Is the output right?
How often the output is correct or acceptable without edits, on the test set and then in real use.
You never commit to more than the next stage. Each one ends with something concrete in your hands and an honest recommendation about whether to continue.
Map where time goes and which tasks suit AI.
Ranked shortlistAgree the workflow, controls and success measures.
Written design & quoteBuild in short increments, connected to your tools.
Working system in testRun your own examples and fix what fails.
Test resultsRoll out carefully, often with every output reviewed at first.
Live systemTrain your team and hand over everything they need.
Docs, runbook, access listUnderstand what your business is trying to achieve, how the work is done today, and which tasks are realistic candidates for AI.
Agree exactly what will be built, how it fits the way your team works, and how you'll know whether it's working.
Build in short increments, showing working progress regularly rather than disappearing until the end.
Timeline set in the written quote
Run the agreed set of real examples, including the awkward ones, and fix what fails before anything is used for real work.
Put the system into real use carefully, usually with one team first and every output reviewed until you're confident in it.
Make sure your team can run, troubleshoot and change the system without depending on any one person.
A good first project builds confidence and shows your organization how AI systems behave, without betting anything important on it.
Frequent tasks give enough volume for the time saved to matter, and enough examples to test with.
These are the things current AI models handle best.
A person can check the output before it affects a customer, a payment or a record.
The information needed lives in systems that can be connected to safely.
A person on your side cares about the outcome and will use the result.
Measures are agreed in the design stage and compared with how the work is done today, so results can be judged honestly.
How often the output is correct or acceptable without edits, on the test set and then in real use.
How long the task takes end to end, and how much human time it needs, compared with today.
Whether the people it was built for actually use it, and what they would change.
Start with a conversation about how the work is done today. You'll leave with an honest view of whether AI is worth pursuing for it.