What this covers
An AI agent with a short list of allowed tools, logged actions, and a person approving anything important.
Tools
Tool calling, Structured outputs, APIs, Audit logs. The exact stack depends on what you already use.
Where it helps
- Prepare a CRM follow-up for someone to approve
- Collect details for an internal service request
- Draft a task update from project records it’s allowed to read
We agree what “done” means before work starts.
What you get
- An agent with a fixed list of allowed tools
- Logs of every action, and an approval step for important changes
- Source code, setup notes and an agreed definition of done
The final list depends on the agreed scope. The estimate and schedule follow from it.
Data and security
Text the agent reads is treated as untrusted. Permissions are enforced in code, tools are limited, and a person approves anything written to outside systems.
How I work on it
Map the task, decide which tools the agent can use, test it against tricky and failure cases, run it in review-only mode, and expand only when results justify it.
What helps me get started
- A short description of the process and who uses it.
- Sample data, any API documents, and what access is available.
- How much it will be used, who reviews the work, and your release priorities.
A short outline is enough for a first call. Please send sample data, not real records or passwords.
Related services
Read the planning guideQuestions
Can an agent send messages or change records?
Only if we explicitly allow it, with checks and approval in place. It’s best to start with read-only access or drafts.
How do you price the work?
I look at the workflow, the systems involved, the access needed and what “done” means, then send an estimate. If there are big unknowns, we start with a short discovery or a prototype.