A new collaborator has been three weeks in the position. His manager calls him because he authorized something he should not have: he promised a client a delivery for Friday. "He asked me when, and Friday seemed reasonable to me. That is five days", he explains. "That product takes three weeks. Always", the manager answers. "I did not know. Where is that written?". And then comes the answer that portrays half of the organization: "Everybody knows it".
Nobody acts in bad faith, and that is exactly the point. The manager is right on substance: the date was impossible. The collaborator is right on form: nobody told him, and he acted with the judgment he was able to build. The conversation ends with the manager concluding that the new guy "has no judgment": the most comfortable conclusion and the least useful one.
Now replace the collaborator with an artificial resource and the exchange becomes impossible. The system is not going to ask where that is written. It is simply going to keep promising Friday.
The trial that is impossible to win
Many AI implementations reproduce that first day. The agent receives access, a general instruction and a disproportionate expectation. It does not know the historical exceptions nor the hierarchy of priorities, and it does not know which decision requires approval. So the organization submits it to a trial it cannot win: if it asks too much, we say it is not autonomous; if it decides and fails, we say it is not reliable. In reality we demand performance from it inside a position we never designed.
The pattern is seen at scale. A report from the NANDA project, of the MIT Media Lab, studied enterprise initiatives of generative AI and found that the enormous majority did not produce measurable return. Its most cited figure, the 95% without return, was much discussed, and the book does not rest on it. What resists the debate is the diagnosis: the systems did not retain context, did not remember what they had learned and did not adapt to the real workflow. The report called that the learning gap. It is the exact description of a collaborator nobody onboarded.
Would you hire a person by telling them only "help sales and do it well"? You would explain to them what result they must produce, which promises require approval and whom to go to when facing an exception. A prompt can orient an execution. It does not replace the design of a position.
A prompt gives instructions. A position gives responsibility.
An operational contract, not a legal piece
The instrument that converts a technological intention into a manageable unit of work is the AI Role Contract, the role contract: the central document of the whole method. It does not pretend to humanize software nor does it have juridical ambition: it is the functional combination of a good job description and an explicit agreement of operation, to convert technological capability into manageable responsibility.
Let us go through its pieces with the thread the book uses: the responsibility of collection reminders that the HWFA assigned to an artificial occupant.
The pieces of the contract
Mission and results
What the role exists for and what value it must produce. Not a list of movements: a purpose. The book exemplifies it like this: "To convert early overdue invoicing into payments, protecting the relationship with the client, for accounts less than 45 days late." A well-written mission already excludes things: this role does not negotiate agreements, does not touch accounts more than 45 days overdue, does not decide about clients in dispute. Every result needs indicators, and a good KPI also watches what must not happen: "zero contacts to accounts flagged as in dispute" is as measurable as a response rate.
Authority and exclusions
What it can consult, create, modify, communicate, approve, commit or spend. In practice it is a table, and its value is in the discomfort of writing it. It can: consult balances, send reminders from approved templates, record payment promises. It cannot: offer discounts or write-offs, modify due dates, commit payment plans, contact accounts in dispute. The exclusions are as important as the permissions. Responsibility and authority travel together: you cannot demand accountability where you never defined authority.
Escalation
Every exception needs a destination: a person, a team or a return rule. The client answers with a complaint or a dispute: it transfers to the analyst with the complete file, within two business hours. The client mentions serious economic difficulty: it stops the sequence and flags it for human handling. A role without escalation destinations does not have governed exceptions. It has buried exceptions.
Suspension and kill switch
If the error exceeds a certain band, an incident occurs or a critical dependency disappears, the manager must know when to reduce autonomy or to stop the role. The kill switch without anyone responsible is barely a technical function; the contract defines who can use it and under what condition. And a switch that has never been tested is decoration.
