What happens every day
Frequent questions, documents and reference answers.
How to prepare an AI project
What does your AI projectneed to get started?
A task with clear boundaries, useful examples and people available to make decisions. Organize what is needed for the next step and make clear what still needs to be resolved.
From the objective
to the next step.
Define the work, the desired improvement and the boundaries of the test. This helps avoid spending effort on what does not need to be built yet.
Sample starting plan
Look up guidance
for customer service.
- What to evaluate
- Find the current guidance in less time, with a source the team can check.
- Materials needed
- Everyday questions and approved guidance, with the version and responsible person identified.
- Test boundaries
- Suggest a response for human review. Without sending it to the customer or changing records.
Before testing
The person responsible for the content needs to clarify which guidance applies when two versions conflict.
Examples that explain
how the work is done.
Start with a sample the team can examine. For each input, identify the expected response and the rule for judging it.
What falls outside the norm
Conflicting versions, missing information and situations with no reliable answer.
What each case teaches
Why the response is appropriate or requires someone’s help.
What context needs to accompany the data?
Record the source, the responsible person, the purpose and permission for use. Identify older versions and situations not represented. To discuss the method, a reduced dataset or prepared examples may be enough.
Define how to compare quality, time and review effort. Keep evaluation cases separate from those used to adjust the solution.
Prepare the comparisonWho explains.
Who decides. Who makes it possible.
Agree on availability and how to resolve disagreements. One person can take on more than one role, provided they have the capacity and authority.
The people doing the work
Explain the routine, identify exceptions and check whether the output is useful.
The person responsible for the process
Confirms rules, resolves business questions and accepts deliverables.
The people managing data and systems
Confirm approved tools, access permissions and technical dependencies.
Reading, suggesting and acting
call for different decisions.
List only the actions needed for the test. Looking up guidance does not, in itself, require permission to change the system.
What should you agree on about the test environment?
Define where the material will be stored, which tools may process it, who will have access and how to revoke that access. Record relevant versions, settings and limitations.
Agree on how to stop the test and whom to notify about an incident. Using the solution in everyday work requires a separate decision on integration, responsibilities and operating conditions.
What needs to be resolved
before the next step?
You don’t need to get the entire company ready before starting. Work with the person responsible for the task to distinguish what prevents testing from what can wait.
- Task, objective and boundaries of the work defined.
- Examples with context and permission for use.
- People available to answer questions and evaluate.
- Access and environment suitable for the test.
- Outstanding issues with a responsible person and a next step.
Resolve first
A conflicting rule that prevents judging the response.
Plan for later
An integration that the initial test does not use.
A reference for further reading
The NIST AI RMF 1.0 · Govern 2.1 calls for documenting responsibilities and communication in AI risk management.
Talk to Gaulett
What does your idea need
to move forward?
Bring the objective and what your team already has. Let’s discuss the conditions needed for the next step.