The best AI teammate starts with a job description.
What building a command centre taught us about turning an impressive AI answer into useful, repeatable work.
Define the work before you configure the agent.
A good answer still needs somewhere to go.
A weekly report rarely begins with a blank page. It begins with updates scattered across conversations, a spreadsheet someone maintains, and a few decisions that only the project lead remembers. Writing the report is only one part of the job. Finding the right context and knowing who can confirm it take just as much care.
The opportunity is to reduce the time a project lead spends collecting updates and chasing missing owners. To do that, the AI teammate needs a specific job: assemble the draft, identify what is missing and return it to the lead. “Help with operations” leaves too much of that contract unstated.
Building the Command Centre has made the importance of a defined role much clearer to us. An agent with a broad instruction to “help with operations” is difficult to assess. A role that prepares a weekly update, flags missing owners and hands the draft to a named reviewer gives the team something concrete to evaluate.
Take this with youA useful first role has one repeatable task, one owner and a result the owner can check.
From a vague request to a defined role
Define the job
Prepare a weekly project update for the project lead. Name the owner and what a useful update includes.
Under the hood
Keep purpose, ownership and working instructions in the role definition. Include the expected output and the person responsible for reviewing it.
Equip the role
Choose the project updates and connected tools the role needs. More access is not automatically more useful.
Under the hood
Bind selected skills, connections and knowledge to the role. Check that the required connection is actually usable.
Review the work
The project lead checks owners, deadlines and blockers before sharing the draft.
Under the hood
Define a review point around the consequential action. Keep preparation separate from sending or publishing.
Learn and adjust
Review missing context and corrections, then update the role before the next reporting cycle.
Under the hood
Use representative task results and recorded activity to refine instructions, tools and review boundaries.
Make the job visible in the product.
Our setup flow brings together purpose, working instructions, connected tools, review and deployment. These belong together because the work depends on their combination. A well-written instruction cannot compensate for an unavailable connection, and a working connection cannot decide what the business considers a good result.
That also changes the conversation between business and engineering. The business owner describes the handover: “Give me a draft with progress, blockers and missing updates.” The technical owner can then decide which tools and information support that handover. Both are working toward the same observable outcome.
Under the hood +
Keep the job description, tool selection and review points together. When the work changes, update the role deliberately so the next task uses the revised instructions and connections.
Start small enough to learn something.
A bounded first role is easier to improve. Try a normal reporting cycle, one with missing information, and one with contradictory updates. Ask whether the role makes uncertainty visible. A polished report that invents a deadline is less useful than a draft that asks the lead to confirm it.
Assess the handover itself. Can the lead check the report without reopening every conversation? Are missing owners easy to find? Those questions connect the role to the effort it is meant to remove, and give the next reporting cycle something specific to improve.
If you are choosing your first use case, look for work that repeats often enough to learn from, has accessible inputs and has a clear reviewer. Those conditions create a better starting point than choosing the task with the most impressive demo.
Take this with youWrite the job description first. Use real examples to decide whether the role is ready for more responsibility.
What should an AI agent job description include?
Name the job, the person responsible, the information and tools it needs, the expected output and the decisions that need review.