No. In the work we propose, a person approves anything that leaves the business or changes a record you rely on. An agent that cannot be stopped is out of scope.
Agentic AI and automation
Agents that prepare the work. People approve it.
We design AI agents, fixed automations, and the integrations between systems you already run. The point is a workflow with a start, a draft, and a person who can say no.
- Support triage
- Lead qualification
- Document processing
- Reporting
- Internal knowledge retrieval
01 — Definition
What an agent does here.
An agent is given a goal, a short list of tools, and a rule for when to stop. It can classify, retrieve, and draft. It is not a general employee, and it is not allowed to invent a tool you did not name.
What enters is a message, a file, or a record. What it prepares is a classification, a retrieved passage, or a draft. Where a person approves is before the send, the post, or the status change. If it cannot retrieve, it stops.
We do not describe this as autonomous, and we do not promise that the draft is right.
02 — Choose the tool
Chatbot, fixed automation, or agent.
| Chatbot | Fixed automation | Agent | |
|---|---|---|---|
| Best at | Answers you already wrote, or a passage from a known source. | A path you can draw: status in, action out. | Unstructured input where the next step depends on reading it. |
| Stops when | The script or the retrieval ends. | The path ends or a field is missing. | The stop rule you wrote is reached, or a tool fails. |
| Should not | Change a record or make an offer. | Invent a branch that is not on the path. | Send, pay, delete, or hire without review. |
03 — Uses
Five workflows we will actually scope.
- 01
Support triage
Classify a ticket against answers you already stand behind, draft a reply, and stop. A person sends it. Unmatched tickets stay in the human queue.
- 02
Lead qualification
Read an enquiry, pull the fields you sell against, and prepare a summary for the owner. The prospect does not receive a message you have not approved.
- 03
Document processing
Read one repeating document type into fields. A person confirms the rows that fail the check before anything posts to the system of record.
- 04
Reporting
Assemble a report from a named source, on a schedule, in a shape someone already knows how to audit.
- 05
Internal knowledge retrieval
Answer a staff question from a corpus you choose, and show the source. It is not a search of the public web, and it is not an authority over policy.
04 — A worked example
One ticket, then a stop.
A customer writes about an order. The message enters. The system classifies the request, retrieves that order from the system of record, and drafts a reply from the order and an explanation you have approved. A person on your team sends the draft or discards it.
- Classify against a list written before the pilot.
- Retrieve only from the systems named in the scope.
- Draft the reply. Do not send it.
- If the order is missing, queue the ticket and stop.
This is an illustration of the pattern. It is not a claim that a percentage of your inbox can be handled this way.
05 — Integrations
The systems you already run.
A useful agent is mostly an integration project. Typical connections are a helpdesk, a CRM, a file store, and an internal API that already represents the record. We use the access those products offer: an official API, a documented export, or a permissioned account.
We do not claim universal compatibility. If a tool has no supported connection, the honest options are a different boundary, a manual step, or dropping that tool from the pilot. Screen-scraping a product against its terms is not a method we propose.
06 — Guardrails
Permissions, review, logs, exceptions.
- 01
Permissions
The agent gets the least access that completes the draft. Read is the default. Write is a separate decision, usually reserved for a person or a later, narrower automation.
- 02
Review
Outward messages and record changes wait. The reviewer can edit, send, or reject. Rejection is a normal outcome, not an error in the log.
- 03
Logs
You should be able to see what entered, what was retrieved, and what was proposed. A log is how you audit a pilot. It is also personal data, so retention belongs in the project terms.
- 04
Exceptions
Low confidence, a missing record, or an out-of-slice request goes to a person. The agent does not “try anyway.”
07 — Pilot
Evaluate a slice, not a slogan.
Pick a slice you can define: a document type, a queue, a kind of enquiry. Record a baseline for two ordinary weeks. Run the pilot with the same definition of done. Compare time, edit rate, rejection rate, and anything you had to undo.
A pilot that excluded the hard cases is not evidence you can scale. Expand only the slice that held up, and keep the off switch.
08 — When not to
Use the simpler automation.
Use a form if a form would collect the facts. Use a rule if the branch is a status you already store. Skip the agent if nobody has time to review the queue, if you have no system of record, or if a wrong outward action is too expensive for a draft to sit in front of a customer.
An agent is the wrong first project when the policy itself is undecided. Software will not supply the missing rule.
09 — Engagement
From a workflow to a supported slice.
- 01
Discovery
The workflow, the volume, the system of record, and the cost of a wrong step.
- 02
Pilot
One slice, one queue, one reviewer. The agent prepares. It does not own the outcome.
- 03
Evaluation
The same definitions as the baseline: time, exceptions, edits, and anything that had to be undone.
- 04
Deploy
Only the slice that passed, with permissions, logs, and a way to turn it off.
- 05
Support
Agreed changes, and a clear view of model or API usage if the workflow calls one.
10 — Scope and running cost
What the estimate has to include.
- How wide the slice is, and what is excluded
- Which systems are in bounds, and the access they allow
- Review load on your side during the pilot
- Model or API usage at the volume you expect, and at double that volume
- Logging, retention, and who may read the log
- The support window, and what counts as a new project
No package price is published. Ongoing usage is not “free once it is built” when a model is called per item.
11 — Products and experiments
Patterns the studio has built.
These are not subscriptions. If you want the same pattern on your systems, start a project.
01
Bellwether PA
Studio build. Not a subscription. We can put the same pattern on your systems as a project.
Read 02AIMEE
Studio prototype for course teams. A teacher still approves the answer.
Read 03Joyquill
Studio tool for turning a brief into a PDF a person edits.
Read 04Instagram AI Avatar
Studio experiment for image drafts. A person decides what is published.
Read 05AI Trading Bot
Not a service. Not financial advice. This page is kept out of search.
Read12 — Questions
Before a pilot.
Accurate enough is a measurement you take on your data during a pilot. We do not guarantee a percentage. If the source is wrong, the draft will be wrong with confidence.
Only through the connections the project names. A model API is a third party. If a workflow must stay inside a boundary you set, that constraint is part of the design, not a footnote.
Hosting, any model or API usage tied to volume, and the support period you agree. Usage cost should be visible before the pilot is called a success.
Not as a login or a subscription on this site. They are patterns the studio has built. If you want one on your own systems, that is a project, scoped like any other.
Start
Describe the workflow and where a person must approve.
Tell us what enters the system, what a good result looks like, and which tools already hold the record. We will say if a simpler automation fits.