The job
What starts it, what it must produce and when it is finished. A job that takes more than a few sentences to describe is split first.
SERVICE / AI AGENTS
AI agents that run recurring jobs and stop where a person decides.
An agent follows one job from start to finish: it checks the sources, works through the steps and leaves the result ready for your team, within written limits on what it may read, prepare and never do alone.
JOBS AN AGENT TAKES ON
Three kinds of recurring job an agent can take off your team.
Each agent is quoted on the hours it frees each month, measured in a diagnosis that is free with an active bag. Paid before the build, used under licence; AI usage is billed to you.
Monitoring and research
Competitor listings, prices, reviews or supplier catalogues checked on a schedule, with a short note when something relevant changes and each claim linked to its source.
Reading documents
Invoices, technical sheets or supplier price lists read as they arrive, the agreed fields copied into a sheet and doubtful values marked for a person.
Sorting incoming work
Emails, forms or marketplace messages sorted by type and urgency, with a reply drafted for someone on your team to send.
WHAT AN AGENT IS
An agent carries out a whole job, then hands it over.
An AI system takes one piece of work at a time from a person. An agent is given the whole job, such as checking competitor listings every Monday: it works out the next step, uses only the tools it was given and stops when it is done or a case falls outside its limits.
- Same steps every time: plain automation is simpler to maintain
- A first draft inside someone's daily work: an AI system
- Variable input and several steps: an agent
LIMITS SET BEFORE IT RUNS
Four limits are written down before an agent does real work.
What it may read
Sources and accounts, read-only wherever that is enough, set in the tools themselves so the agent cannot widen its own access.
What it may do
What it prepares alone, such as a note, a filled sheet or a draft, and which steps always wait for a person's approval.
When it stops
Missing data, an unsure result or anything outside the job goes to a named person. At first every run is checked; your team decides when to check less.
WHAT WE DO NOT BUILD
Where agents stop, and where to look instead.
No agent that sends, publishes, pays or edits a live account on its own, and no hours-saved figure until the job has been measured.
Every run follows the same steps
Business process automation, with no model involved and someone named for the odd cases.
The work needs a draft to review
AI systems inside daily tasks. That page also explains how a model change is tested.
You are not sure an agent fits
One consulting session to decide between an agent, a simpler automation or your own team.
You want to read more first
Guides on fitting AI into daily work and reviewing what it produces.
FAQ
Questions about AI agents
Is an agent the same as a chatbot?
No. A chatbot answers people in a conversation. Our agents work in the background on one job and report to your team; conversations with your customers stay with your people.
Can the agent learn from corrections?
Yes, under control. Corrections update its instructions and examples, and each update is tested on saved runs before it goes live. The agent never rewrites its own instructions.
What happens if the agent gets stuck?
It stops instead of guessing. A stopped agent costs a delay; a guess can send a mistake to a customer. Its log shows where it stopped, so the person picks up from there.
Can an agent work with our Amazon account?
Yes, through an application you authorise in Seller Central and can switch off at any time, never through a person's login. Changes to a live listing, price or campaign wait for approval.
NEXT STEP
Which recurring job would you hand over first?
Say what the job involves, how often it comes round and who does it now. Our reply says how to start: the diagnosis if you have a bag of hours, or an hour of consulting.
Tell us about the job