Persona Agents. Agents that know how your business works.
Persona Agents are AI agents that each hold one specific job in your business — an estimator, an admin, a quoter, a scheduler — at a named autonomy tier. An agent with no context is an expensive way to generate confident nonsense. Persona Agents run on top of Org Brain, the context layer, which is the only reason one can hold a real job in your business instead of a demo in a slide deck.
- L1 AssistDrafts and answers. A person does anything that changes a record.
- L2 ProposePrepares the complete action and parks it for approval. The default.
- L3 Act in boundsCompletes it inside a written boundary. Outside it, drops back to L2.
- L4 Always approveMoney out, safety, people. A permanent floor, not a starting point.
Trust is earned per task, from that task's own shadow-mode record. An agent reliable at quoting has earned nothing about scheduling.
Context comes first. There is no version of this that skips it.
An agent that cannot see your precedents, your entity graph, your current capacity or your policy will still answer confidently. It just will not be right, and it will not know that it is wrong. That is the specific failure mode of an ungrounded agent: not silence, but fluent error.
Org Brain is what gives an agent a real job to be good at: read the argument for why context is the bottleneck before you look at agents at all.
Five roles, and what each one has to know to be trusted with it.
What an agent does is the easy half to describe. The interesting half is what it has to know before its output is worth reading.
Show the 5 roles and what each has to know
- 01EstimatorDrafts a job estimate from similar past work, current materials and labour rates. It must know the precedent index — what a comparable job actually cost and how it was scoped — and the entity graph linking that job's assets and site conditions to this one. An estimator with no precedent is guessing with confidence.
- 02AdminDrafts routine correspondence, status updates, reminders and data entry across systems that should already agree with each other. It must know the entity graph, so it updates the one record for a customer or job rather than three half-matching ones, and the policy on who is allowed to send what.
- 03Quoter / bid coordinatorDrafts a quote or a bid pack from your pricing history and current cost inputs. It must know the precedent index of what you have quoted, won and lost before and why, and the policy that sets margin floors and who has to approve an exception to them.
- 04Service schedulerProposes a service or maintenance run against current technician and equipment availability. It must know current capacity in real time, not a plan from last week, and the entity graph tying an asset's service history to the site and the customer contract that covers it.
- 05Compliance checkerFlags a job, quote or document against the standards and regulatory obligations that apply to it. It must know policy — your standards and the regulations that bind you — and precedent, so it can point to how a similar case was actually handled, not just what the rule book says in the abstract.
The default is L2 Propose.
Most of the work on this page runs at L2: the agent prepares a complete draft, a named person reviews it against the source it cites, and only that person sends, posts or commits it. Nothing crosses from draft to action without that step.
A narrow task can move to L3 on its own shadow-mode record. Money out, safety and people stay at L4 for good. All four tiers are defined once, in section 04.
Four levels. Named, not implied.
"The agent proposes, a person approves" is not a governance model on its own — it is a slogan that hides where the actual boundary sits. Every persona agent on this page runs at one of four explicit levels, and moving between them is a decision on the record, not a drift nobody signed off on.
| Level | What happens |
|---|---|
| L1 Assist | Always on. The agent drafts, suggests and answers, and a person does everything that changes a record. Nothing needs approval, because nothing happens without a person acting on it first. |
| L2 Propose | The agent prepares a complete action — a quote, a reply, a schedule change — and parks it for approval. A person approves it or edits it. This is the default tier for most of the work on this page. |
| L3 Act within bounds | The agent completes the action itself, but only inside a boundary written down in advance — a value ceiling, a named set of customers, a specific record type. Anything outside that boundary drops straight back to L2, and every boundary is agreed in writing and logged. |
| L4 Approval required, always | Anything touching money out, safety or people needs a named person's approval on that specific action. These tasks never move up a tier, no matter how well the agent has performed. This is a permanent floor, not a starting point. |
A task earns its way from L2 to L3 on the evidence of its own shadow-mode record, one task at a time. An estimator agent proven on repeat-customer quotes has earned nothing on compliance flags. L4 is not on that ladder at all.
Owner and blocker are not the same person.
Most stalled work is not unassigned. It has an owner, a deadline and a place in someone's queue — it is just waiting on a different person than the one who owns it, and that person usually has no idea they are the reason it has not moved. A persona agent that only tracks ownership will chase the wrong name. One that tracks who a task is currently pending with escalates against the actual blocker instead.
A repeat-customer quote has been sitting in the customer's inbox for two days, but the actual hold-up was a margin exception a manager had to sign off before it should have gone out at all. The task's owner is sales. The task was waiting on the manager, who had no idea he was the reason it stalled.
A maintenance job is assigned to a technician for Thursday, but it cannot actually be locked in until a foreman confirms a piece of specialised equipment is back from another site. The task's owner is the scheduler. The task is waiting on the foreman, and chasing the technician fixes nothing.
A staged build, not a switch you flip.
- 01Scope one jobPick a single, specific job an agent will do — not "help with quoting", but "draft a quote for a repeat customer from a matched precedent job". Narrow enough to be provably right or wrong.
- 02Encode the context it needsConnect and structure exactly the precedent, entity and policy data that job depends on. If Org Brain does not already hold it, this is where it gets built — not guessed at inside a prompt.
- 03Run it shadow-modeThe agent drafts its proposal on real, live work, alongside the person who would normally do the job. Nobody acts on its output yet. We compare what it proposed against what the person actually decided.
- 04Run it at L2 ProposeOnce the shadow comparison holds up, the agent's draft goes to the person doing the job as a real proposal to approve, edit or reject, on the record every time.
- 05Widen the jobOnly after a specific proposal type has earned a track record of approvals does its scope widen — more job types, or L3 within written bounds on that one task, never a blanket expansion to "handle quoting" in one step. Anything at L4 stays there.
What agents are bad at.
Agents are poor at judgement calls with no real precedent behind them — the first job of a genuinely new kind, a dispute where the right answer depends on reading a person rather than a record, or a decision where being wrong once is more expensive than the entire programme is worth automating.
We would talk you out of automating a workflow that changes every few months faster than its context can be encoded, or a job whose real value is the relationship it maintains rather than the output it produces. Automating the paperwork around a relationship is fair game. Automating the relationship is not.
If a job cannot be scoped narrowly enough to shadow-test, it is not ready for an agent yet — that is a sequencing problem to fix, not a reason to widen the agent's licence.
Put an agent on one real job, once the context exists for it.
If Org Brain is already live, we can scope a first persona agent in weeks. If it is not, we will tell you that honestly and start there instead.