The Agent Authority Model
A deterministic way to set how much authority each agent handover may exercise on its own.
Assess one handoverThe handover, not the agent.
Govern the handover, not the agent. Knowledge, Decision and Action describe individual moments when an agent passes something to a person, another agent or a system. One agent can make several handovers, and each can carry a different exposure and authority ceiling.
These are not permanent agent classes. A single agent may answer a question, fix a case outcome, and change a system record—three handovers that can require three different authorities.
Place handovers into the operating model- 01
Knowledge
Someone simply knows more.
A telecom care agent answers a tariff question.
- 02
Decision
It fixes an outcome for a specific case or person that something downstream acts on.
The same agent scores churn risk.
- 03
Action
A record, system, or physical thing is different after it runs.
The same agent applies a retention credit.
Six answers become a control specification.
The process owner answers the questions. The application calculates the ceiling. Narrative can be edited and reviewed in the CMS; the scoring rules remain deterministic application logic.
- 01Classify the exact handover: Knowledge, Decision, or Action.
- 02Score how the operation can reverse a wrong result, R1–R4.
- 03Score who is affected or would find out, H1–H5.
- 04Compare requested authority with the calculated oversight ceiling.
- 05Name one accountable operating role—not a team.
- 06State promotion evidence and the condition that demotes authority automatically.
Exposure sets the ceiling
Classify what is handed over, then score reversibility from R1 to R4 and reach from H1 to H5. The more severe answer determines E1 to E5 and therefore the permitted oversight.
- E1: out of the loop for internal, reversible handovers.
- E2: on the loop, with a stated intervention window.
- E3: in the loop before the handover acts.
- E4: in the loop with an independent second control.
- E5: in the loop with external safety sign-off.
Evidence earns the climb
Name one accountable operating role, the evidence required for promotion and the conditions that automatically demote the handover. Authority requested above the ceiling must sit in an approved artefact such as a template, whitelist, rule set or blocking gate rather than in the model itself.
Airline disruption re-accommodation.
The agent rebooks one passenger and issues a boarding pass. Reversal needs the passenger or receiving carrier, so the intervention window must be shorter than the time the released seat remains available. The accountable role is the Duty Manager, Operations Control Centre.
- Handover
- ACTION
- Reversibility
- R3
- Reach
- H2
- Exposure
- E2
- Ceiling
- On the loop, with a stated intervention window
- Requested authority
- On the loop, with a stated intervention window
- Owner
- Duty Manager, Operations Control Centre
- Promotion evidence
- 500 consecutive rebookings with zero disputed reversals and no complaint uplift against the manual control.
- Automatic demotion
- Any involuntary downgrade or caused missed connection.
Binding condition: Before the released-seat inventory expires. No carrying artefact is required because the requested authority sits at the calculated ceiling.
Technology does not determine authority.
The same underlying capability can sit at different authorities because operational exposure differs.
| Sector | Handover | Type | Exposure | Authority / control |
|---|---|---|---|---|
| Telecoms | Telecom tariff answer | Knowledge | E1 | Out of the loop |
| Financial services | Credit decline and adverse-action reason | Decision | E3 | In the loop |
Six questions · one handover
Set an authority ceiling your operation can defend.
Nothing is sent to Cognirise or saved in your browser. Answers remain in this page only and are cleared when you leave or reload. Assess one handover, download the local control brief, then restart for every other handover the same agent makes.
1. What is handed over?
Ask in order. First yes wins. This classifies the handover, not the agent.
Guardrails are not an authority model
Most teams that have built guardrails believe they have governance. They have not, and the gap is expensive.
A guardrail is a mechanism: an output filter, a rate limit, a system prompt, an approval step. An authority model is the thing that decides which mechanisms are required, where each one sits, and who is answerable when one fails. The first is a component. The second is a delegation.
Every bank already runs this distinction without thinking about it. Its delegation of authority says a relationship manager may approve up to a limit, a committee above it, and anything touching a sanctioned party goes to compliance regardless of size. The four-eyes check and the screening filter are the guardrails. Nobody would say "we have four-eyes checks, therefore we have a delegation of authority." That is precisely what is being said when a team points at a content filter and calls it AI governance.
Three differences that matter
| Guardrails | The Agent Authority Model | |
|---|---|---|
| What it attaches to | GuardrailsThe agent or the model — one filter, one prompt, one limit, applied to everything it does | The Agent Authority ModelThe handover — the moment an output leaves the agent and becomes consequential for someone else |
| Where it comes from | GuardrailsChosen, usually from a vendor's feature list or from whatever has already gone wrong | The Agent Authority ModelDerived from the handover's profile: what is handed over, who is present, and what is at stake |
| What it answers | GuardrailsWhat stops it doing something bad? | The Agent Authority ModelWho authorised it to do this — and what must it prove before it is allowed to do more? |
Why the unit is the whole argument
A single agent does several things of very different consequence. A front-desk agent answers questions, books appointments, cancels them and issues refunds. Guardrail thinking gives all four the same protection, because the protection was attached to the agent. The refund is then defended exactly as well as the opening-hours question — which is to say, the riskiest thing the agent does inherits the posture appropriate to the safest.
Govern the handover instead and each of the four is rated on its own: what kind of thing is being handed over, who is standing there when it happens, how hard it is to undo, and who is exposed if it is wrong. Same agent, same model, same guardrail technology — but the refund is now governed as a refund.
