a first step. What drives the cost of building it is covered in the guide to AI implementation costs.
This week: list five recurring tasks with the people who do them, score them, and write the one-page rules for the winner. Success is a single page that someone else on the team could follow.