Your first AI employee needs a job description, not another prompt
AI agents need defined outcomes, evidence, authority, escalation and ownership before they can become accountable operating roles.
Treating an agent as a reusable prompt leaves purpose, authority, evidence requirements, escalation and accountable ownership undefined.
A new AI agent often enters a business through a deceptively simple request: write a better prompt, connect a few tools and let it get to work.
That may be enough for a demonstration. It is not enough for an operating role.
The moment an agent can influence customers, records, decisions, budgets or other employees, the prompt is no longer the most important document. The important document is the job description around it: what outcome the agent owns, what evidence it may trust, which actions it can take, where it must stop and who remains accountable.
Without that structure, the business has not delegated work. It has accelerated ambiguity.
A prompt is not a role
Prompts tell a model how to respond in a moment. Roles tell an organisation what work is being done and under whose authority.
That distinction matters because business work is rarely a single exchange. It moves across systems and people. Information changes. Exceptions appear. A reasonable next step in one situation may be unacceptable in another. The agent needs more than an instruction. It needs a stable understanding of its purpose and boundaries.
Consider an agent asked to follow up with sales prospects. A prompt might define tone and ask for concise messages. A job description would go further. It would identify which prospects are eligible, which source records are current, what claims are permitted, when a message requires review, how opt-outs are handled and who owns the commercial outcome.
The prompt shapes the words. The role governs the work.
The cost of undefined authority
An agent without a clear job description creates two opposite risks.
The first is overreach. The agent acts beyond the authority the business intended. It sends a message, changes a record or advances a process without the right evidence or approval. Even when the output looks sensible, the organisation cannot show why that action was allowed.
The second is paralysis. Every meaningful action is routed back to a person because nobody has defined what the agent may do independently. Employees become permanent reviewers, copying information between systems and approving routine steps. The business appears to have automated the workflow while retaining most of its coordination cost.
Both problems come from the same source: authority was never designed.
This is why “keep a human in the loop” is incomplete advice. The organisation must decide which human, at what point, with what evidence and for what class of decision. Otherwise the human becomes a vague safety blanket rather than an accountable control.
What an AI job description must define
A useful AI job description does not need to resemble a legal contract. It needs to make the operating expectations explicit.
Start with six questions:
- What outcome does the agent own?
- Which sources may it use, and how current must they be?
- What can it do without approval?
- Which conditions require escalation?
- How will its work be checked and recorded?
- Which person remains accountable for the result?
These questions separate useful delegation from hopeful automation.
The outcome should be stated in business terms, not as model activity. “Draft emails” describes output. “Prepare accurate, policy-compliant follow-ups for eligible prospects” describes work.
The evidence requirement should identify authoritative sources and what to do when they conflict. The authority level should distinguish recommendations from reversible actions and high-consequence decisions. Escalation rules should name the trigger and the owner, not simply say “ask a human if unsure”.
Finally, accountability must remain visible. An agent can perform work, but it does not absorb organisational responsibility. A named owner must be able to review the role, change its boundaries and answer for its results.
From isolated tool to accountable operating role
Once the job description exists, the agent becomes easier to integrate with the rest of the company.
Its permissions can match its responsibilities. Its context can be limited to the information required for the role. Its work can produce a useful decision history. Corrections can improve the shared operating standard instead of remaining trapped in one employee’s chat history.
This is where company memory and governance become practical rather than abstract. The agent needs access to current decisions, policies and relevant relationships. It also needs standing rules that define how that context may be applied.
Walter’s point of view is that useful enterprise AI is not created by adding more isolated intelligence. It comes from connecting intelligence to company context, clear authority and accountable work. The organisation should be able to change the underlying model or tool without losing the role it has designed.
A durable role survives a prompt rewrite. It can survive a vendor change. It can also be understood by a new manager who did not build the original automation.
The practical test
Before introducing an agent, ask whether a capable employee could read its job description and understand what the organisation expects.
Could they explain the outcome? Could they identify the approved evidence? Could they tell which actions are permitted, which require review and who owns the consequences?
If not, the agent is not ready for operating responsibility.
The next generation of AI adoption will not be won by the company with the longest prompt library. It will be won by organisations that can define work clearly enough to delegate it without losing control, continuity or accountability.
Your first AI employee still needs intelligence. But intelligence becomes useful only when the job around it is designed.