In engineering terms: authority attaches to the tool, not the agent.
The common pattern
Guardrails govern the agent
One filter. One prompt. One limit. Applied to everything it does.
Guardrail perimeter
- Answer a clinic questionsame authority
- Book an appointmentsame authority
- Cancel an appointmentsame authority
- Issue a refundsame authorityWarning: the riskiest act inherits the posture of the safest.
The riskiest act inherits the posture of the safest.
Governance attaches to the wrong object, so the refund is protected exactly as well as the opening-hours question.
The Agent Authority Model
Authority governs the handover
Each moment the output leaves the agent is governed on its own.
Separate handover profiles
- Answer a clinic questionKnowledge · R1 / H1out of the loop
- Book an appointmentAction · R2 / H2on the loop
- Cancel an appointmentAction · R3 / H2 · defined intervention windowon the loop
- Issue a refundAction · R4 / H3in the loop
Authority attaches to the tool, not the agent.
Same agent, same model, same guardrail technology — but the refund is now governed as a refund.
How the two interact
Two rules bind them.
Exposure sets the ceiling. How hard an output is to undo, and how far its effects reach, determine the maximum authority a handover may hold. Capability does not enter into it. A more accurate model does not earn more authority; a smaller blast radius does.
Evidence earns the climb. Every handover launches one level below its target authority and is promoted only on measured performance, with demotion automatic on incident. Guardrails have no concept of promotion — they are static by nature. This is the question no published framework asks, and it is the one that turns a classification into an operating model.
Guardrails then enter in two distinct ways, and keeping them apart is the whole of the discipline.
As required controls. Once a handover has a profile, its control set follows from that profile rather than from preference — universal controls, plus those set by type, by authority level and by exposure band. Each must be provable by a test, a query or an artefact, never by an assurance. "The team ensures the agent does not give medical advice" is not a control. "An output filter independent of the model blocks each prohibited class, and a test suite attempts every one of them" is. A guardrail you cannot test is not a guardrail; it is an intention.
As compensating controls that raise the ceiling. This is the mechanism that makes the model workable rather than merely restrictive. A handover may hold authority above its exposure ceiling where the content of the handover is constrained by construction rather than by trust in the agent. An agent sending messages to patients with no human present sits above the ceiling for an irreversible, customer-affecting handover. It can still be correct — if it writes no free text, renders a clinician-approved template through a whitelist of variables, and passes a blocking gate on every send. The authority there is carried by the approved template. The agent is a dispatcher.
A chart with authority rising vertically from in the loop to out of the loop, and exposure rising left to right. A descending staircase marks the ceiling exposure sets on authority. Required controls sit in the permitted region below it, an arrow shows a handover promoted upward on measured evidence, and a marker above the ceiling shows a compensating control where authority is carried by an approved template rather than by the agent.
Exposure sets the ceiling. Evidence earns the climb. Guardrails do both jobs.
The five canonical exposure bands run left to right. Authority is shown in three human-oversight rows; the pink staircase is the exposure ceiling, and everything beneath it is permitted.
Read the plot from the profile, not from the agent's confidence. A handover can climb on evidence, while an approved compensating artefact is the only way to carry authority above its exposure ceiling.
Scroll horizontally to explore the chart. Keyboard: focus the chart, then use Left and Right arrows.
Illustrative exposure buckets — annotations, not axes
Five canonical bands; illustrative buckets below are examples, not alternative axes.
A Permitted
Everything under the staircase is permitted. The controls required here are read off the profile, not chosen. Required controls sit in here — the guardrails the profile makes mandatory — universal, by type, by authority, and by exposure.
B Promotion
Evidence earns the climb. Guardrails as required controls: meet them, hold the evidence, and the handover climbs one level — launching one level below target. A better model does not raise the ceiling; promotion never moves it.
C Compensation
A compensating control raises the ceiling. Guardrails that carry the authority themselves. Only these move the ceiling: only compensation, through an approved artefact, may carry authority above it — never unrestricted agent discretion. Example: recall message to a patient — no free text, approved template, blocking gate on every send.
Reversibility R1–R4 and reach H1–H5 as defined in the Agent Authority Model. Handovers shown are illustrative.
The design rule this produces
Where an agent operates above its exposure ceiling, the design must name the artefact that carries the authority instead — an approved template, a whitelisted parameter range, a deterministic rule set, or a gate with the power to block.
That is the cleanest test of the relationship between the two. A guardrail that merely constrains the agent does not move the ceiling; it has made a risky thing somewhat less risky. A guardrail that carries the authority itself does move it, because the consequential content no longer originates with the model at all. Most teams cannot say which kind theirs is. That, usually, is the finding.
The failure runs in both directions. An authority model with unenforceable controls is a register of good intentions. Controls without an authority model are a pile of features nobody can justify to a regulator. Each is load-bearing for the other.
Authority is earned, not configured.
Published practice, not invented vocabulary.
Human in, on, and out of the loop is established oversight language. The model combines that vocabulary with published practice on human oversight, reversibility, reach, and levels of automation.
- Cognirise Agent Authority Model, approved source presentation
- EU AI Act, Article 14 — Human oversight
- NIST AI Risk Management Framework
Verified 2026-09-06. Next editorial review 2027-03-06.
Bring one handover under pressure.
Use the assessment privately now. When you need to translate the brief into architecture, controls, and accountable delivery, bring us the process.