How it works
How managed AI agents work, with your team in charge.
AI agents do the repeatable work. Ergon operators watch them every working day and fix what they get wrong. A named person on your team approves anything that matters.
Who does what
Agents, operators, and your team.
Each party has work it owns, and a limit it never crosses.
- 01
Agents
Software that carries the repeatable tasks of a role.
- Read what arrives and sort it
- Enter records and draft replies
- Reconcile, schedule, and follow up
Never act outside their written permission level.
- 02
Ergon operators
People on our team who run your agents every working day.
- Watch the queues and resolve exceptions
- Correct outputs and log each correction
- Report volume, exceptions, and cost
Never make a decision that commits your business.
- 03
Your people
The named approvers who own the business decisions.
- Set the rules agents and operators follow
- Approve payments, commitments, and exceptions
- Approve every change before release
Never lose the final word on anything consequential.
A working day
Our operators run the agents. Your team signs off.
The same five steps run every day. The approval step never moves to an agent.
- ArriveWork arrives in the systems you already use.
- WorkAgents carry the agreed tasks inside written permissions.
- ReviewOperators resolve exceptions and correct outputs within policy.
- ApproveYour approver decides payments, commitments, and anything outside policy.
- ReportYou see volume, exceptions, corrections, and cost.
Permission levels
Five levels, written down for every agent.
Each task on the Role Card carries one of these levels. Raising a level is a written decision by your approver.
| Level | What the agent may do | Who approves | Example |
|---|---|---|---|
| Read | The agent reads the sources it was granted, and nothing else. | The access is approved once, when the system is scoped. | Reading open invoices in the accounting system. |
| Draft | The agent prepares work, and a person decides whether it is used or sent. | An Ergon operator or someone on your team reviews each draft before it leaves the system. | A reply to a customer, waiting in the review queue. |
| Pre-authorized routine write | The agent makes a bounded, reversible change that a written policy already allows. Every change is logged. | A named owner on your team approves the policy in advance. Operators watch the changes, and anything outside the policy waits for a person. | Correcting a duplicate customer address under the approved conflict rule. |
| Consequential action | Payments, commitments to customers, changes to a record of truth outside an approved policy, and anything that cannot be undone. | A named person approves each one, every time. If nobody approves, nothing happens. | Paying a supplier, or changing a price a customer was quoted. |
| Release | A new model, agent, prompt, or rule version goes live. | The client’s named approver approves each release after the agreed tests pass. Contractual acceptance of the work is a separate step. | Moving an improved extraction model from testing into daily use. |
Trust in steps
An agent earns each permission with evidence.
Every task starts in shadow. It moves up only when the run records support it and your approver agrees in writing.
- ShadowThe agent works beside your team on real inputs. Its drafts are compared with what your people did. Nothing it prepares is sent or saved.
- Approve every actionThe agent’s work goes live, but an operator or your approver confirms each action before it happens.
- Routine within a policyWhen run records support it, your approver signs a written policy. Routine, reversible actions inside it proceed, and operators watch each one.
- Consequential stays humanPayments, commitments, and anything outside the policy wait for your named approver, at every stage.
Always yours
Decisions that stay with your team.
These decisions stay with a named person on your team, every time. The agreement says so in writing.
- Payments and refunds outside an approved policy
- Prices, discounts, and commitments a customer relies on
- Hiring, clinical, legal, and credit decisions
- Complaints, disputes, and anything sensitive
- Changes to records outside an approved policy
- Every new release of an agent, prompt, model, or rule
Getting better
Every correction becomes a test.
Improvements come from the work itself, and none reaches daily use without your approval.
- CorrectionAn operator or someone on your team corrects an output. The correction is logged with its input and reason.
- Test caseThe corrected case joins the test set for that task, so the next change is checked against it.
- ChangeAn Ergon engineer adjusts the agent, its instructions, or its model, and runs the agreed tests.
- ReleaseYour approver sees the test results and decides whether the change goes live.
Questions
Are the operators your employees or ours?
Ours. Ergon operators are part of the Ergon team and work from your written rules. Your people keep the decisions.
What happens outside operator hours?
Coverage hours are set in the agreement. Outside them, agents continue routine work inside policy and hold anything that needs a person.
Can we see everything the agents do?
Yes. Every run is recorded with its input, output, corrections, and cost. Reports summarize them at the agreed interval, and the records are available on request.
Who decides when an agent gets more permission?
Your named approver, based on the run records. Moving from approval on every action to a pre-authorized policy is always a written decision.
Ready when you are