The day an AI agent goes live, most job descriptions in the affected team become wrong. They describe work the agent now drafts and leave out what people actually own: reviewing held actions, deciding refusals, keeping certificates current. Update them before go-live, not after, because accountability follows the job description whether it is accurate or not.
What changes in a job the moment an agent goes live?
The verbs. A claims handler's description says "assesses and settles claims". After go-live, an agent reads the claim, drafts the assessment and proposes a settlement at depth 2, perhaps executing small settlements at depth 3 once certified. The handler now reviews held settlements, decides the borderline ones, and handles everything the agent refused or was not allowed to touch.
That is a different job. It needs more judgement, less keying, and a new set of skills: reading provenance, understanding verdicts, spotting when a confident draft is wrong. The old description still says "assesses and settles", and so performance reviews, grading and training plans all aim at the wrong target.
Why does an out-of-date job description create risk?
Because accountability lives in it. When something goes wrong, organisations look at who was responsible for what. If the description says the handler settles claims, but in practice the agent settled it at depth 3 and nobody was named to review that category, there is a gap no one owns.
The record will show what happened. Every verdict and every human decision is filed in an append-only record, so it is clear which actions the agent took, which a person approved and who that person was. What the record cannot show is whether that person was supposed to be checking. Only the job description, and the delegated authority that sits with it, can answer that. Regulators asking about named accountability, such as under the UK's Senior Managers and Certification Regime, will want both to line up.
What should a post-agent job description include?
Five additions cover most roles:
- Review authority. Which held actions this person may approve, up to which limits, and which must go higher.
- Override authority. Whether they may override a refusal, under what conditions, and with what record.
- Certificate requirement. If the business makes approval rights depend on a current certificate, say so, with the consequence of a lapse.
- Agent oversight duties. Reporting stale sources, recurring draft errors and anything that suggests the agent's brief needs change.
- What stays with the person. The decisions the agent will never take, such as red lines and withheld actions handled manually.
A reviewer job description template sets these out in a form HR can adapt.
What is the trade-off in rewriting job descriptions?
Time, consultation and money. Changing job descriptions can trigger grading reviews, consultation with staff representatives and, in some places, formal processes. A role that moves from processing to judging may deserve a higher grade. Finance will notice.
The alternative is worse but cheaper in the short term: leave the descriptions alone and let practice drift. People then do one job while being measured on another, and when an error happens nobody can say whether it was theirs. The cost appears later, in a complaint, an audit finding or a disciplinary process that collapses because the job description did not say what the person was meant to do.
A middle course works for many organisations: add an oversight annex to existing descriptions at go-live, covering the five points above, and schedule a full rewrite after the first few months, once the shape of the work is clear.
How should HR and operations work together on this?
Operations knows what the work now looks like; HR knows how to describe and grade it. Neither can do it alone. The best moment to start is during the shadow period, when the agent runs alongside people without acting and the future shape of each role becomes visible. The guide on how to shadow an AI agent before it goes live explains that period, and how to test an AI agent before it goes live covers the evidence that comes with it.
The underlying shift is described in the new frontline job is judging exceptions. For a practical walk-through of how individual descriptions change, see how job descriptions change when AI agents do part of the work, and for the skills each role needs, AI agent skills by role.
What should you do differently?
Treat job descriptions as part of go-live, with an owner and a deadline. List every role that will review, approve or override agent work. For each, write the five additions above before the agent acts on live cases. Tie review authority to training and, where you choose, to a current certificate. Then revisit after the first month, once the record shows what people actually spend their time on.
Do not wait for the annual review cycle. By then, people will have been accountable for months for work no document describes, and some will have been appraised against targets that no longer make sense.
Where the Academy fits
Pryme Intelligence Academy maps cleanly onto revised job descriptions. Foundations suits anyone affected; Certified Pryme Operator suits reviewers and team leads; Certified Pryme Builder suits those who design and certify agents. Certificates expire, so a job description that requires a current one stays honest about competence over time.
Questions people ask
Should job titles change when an agent goes live?
Not always. A claims handler may still be a claims handler. What must change is the description of responsibilities and authority, because that is what accountability rests on.
Do people need to agree to job description changes?
That depends on employment contracts and local law. Many changes need consultation, and some need agreement. Involve HR early and take advice where needed.
How often should these job descriptions be reviewed?
Whenever an agent's depth changes for a capability the role oversees, and at least once a year. A depth increase moves work from review to sampling, and the description should reflect that.