Insights
How to choose your first AI workflow
A practical way for business leaders to pick a first AI project that is small, checkable and genuinely useful, plus the warning signs of a poor first choice.
Most organizations are past asking whether AI could help. The harder question is where to start. A good first project builds confidence and teaches your team how to work with these tools. A poor one burns budget, produces a demo nobody uses, and makes the next proposal harder to approve.
Here is the approach we use with clients to pick a first workflow.
Start with tasks, not technology
Skip “how can we use AI?” and ask a narrower question: which repetitive tasks take our team the most time, and which of those are tedious but not difficult?
Good candidates usually involve reading, sorting, extracting or drafting:
- Pulling the same fields out of incoming documents and typing them into a system
- Reading requests in a shared inbox and deciding who should handle each one
- Turning a week of notes or tickets into a status summary
- Looking up how a policy or procedure applies to a specific situation
These tasks share a pattern. A person can do them well, but they happen often, they follow recognizable rules, and the work is easy to check.
Score each candidate on four things
Once you have a short list, rate each task honestly.
- Frequency. How often does it happen? A task done fifty times a week is a better pilot than one done twice a month.
- Checkability. Can someone quickly tell whether the output is right? If verifying the result takes as long as doing the work, the benefit shrinks.
- Cost of a mistake. What happens if an error slips through? Start where mistakes are caught easily and fixed cheaply, not where they reach customers or regulators directly.
- Where the work lives. Does the task start and end in systems you can connect to, such as an inbox, a shared drive, a CRM or a database? A workflow that requires copying results by hand will not stick.
The best first project scores well on all four. It is frequent, easy to check, low risk and close to systems you already use.
Define success before anything is built
Agree on what good looks like while the project is still on paper. Gather a set of real past examples, including the awkward ones, and decide how results will be judged. For example: the right fields are extracted in most documents, anything uncertain is flagged for a person to check, and review takes a fraction of the time manual entry did.
This makes the pilot a fair test, judged on your own work. It also gives you a clear decision at the end: expand it, adjust it, or stop.
Keep a person in the loop
For a first workflow, design the human checkpoint on purpose. Decide who reviews the output, what they are looking for, and how they correct it. Make that step quick and comfortable, and make it obvious when the system is unsure.
People stay responsible for the outcome. The AI handles the repetitive part and prepares the work for their judgment.
Warning signs of a poor first choice
- The goal is a feeling. “Make us more efficient” or “use AI across the company” cannot be tested.
- Nobody owns the process. If no one can say how the task works today, a pilot will not fix that.
- The data is off limits. If the information cannot be shared with an AI provider under your policies, choose a different task or settle the data question first.
- Mistakes are expensive and hard to spot. Leave those for later, once you have experience and monitoring in place.
- It lives outside every system. If results have to be copied by hand into where work happens, adoption will fade.
Plan for after the pilot
A pilot that works becomes software you depend on. It needs the same care as any other system: access controls, logging, monitoring of quality over time, and someone responsible when models or your processes change. Budget for that from the start so a successful pilot does not turn into an unsupported tool.
A simple next step
Write down three repetitive tasks your team handles each week. For each, note how often it happens, who checks it, and which systems it touches. That short list is usually enough to identify a strong first workflow, and it is a good starting point for a conversation with us or anyone else helping you plan.
