ShapedSI Journal · Applied AI
AI readiness: choose a first pilot you can judge
AI readiness is specific to a job. A team may be ready to test an internal drafting task while still needing work before connecting customer records or changing a business system. Choose one workflow, inspect the information it depends on and decide how the people responsible will judge the result.
In this guide
01
Choose a useful, bounded job
Look for a recurring task with a clear input and finish. Drafting a short internal summary is easier to judge than an instruction to transform the whole business. Choose something the process owner understands well and can review during the pilot. Record what the system should prepare and what remains a person’s decision.
Keep the first version small. Limit the sources, the users and the operations. A narrow pilot can reveal whether the idea helps without requiring a large integration project first. If a task has no agreed finish or depends on constantly changing instructions, clarify the process before adding AI.
02
Inspect the information and the permissions
Find the approved source material and check whether it is current and understandable. Identify missing documents, conflicting versions and information that should remain outside the pilot. The tool cannot resolve an internal disagreement about which document is authoritative simply by producing a confident answer.
Confirm who may provide the input and who may see the output. Ask where processing and storage take place, and involve the relevant owner before using sensitive information. Early examples can be fictional or appropriately anonymised. That lets the team test the working method while the data arrangements are still being reviewed.
03
Define the test before the demonstration
Prepare a small, representative set of examples, including incomplete and difficult cases. Write the expected result or the criteria a reviewer will use. Check factual accuracy, missing information and the time needed to correct the output. Keep a record of the manual approach so the comparison is meaningful.
Agree on a stop condition. If the tool repeatedly needs more correction than the task is worth, if the sources cannot support the answer or if the review burden is too high, the pilot may need a different scope. A clear reason to stop is part of a good experiment, not an admission that the team failed.
04
Decide what happens after the pilot
Review the results with the people who will operate the workflow. Ask which examples helped, which failed and whether the handoff was clear. Separate a promising pattern from an occasional impressive answer. Include the effort to maintain the sources and instructions in the decision.
If the team proceeds, name the owner, document the fallback and agree on how changes will be reviewed. If it pauses, keep the findings and the process map. Readiness can improve through clearer source information, better training or a simpler workflow. The first pilot should make the next decision easier to explain.
Bring this to the conversation
A working brief.
- A narrow recurring task
- Approved source information
- A reviewer with time to judge it
- Representative and difficult examples
- A proceed, revise or stop decision
A few useful questions.
Do we need perfect data before testing?
You need information that is adequate for the chosen task and a clear account of its limits. A pilot can help identify which gaps matter.
What if the pilot shows that AI is not useful here?
Keep the findings. A clearer process, source list or handoff may be the more useful next improvement.
Original sources
Sources and further reading.
OpenAI’s agent safety guidance treats approvals and input controls as part of the workflow design. For a first pilot, establish the review boundary before granting a tool the ability to act.
Shape the next step