AI agent governance

Keep Humans in Control of AI Sales Agents

Give AI enough authority to be useful. Not unlimited authority because it can.

An AI agent can read information. Make recommendations. Prepare work. Use tools. Update systems. Send communications. And, with the right access, take actions without waiting for somebody each time.

The important question is not simply: "What can the AI do?" It is: "What should this AI be allowed to do?"

We design AI sales workflows with permissions, approval, limits and escalation built into the process.

Design a controlled AI workflow See how we work
The authority ladder lighting from READ up to ESCALATE
Autonomy is not an on or off switch

Autonomy is not an on or off switch.

There is a lot between chatbot and autopilot.

AI is often described as either: Something that waits for instructions. Or: Something autonomous that gets on with the job. Real business workflows need more nuance.

An AI system might be trusted to read a sales enquiry but not respond to it. It might prepare a CRM update but not make the change. It might send one type of routine communication automatically but require approval for another. It might handle normal situations independently and immediately escalate anything unusual.

Different actions deserve different levels of authority. That is why we use an authority ladder.

The AI authority ladder

The AI Authority Ladder

Start with the lowest level that makes the system useful.

01Read

The AI can access agreed information. It can understand what is happening and use that information as context. It cannot change anything.

ExampleAn AI agent reads a new enquiry and relevant CRM information, then presents the context to a salesperson.
Human controlThe salesperson decides what happens next.
↓
02Recommend

The AI can analyse the information and suggest a next action. It still cannot carry that action out.

ExampleThe agent recommends that an enquiry should be routed to a particular team and explains the information behind that recommendation.
Human controlA person accepts, changes or rejects the recommendation.
↓
03Prepare

The AI can prepare the work required for the next step.

ExampleIt drafts the response, prepares the CRM update and creates a suggested follow-up task. Nothing has been sent or changed yet.
Human controlA person reviews what has been prepared.
↓
04Act with approval

The AI can carry out an action once a person explicitly approves it.

ExampleThe salesperson reviews a prepared follow-up and approves it. The agent sends the message and records the activity.
Human controlThe person makes the decision. The agent carries out the work.
↓
05Act within limits

The AI can carry out specific actions independently when agreed conditions are met.

ExampleA known type of website enquiry arrives with all required information. The workflow categorises it, creates the appropriate CRM task and sends an agreed acknowledgement without waiting for approval.
Human controlThe business has already defined what the agent is allowed to do and under which conditions.
↓
06Escalate

When the situation falls outside the rules, the agent stops. It gathers the relevant information and passes the situation to a person.

ExampleThe enquiry contains conflicting information or an unusual commercial request. Rather than deciding anyway, the agent flags it for review.
Human controlThe exception returns to somebody with the authority and context to decide.
An agent can sit on several levels at once

An agent can sit on several levels at once.

Authority belongs to the action, not the agent.

This matters. You do not necessarily decide: "This agent is Level 4." You decide what authority it has for each action.

For example, an enquiry agent might be allowed to:

One agent. Different authority for different actions. That is much more useful than a single switch marked:

AUTONOMOUS: ON
Permission should follow the job

Permission should follow the job.

Access is not a prize for being intelligent.

When businesses experiment with AI agents, it can be tempting to connect everything first. Email. CRM. Calendar. Files. Internal documents. Customer information. Then work out what the agent can do with it.

We prefer the opposite approach. Define the job. Then ask what access that job actually requires.

If the agent only needs information from three CRM fields, why give it unrestricted access to the entire system? If it only needs to prepare emails, why give it permission to send them? If it only needs to read a calendar, why give it permission to change meetings? See how this plays out with CRM permissions.

Capability does not automatically justify permission.

Read access matters too

Read access matters too.

"It can't change anything" does not mean access is irrelevant.

An agent that only reads information can still potentially access sensitive or commercially important data. So read permissions should be considered deliberately.

The principle remains the same: Give the workflow the information required for its job, not everything that happens to be available.

Human approval should be meaningful

Human approval should be meaningful.

A button does not automatically create good oversight.

It is easy to add: Approve / Reject to an AI workflow. But approval only works if the person approving has enough information to make the decision.

If somebody is expected to approve fifty AI actions without understanding why they were proposed, approval can become another repetitive task.

Good human review should provide useful context.

Where practical, the workflow should make the human decision easier rather than simply move responsibility to an approval button.

Escalation is part of the design

Escalation is part of the design.

Not an admission that the AI failed.

A well designed agent should encounter situations it is not allowed to handle. That is normal.

Perhaps:

The agent's job in those situations is not to improvise. It is to recognise the boundary, gather useful context and escalate. Knowing when to stop is a capability too.

Define the exception before it happens

Define the exception before it happens.

"Ask a human if necessary" is not enough.

Where possible, we identify escalation conditions during workflow design. That might include:

Not every exception can be predicted. But the more clearly the boundaries are defined, the less the agent has to invent its own interpretation of what is acceptable.

Put limits around independent actions

Put limits around independent actions.

Autonomy works better when the boundaries are specific.

Suppose an AI agent is allowed to follow up with sales leads. That instruction is too broad. A more controlled workflow might define:

The useful part is not simply that the AI can send an email. It is the system around that action.

What about mistakes?

What about mistakes?

Design as though they can happen.

