AI and automation
AI agents
Good-fit situations
- Staff repeatedly answer questions from the same approved material.
- Incoming requests or documents need consistent classification and routing.
- A person spends substantial time extracting the same fields from documents.
- Teams need a first draft or summary that will always be reviewed.
- An existing chatbot lacks reliable sources, boundaries, or a clear owner.
What you receive
- A use-case and risk assessment
- Written boundaries for what the agent may do, must escalate, and must never claim
- Source, permission, and data-handling requirements
- A working prototype with a human-review path
- An evaluation set, acceptance criteria, and a plan for ongoing checks
- Documentation and a clear way to pause or disable the system
Process
How the work proceeds
Find a narrow task
Identify work with enough volume to justify the build and a scope tight enough to evaluate.
Define the boundaries
Sources, permissions, failure modes, and review requirements are written down before anything is built.
Prototype and test
The agent is tested against representative examples, including the ones it should refuse.
Measure before expanding
Accuracy, exceptions, and operational value are measured before the agent is given more authority.
Where an agent is the wrong answer
An agent should not begin with irreversible authority, an undefined task, no accountable owner, or no practical way to review its work. In those cases, the useful engagement may be a process redesign rather than an AI build.
Related
The rest of the practice
- Workflow automation
Routine work that moves without manual re-entry.
- Custom internal applications
Focused software for a process products cannot support.
- Data and reporting systems
One reliable view of the numbers behind decisions.
Start with the specifics.
Describe the task, the systems involved, and who is responsible for it today. That is enough for a first conversation.