The Dealer Still Decides · Article 2 · Every Agent Needs a Job Description
Nobody would hire an employee, hand them every password in the dealership and say, “You seem smart. Go find something useful to do.”
There would be a role. A set of responsibilities. Systems the employee could access. Decisions they could make without asking. Decisions requiring approval. People to escalate to when the situation exceeded their authority.
The employee would not need to understand the company’s entire org chart to do useful work. They would need enough context to perform a defined job well.
Agentic AI deserves at least that much organizational clarity.
As intelligent systems move beyond answering questions and begin participating in workflows, the useful unit of design is no longer simply “the AI.” It is the role the AI has been asked to perform inside the business.
An AI agent should not be defined by how intelligent it is. It should be defined by the job the organization has authorized it to do.
That distinction sounds obvious until a prototype becomes useful.
A dealership builds an assistant to analyze inventory. It works. Then someone connects pricing data. Then market intelligence. Then the ability to draft merchandising recommendations. Then somebody reasonably asks why it cannot simply make the changes itself.
Nothing about that progression is inherently wrong. In fact, it is how useful systems evolve.
The risk appears when capability expands faster than the role around it.
A job description solves that problem before it becomes an access-control problem.
The current AI market encourages organizations to begin with capability. A new model appears, a vendor demonstrates something impressive, or an internal builder discovers that an agent can suddenly use a collection of tools. The natural response is to start looking for places to deploy it.
That approach can produce excellent experiments. It is less reliable as an operating model.
Organizations do not normally begin workforce design by finding the smartest available person and then searching for something they might be able to do. They begin with an operating need.
The dealership needs someone to monitor aged inventory and surface meaningful exceptions. It needs help preparing customer-facing vehicle content. It needs an analyst that can reconcile performance signals before the Monday meeting. It needs a system capable of identifying repeated service questions so the organization can create better customer education.
Those are jobs.
Once the job is clear, model selection becomes one implementation decision among several. The organization can determine which resources the agent needs, which tools are appropriate, how often the work should occur, what success looks like and which outputs should reach a person before anything happens downstream.
This framing also prevents the agent from quietly becoming responsible for everything adjacent to the original task.
An inventory analyst does not automatically become a pricing manager. A content assistant does not automatically become a publisher. A customer-response agent does not automatically gain authority over the CRM simply because it can read from it.
The job should define the agent’s authority. The agent’s capability should not be allowed to define the job.
That reversal matters because AI capability will continue expanding. The role gives the organization a stable boundary even while the underlying technology gets better.
Autonomy becomes substantially less intimidating when it is attached to a well-defined scope.
A dealership may be perfectly comfortable allowing an agent to review every vehicle in inventory overnight and flag the ten that deserve human attention. The same dealership may be uncomfortable allowing that agent to adjust every retail price autonomously.
Those positions are not inconsistent.
They reflect different scopes of authority.
A useful agent job description should make several dimensions explicit: which objects the agent can work on, which systems it can access, what actions it may take, the conditions under which it can act without approval and the point at which the work should move to a person.
That turns autonomy from an ideological question into an operating decision.
Consider a dealer group using an intelligent system to improve content around aged inventory. The agent might be permitted to identify candidate vehicles, assemble current vehicle and market context, recommend a content angle and prepare a draft. The dealership could choose to let that workflow proceed automatically until publication, where an employee remains responsible for the final customer-facing output.
Another organization may be comfortable automating publication for certain content classes while requiring approval for others.
Neither organization is inherently more advanced.
The sophistication is in knowing why the boundary exists.
Autonomy works best when the organization can describe its boundaries more clearly than the agent can discover them.
That clarity also makes experimentation easier. Authority can expand intentionally as the organization gathers evidence that the agent performs well inside its existing role.
The result is not a binary choice between manual work and full autonomy. It is progressive delegation.
A defined job also answers one of the most difficult questions in AI architecture: how much organizational context should the system receive?
The answer should depend on the work.
An agent responsible for helping a salesperson prepare a vehicle explanation may need current vehicle specifications, inventory context, dealership brand guidance, local market information and recurring questions associated with the model. It probably does not need unrestricted access to accounting, payroll or every customer record in the CRM.
An executive analysis agent may legitimately require broader operating context while still having no reason to modify the underlying systems it reads.
This is the principle we established in Keep the Ledger Where the Ledger Belongs: useful intelligence does not require making every intelligent environment another permanent home for the dealership’s authoritative records.
A job description makes that principle actionable.
The organization can begin with the responsibility and work backward into the minimum set of resources necessary to perform it well.
That is considerably healthier than connecting everything first and deciding later what the agent probably should not have seen.
It also creates a cleaner relationship between the permission layer and the dealership intelligence layer.
