Most AI projects should not start with a model.
They should start with an audit.
Not a generic readiness survey. A real implementation audit that answers five questions:
- What workflow is actually broken?
- What data can the system trust?
- What actions should AI never take alone?
- What does quality mean?
- What would make the project worth the cost?
The audit turns a messy workflow into a ranked build plan. The point is not to prove AI can be used. The point is to decide where it should be used first.
Start with the workflow
The workflow tells you whether AI belongs there.
Look for repeated decisions, repeated writing, repeated triage, repeated lookup, repeated handoff, and repeated follow-up.
Then ask what happens when the system gets it wrong.
If a wrong answer is annoying, you can automate more aggressively.
If a wrong answer touches money, customers, legal exposure, health, safety, or trust, the system needs review, evals, and escalation.
Map the data
AI systems are limited by source quality.
The audit should identify:
- where the source data lives
- whether it is current
- who owns it
- who is allowed to see it
- what the system should cite
- what the system should refuse to answer
If nobody owns the source, the AI will inherit the mess.
Define the no-fly zone
Every AI system needs a no-fly zone.
Examples:
- refunds over a threshold
- legal advice
- medical advice
- firing decisions
- customer-facing promises
- price exceptions
- production writes
- destructive file operations
The audit should decide what requires human review before the first prototype exists.
Decide what to build first
The first AI project should usually be narrow.
Good first builds:
- support triage
- quote drafting
- document extraction
- internal knowledge assistant
- lead qualification
- customer follow-up
- report generation
Weak first builds:
- “AI for everything”
- autonomous sales agent with no guardrails
- executive dashboard with no source discipline
- chatbot on top of unmaintained docs
The output should be a build plan
A good audit ends with a ranked plan:
- build now
- build later
- buy instead
- skip entirely
That last category matters.
The strongest AI strategy often includes the work you deliberately do not automate.
