The HAR method
Human and AI agent roles, decided one decision at a time.
HAR (Human Agent Relationship) is a method for dividing work between people and AI agents. It looks at work as decisions and settles each one with three questions, a supervision mode and a named accountable person.
Decision paths
The unit of work is the decision.
A business outcome comes from a series of decisions: what is let in, what is prioritised, what is approved, what is escalated. We follow that series from the event that starts it to the consequence that follows. The method calls it a decision path.
A decision path
- Trigger
- Inputs
- Decision point
- Outcome
- Execution
- Consequence
Decision point: a point on the path where at least two outcomes are possible and one is chosen.
Three questions
Each decision is scored on three axes.
The three answers stay separate, so a high consequence cannot hide behind a low recurrence.
-
How often is the decision made?
Counted in decisions, from fewer than five a year to continuous. A process that runs once a week can hold hundreds of decisions.
In the method: Recurrence
-
Can it be done reliably the same way?
How far the decision can be written as rules and data in your organisation. One test is to let two experienced people decide the same five recent cases and compare.
In the method: Standardizability
-
What does an error cost?
The stakes, whether the decision can be reversed, and how soon an error is noticed. The supervision mode follows from this answer.
In the method: Consequence
From consequence to supervision
What an error costs sets how closely a person supervises.
Consequence combines the cost of an error, whether it can be reversed, and how soon it is noticed. Each level has a default supervision mode. A different choice is recorded with its reason.
| Level | What an error is like | Default supervision mode |
|---|---|---|
| C1 | Cheap, easy to reverse and visible at once | Audit |
| C2 | Cheap and easy to undo, and would go unnoticed without someone watching | Monitor |
| C3 | Costs something, and shows within the same cycle | Approve |
| C4 | Material and noticed late, or severe and visible at once | Approve |
| C5 | Severe or irreversible, and invisible when it is made | Command Collaborate |
Five supervision modes
Command, Collaborate, Approve, Monitor, Audit.
Ordered from most to least human involvement per instance. The names are fixed.
-
Command
A person directs each instance, and the agent carries out the intent.
High attention per instance
An error would be severe or irreversible, and nothing downstream would catch it in time.
-
Collaborate
A person works each instance together with the agent.
High attention per instance
The stakes are as high as for Command, and the person and the agent each add something to every case.
-
Approve
A person signs off each output before it takes effect.
Moderate attention per instance
An error costs something or is hard to undo, and a sign-off before it takes effect can stop it.
-
Monitor
A person watches the flow and the exceptions it raises.
Light attention
Something has to be watching, and nothing has to be signed off one by one.
-
Audit
A person samples completed work after the fact.
Light attention
Errors are cheap, easy to reverse and visible at once.
Accountability
Every path has a named, accountable person.
This person is the accountability anchor: the named person answerable for the outcomes of a decision path, with the standing and information to carry it.
If no accountability anchor can be named, the decision is not ready for an agent.
-
Capacity is part of the design
A supervision mode that nobody has time to perform is a rubber stamp. The design checks the supervision load against the hours people have.
-
People decide
The rule proposes, and a named person decides. Proposals are sized to the time of the people who decide them.
-
Freed capacity
Freed capacity counts when it is placed. Released hours become value when they go to named work.
In operation
Autonomy grows on evidence, and supervision is watched for drift.
The outer loop decides the operating model, and the inner loop builds and runs the agents. Evidence from use returns to Reassess, which moves autonomy up or down.
-
Outer loop
The operating model’s loop of Discovery, Design, Scale and Reassess, which decides who does which work under which control. The HAR method runs it.
-
Inner loop
Building and running the agents the design calls for, and measuring them in use. We build the first agents with you to prove a path; production builds and Run go to your team or an engineering partner you choose.
-
Progressive autonomy
An agent’s independence increasing in steps as evidence accumulates. The step back down stays available. The thresholds are agreed at design time, before anyone is under pressure to loosen them: an error rate, an error class, how soon errors surface, and how many decisions were observed.
-
Mode drift
Supervision loosening without anyone deciding it. The operating review is there to catch it. Three signs: approval rates close to 100 %, escalations far below the estimate, and a new accountable person who has not been briefed on what they answer for.
Limits
What the method does not claim.
Saying what is not known is part of the method.
- It does not predict how many agents you should have. Headcount for people and agents is the result of the calculation, and it changes when volume, consequence or capacity change.
- It does not evaluate models, vendors or platforms. The design is independent of any platform up to the engineering specification.
- It does not replace a compliance function. It surfaces exposures and stops the work at a gate until they are handled.
- It does not measure culture. It measures supervision load, error rates and how soon errors are noticed.
Glossary
HAR terms in plain words.
- Human Agent Relationship (HAR)
- A method for dividing work between people and AI agents, one decision at a time.
- Decision path
- The series of decisions that produces a business outcome, from the trigger that starts it to the consequence that follows. The unit of analysis.
- Decision point
- A point on the path where at least two outcomes are possible and one is chosen.
- Recurrence
- How often the decision is made. Frequency only, scored R1 (fewer than five a year) to R5 (continuous).
- Standardizability
- How far the decision can be expressed as rules and data, scored S1 (tacit judgement) to S5 (deterministic).
- Consequence
- How bad it is when the decision is wrong, read from the cost of the error, its reversibility and how soon it is noticed. Scored C1 to C5.
- Supervision mode
- How closely a person supervises the work, from Command through Collaborate, Approve and Monitor to Audit. Derived from the consequence.
- Accountability anchor
- The named person answerable for the outcomes of a decision path, with the standing and information to carry it.
- Supervision load
- The attention a person’s paths and their modes require. A design constraint, checked against the hours people have.
- Progressive autonomy
- An agent’s independence increasing in steps as evidence accumulates. The step back down stays available.
- Mode drift
- Supervision loosening without anyone deciding it. The operating review is there to catch it.
- Resourcing as outcome
- Headcount for people and agents is the result of the design calculation, set after the work is understood.
- Cost per accepted outcome
- Total cost divided by the outcomes that passed supervision. The measure a path is judged by, before and after.
- Hybrid system
- An organisation whose work is done by both people and agents.
- Gate
- An approval point. The business signs off what a gate covers, and the next phase starts after it.
- Outer loop
- The operating model’s loop of Discovery, Design, Scale and Reassess, which decides who does which work under which control. The HAR method runs it.
- Inner loop
- Building and running the agents the design calls for, and measuring them in use.
Questions
Questions about the method.
Does the method replace our compliance function?
It works alongside it. The method surfaces exposures and stops the work at a gate until they are handled. Every decision path has a documented supervision mode and a named accountable person, and gate decisions are kept on record with the reasons.
Which AI platform do you recommend?
We leave that choice with you. The design is independent of any platform up to the engineering specification. Your team can build it on the platform and with the vendor you choose.
Start with a 30‑minute maturity check.
We place your company on the five-level adoption scale, match your situation to a typical buying trigger and recommend one next step. No charge, and no obligation to continue.