The intelligence layer makes organizational context durable and interoperable. The permission layer determines which slice of that context belongs inside this particular role.
Hrizn MCP provides a practical expression of that architecture: authorized intelligent environments can reach useful dealership context without requiring the dealership to rebuild itself independently inside each interface.
The right question is not, “How much data can this agent reach?” It is, “What does this job need to know in order to perform responsibly?”
That shift reduces unnecessary access while usually improving the quality of the work. More information is not automatically more relevant information.
Human escalation is often discussed as though it represents failure.
A good operating organization knows better.
Escalation exists because some conditions fall outside the authority or judgment assigned to a role. An employee who recognizes those conditions and brings the issue to the right person is not failing to perform the job. They are performing the job correctly.
An intelligent agent should be designed with the same expectation.
The useful question is not whether the agent can eventually handle every edge case. It is whether the organization has defined the situations in which continuing autonomously would exceed its authority.
A customer request may involve a representation the agent should not make. An inventory recommendation may cross an agreed financial threshold. A content workflow may encounter a compliance concern. An agency automation may discover conflicting instructions between the dealership’s policy and the task it has been asked to perform.
Those are not necessarily reasons to stop using the agent.
They are reasons to know where the agent stops.
Good escalation architecture can also preserve momentum. Instead of failing silently or handing the entire task back to a human, the agent can prepare the relevant context, explain what triggered escalation and present the decision in a form that makes human review faster.
That is a much more useful version of “human in the loop.”
The human does not exist to babysit every ordinary action. The human receives the decisions where discretion, consequence or ambiguity justifies the interruption.
An agent that knows when to stop can be more valuable than one designed to continue at all costs.
This becomes increasingly important as dealership agents operate continuously. The system that knows its boundaries is substantially easier to trust with more responsibility over time.
There is a tendency to describe autonomous agents as though they will become a parallel workforce hovering above the existing organization.
That framing obscures the more practical opportunity.
The best agents will likely fit into the dealership’s operating model much the way good technology always has: by extending the capability of the people responsible for the outcome.
A salesperson using Hrizn Creator does not stop being responsible for communicating credibly with the customer because AI helped prepare the script. The technology makes the employee more capable while the customer-facing role remains recognizably human.
An inventory agent can reduce hours of analysis without replacing the manager’s responsibility for inventory strategy. A marketing agent can prepare recommendations continuously without turning campaign accountability into a property of the model. An executive assistant can synthesize organizational intelligence without becoming the executive.
The job description tells the agent where it fits.
It also tells everybody else.
That matters as organizations begin operating with multiple intelligent systems. Without clear roles, agents can overlap, issue conflicting recommendations, duplicate work or act on assumptions another system was responsible for maintaining.
The problem begins to resemble poor organizational design remarkably quickly.
Because it is poor organizational design.
The difference is that software can reproduce the confusion at machine speed.
The future dealership will not merely have AI agents. It will have an organizational model that explains where those agents belong.
That is why the job-description metaphor is more than a convenient way to talk about permissions.
It forces leadership to answer the questions that matter before capability obscures them.
What is the role?
What outcome does it own?
What does it need to know?
What can it change?
Where does its discretion end?
Who receives the escalation?
And who remains accountable for the result?
Once those answers are clear, the AI becomes easier to govern because the organization has already decided what kind of participant it is supposed to be.
This is also where the next article in the series begins.
A job description establishes the role.
But consequential work still needs a deliberate answer to another question:
When should the organization require approval before the agent acts?
Every Agent Needs a Job Description is Article 2 of The Dealer Still Decides: 52 Weeks Into the AI-Native Dealership.
Earlier in the series:
Automation Is Easy. Authority Is Hard. →
Why capability and organizational permission become different problems once intelligent systems can act.
Next: Approval Is Architecture — why effective human oversight should be designed around consequence, reversibility and organizational discretion rather than added indiscriminately to every AI workflow.
Interoperate with Hrizn MCP →
Give authorized intelligent environments a governed path back to dealership context.
Explore Hrizn Creator →
See how AI guidance and human expertise can participate together inside a governed frontline workflow.
Explore the Hrizn Resource Library →
Practical guidance for operators working through governance, interoperability, AI visibility and responsible automation.
Apply to Hrizn Banditworks →
Build and pressure-test new ideas alongside operators actively working at the edge.
Let Dealers Plug In →
Support an interoperable automotive ecosystem where dealer choice remains part of the architecture.
The goal is not to give every agent every capability. It is to give each intelligent system the context, tools and authority appropriate to the job—and make those boundaries easier to govern as the organization scales.
Define the role. Connect the context. Expand authority when it is earned.
Free Around and Find Out.
We Rise Together.