Governance calendar and versioning
Everything that has a clock in the life of the role lives in a single section with dates and a single owner of the schedule: the review of the contract, the re-justification of the position, the kill switch drills, the revalidations after changes. Without a consolidated calendar, every obligation has a date and none has anyone responsible. The contract is also versioned, because the position changes with products, policies, models and risks. Clause HWF-61 puts a cap on the clock: the contract is reviewed with a declared cadence, never greater than twelve months, and the decision to change it is always human.
Exactly one owner
The contract identifies the accountable owner: the human who answers for the design and the performance of the role. The standard is surgical with this figure, and it is worth inheriting its full precision.
First: exactly one. A role with two owners has none, and "the finance area" answers for nothing. Areas do not sign. It can be a person with a name or a human governance body, but a committee only counts as an owner if it has an identified chair, declared decision rules and the capacity to act in an emergency. Without those three things it is not an owner: it is a mechanism for blurring the blame.
Second: do not confuse the owner with the supervisor. Supervising (routing cases, reviewing outputs, prioritizing, receiving exceptions) is work, and it can be delegated, even to another artificial resource. Accountability is never delegated: every supervision chain, whatever links it has, ends in that identified human or body.
Third: one is primary, not exclusive. The owner does not extinguish the obligations of the security owner, of the data officer nor of the vendor. It answers "who answers for this resource?", not "who else has duties?".
An AI Employee with real credentials, official channels and the capacity to commit the company, but with no declared mission, no written exclusions, no escalation destination and no identified owner, is not an audacious bet. It is an incident waiting for a date. The only unknown is the day.
From paper to operation
A well-written contract still does not work. There remains a verification that almost nobody does and that separates serious design from optimistic design: proving that the authority defined on paper coincides with the real permissions. A resource can be forbidden to approve discounts in its contract and keep, by mistake, a credential that allows it. In the opposite direction, it can answer for a KPI without access to the information necessary to measure it. The distance between the contract and the configuration is exactly the space where incidents live.
And the contract defines the responsibility, but it does not make it executable with judgment: for that the role needs its second half, the governed context. First the position, then the assignment, then the contract, then the context. Only then is autonomy discussed. Nobody serious hands over the keys on the first day.
Frequently asked questions
The AI Role Contract, the role contract, is the operational document that defines the position an AI Employee occupies: what it exists for, what results it must produce, with what authority, inside what limits and under the responsibility of which human. It is not a legal piece: it is the functional combination of a good job description and an explicit agreement of operation. Its purpose is to convert technological capability into manageable responsibility: without a contract, the organization has a system with permissions, but not a governed position for which someone answers.
The glossary of the standard (G-03) defines it with this exact list: mission, responsibilities, results, KPIs, authority, exclusions, tools, accesses, service level, escalation, suspension criteria, governance calendar and accountable owner, all of it versioned. In practice four pieces stand out: the authority table with its explicit exclusions (what it cannot do is worth as much as what it can), the escalation destinations for every exception, the kill switch with someone responsible and a condition of use, and the governance calendar that consolidates all the review dates with a single owner of the schedule.
It is the human who answers for the design and the performance of the role, and the standard demands exactly one. It can be a person with a name or a human governance body, but a committee only counts if it has an identified chair, declared decision rules and the capacity to act in an emergency. "The finance area" answers for nothing: areas do not sign. The owner is not the supervisor: supervising is work and it can be delegated, even to another artificial resource, but accountability is never delegated. And it is primary, not exclusive: the obligations of security, data and vendor survive intact.
With the cadence the contract itself declares, and clause HWF-61 puts a cap on it: never greater than twelve months. The decision to change the contract is always human. Besides the periodic review, the governance calendar schedules the revalidations after relevant changes (products, policies, models, risks), the kill switch drills and the re-justification of the existence of the position. Versioning completes the cycle: it allows knowing what authority existed at a given moment and what modification produced an improvement or a deterioration.
The contract formalizes a position that was already designed and assigned: that order is explained in Work Resource Management. The second half of the role, the governed context, is in Context for enterprise AI.
Want the full method? Read AI Employee. For executive AI consulting or keynotes and workshops.
Go deeper
Want to bring your team to the next belt?
Book a discovery call or explore the full book.