AI systems can misunderstand information. Generate something incorrect. Miss context. Choose the wrong permitted option. Encounter information they were not expecting.

Connected systems can fail too. APIs return errors. Data becomes stale. Records conflict. Automations trigger unexpectedly.

That is why we do not design workflows around the assumption that everything will always behave perfectly. Depending on the workflow, controls might include:

Human approvalRestricted permissionsValidationDefined rulesLoggingEscalationTestingFallback behaviourMonitoring

The appropriate controls depend on what the system is doing and the consequences if something goes wrong.

The higher the consequence, the stronger the case for control

The higher the consequence, the stronger the case for control.

Not every action carries the same risk.

Adding an internal label to a lead is different from sending a contractual commitment. Preparing meeting notes is different from changing a commercial agreement. Suggesting an opportunity stage is different from deleting customer information.

That difference should affect how much authority the AI receives. A useful question is: "What happens if this action is wrong?"

If the answer is: "Someone corrects a low impact internal field." the workflow may tolerate more independence. If the answer involves a customer, money, commitments, sensitive information or significant consequences, stronger controls may be appropriate.

More autonomy is not automatically progress

More autonomy is not automatically progress.

Level 5 is not better than Level 3.

The authority ladder is not a maturity score. A workflow does not need to climb it.

Some AI systems may remain permanently at Recommend or Prepare because that is the right place for them. Others may handle narrow, predictable actions independently.

The goal is not to reach maximum autonomy. The goal is to find the level that makes the workflow useful without giving away unnecessary control.

Autonomy should earn its place. And sometimes it should stay exactly where it is.

Seven useful questions

What should you ask before giving an AI agent authority?

Seven useful questions.
  1. What is the job?Can you describe exactly what the agent is responsible for?
  2. What information does it need?Not what could it access. What does it actually need?
  3. What can it change?Which systems, records and actions are writable?
  4. What can it do without approval?Be specific.
  5. What always requires a person?Define important decisions and actions explicitly.
  6. When must it stop?What conditions trigger escalation?
  7. What happens if it is wrong?Understand the consequence before deciding the level of autonomy.

If those questions do not have clear answers, the agent probably needs more design before it needs more authority. A sales process diagnostic is often the right place to start.

Governance should live inside the workflow

Governance should live inside the workflow.

Not in a document nobody reads.

Policies matter. But controls become much more useful when they are reflected in how the system actually works.

The aim is to turn principles into behaviour.

Human control does not mean humans doing everything

Human control does not mean humans doing everything.

The point is to put people at the right points.

A workflow where a person manually checks every tiny action may remove very little work. A workflow where AI can do anything it likes may create unacceptable uncertainty. The useful space is between those extremes.

That is controlled agentic selling.

How we design control into AI sales workflows

How we design control into AI sales workflows.

01

Define the job

What is the AI actually responsible for?

02

Map the information

What does it genuinely need to access?

03

Break down the actions

What can it read, recommend, prepare and change?

04

Assign authority

Which actions require approval and which can happen within limits?

05

Define escalation

When must the system stop and involve somebody?

06

Test normal and abnormal situations

What happens when information is incomplete, unusual or conflicting?

07

Introduce authority carefully

Start at an appropriate level and increase specific permissions only where there is a reason.

08

Review behaviour

Look at what the agent actually does, where people intervene and whether authority needs to increase or decrease.

This is the methodology we use on every workflow.

Design your AI workflow

Frequently asked questions

Frequently asked questions.

What is AI agent governance?
AI agent governance is the set of rules, permissions, controls and oversight used to determine what an AI agent can access, decide and do. In a practical workflow, this can include permissions, approval requirements, action limits, escalation rules, testing and monitoring.
Should AI agents always require human approval?
Not necessarily. Some low risk, well understood actions may be appropriate to perform independently within defined limits. Other actions may always justify human review. The appropriate level depends on the task and consequences.
Can an AI agent have different permissions for different actions?
Yes. That is how we prefer to design workflows. An agent may have permission to perform one internal action automatically while requiring approval for another.
What does human in the loop mean?
Human in the loop generally means a person remains involved in part of the AI decision or action process. That might involve reviewing recommendations, approving actions or handling situations the AI escalates.
What happens when an AI agent is uncertain?
The workflow should define what happens when the agent cannot safely or reliably continue. In many cases, the appropriate action is to stop and escalate the situation to a person with the relevant context.
Should AI agents have access to all business data?
No. Access should be based on what the workflow actually requires. The fact that information is technically available does not mean an agent needs access to it.
Can we increase an agent's autonomy later?
Yes. A workflow can begin with more human review and later give specific actions greater authority if there is evidence that doing so is appropriate.
Can we reduce an agent's authority?
Yes. Authority should not only move in one direction. If real use shows that an action needs more oversight, permissions or autonomy can be reduced.
Want the deeper thinking?

The Agentic Selling knowledge hub explains the principles behind agent authority, human oversight and responsible AI in sales.

Explore Agentic Selling →
Do not ask how autonomous your AI can be.

Do not ask how autonomous your AI can be.

Ask how autonomous it needs to be.

For this job. For this action. With this information. Under these conditions. With these consequences.

Then give it the authority that makes sense. No more.

Design a controlled AI workflow See how we work