Voices · · 7 min read

Your AI Can Act. Who Gave It Permission?

Giving an AI system access isn't the same as giving it authority. If the action, scope, evidence or timing changes after approval, the approval has expired, even if the software still says yes.

Chief systems operator, C. Misselwitz & C. Wheeler letsrocc GbR

Share

Your AI can have the right credentials, pass every technical check and still make a decision it had no business making. That is the point many organisations miss. Identity and access management, or IAM, can prove that a system was allowed to use a tool. It cannot prove that the business authorised this exact action, against these customers, with this evidence, at this time.

If the action, scope, evidence or timing changes after approval, the authority has expired. A green tick doesn't revive it. I've spent years working across ecommerce, marketing, product, technology and operations, often where the numbers and the operating story have stopped agreeing. The same fault line appears again and again: software can execute a decision long after the evidence or authority behind it has moved on. AI makes that old problem faster. It doesn't make it new.

My commercial interest

I co-founded letsrocc, where we diagnose the gap between reported performance and commercial reality. AI-enabled workflows increasingly sit inside that gap, so I have a commercial interest in the subject. That doesn't prove my argument. The control principles below are supported by public guidance from the NCSC, NIST and OWASP, and by a simple operating test: can you show that the thing executed was still the thing a named person approved? If you can't, you have evidence of activity, not authority.

Permission is not authority

Imagine giving someone a boarding pass for the 08:30 flight to Berlin, then watching them use it to board a different aircraft three days later. The barcode might still scan. The authority clearly shouldn't. AI approvals need the same boundaries. One decision. One scope. One period of validity.

Yet many workflows treat approval as a reusable hall pass. A system observes a change, recommends an action and sends it to a person. The person clicks approve. By the time the action runs, the data may have changed, the audience may have widened or another system may have translated the recommendation into a materially different instruction.

The approval remains green while the decision it covered has quietly changed underneath it. You wouldn't accept that from a finance process. A purchase order for one supplier and one amount doesn't authorise any supplier and any amount. You shouldn't accept it because the person asking is now a model and the paperwork is an audit log.

Four handoffs, not one smooth workflow

The language around AI tends to compress four different decisions into one neat diagram.

Observe

The system reads records, messages, files, market data or operational events. You need to know what it may see, under whose identity and for how long.

Recommend

The system interprets that evidence and proposes a next step. You need to know whether the evidence is current, complete enough and relevant to the commercial question.

Approve

A person accepts, rejects or changes the proposal. You need to know whether that person has the decision rights, evidence and time required to own the consequence.

Act

The system creates, changes, publishes, sends, buys, pauses or deletes something. You need to prove that the executed action is still inside the approved scope.

These handoffs are connected, but they aren't interchangeable. Read access doesn't include authority to recommend. A recommendation doesn't include permission to act. Approval doesn't survive a material change in scope or evidence. Successful execution doesn't prove that the decision was commercially sound. If you record only the final click, you're filming the finish line and pretending you saw the whole race.

Your RBAC cannot own the consequence

Role-based access control, or RBAC, answers a technical question: can this identity perform this operation in this system? Business authority answers a harder one: who is entitled to accept the customer, financial, legal, reputational or operational consequence? Those rights often sit in different rooms.

A marketing leader may control campaign spend but not customer eligibility. Finance may own the budget ceiling but not the promise attached to the offer. Operations may stop a process but not approve a public claim. Technology can configure every permission while owning none of the commercial decision.

I've seen this disconnect without AI. Teams inherit access because somebody needs the work done. Agencies are asked to own outcomes while being denied the margin, returns or customer data needed to judge them. Dashboards become a substitute for the argument nobody has resolved. Add AI and the ambiguity becomes executable. The system inherits whatever authority its credentials happen to expose. That isn't delegation. It is an unresolved organisation chart wearing an API key.

OWASP describes excessive agency as excessive functionality, permissions or autonomy. Its guidance includes restricting tools and permissions, requiring approval for high-impact actions and enforcing authorisation in the downstream system. That last part is crucial. Asking the model to remember its limits is like putting the speed limit in the driver's notes and removing the brakes.

A human in the loop can still be theatre

Human in the loop, or HITL, sounds reassuring. It is useful only if the human can make a real decision. Before you approve a material action, you need to see:

  • the business decision, not merely the tool call
  • the customers, records, systems, budget or markets affected
  • the evidence used and what remains uncertain
  • the strongest alternative explanation
  • what happens if you delay or reject the action
  • how the action can be stopped, contained and reversed

You also need permission to say no. If rejection creates a political penalty, misses a target or leaves you defending a delay without access to the evidence, the approval screen is not collecting judgement. It is collecting a name to put on the invoice when something goes wrong.

This is why polished recommendations can be more dangerous than obviously poor ones. The conclusion arrives clean and confident because the system has already selected, compressed and framed the evidence. You see the plated meal, not what happened in the kitchen.

In a separate analysis, I explain how to follow the evidence behind an AI-generated claim before it becomes action. The point isn't that every output needs a forensic investigation. The evidence burden should rise with the consequence and irreversibility of the decision.

The control must travel with the action

An audit log is necessary. It isn't a get-out-of-jail-free card. A useful log must show more than who clicked approve. It must show what was approved, which version was executed, what scope applied, which evidence supported it and whether a material condition changed before action.

For a consequential workflow, I would bind six fields together:

  • Decision: the commercial or operational choice being made
  • Object: the exact action, customer group, record set, market or budget affected
  • Evidence threshold: the minimum evidence required and the conditions that invalidate it
  • Owner: the named person entitled to accept the consequence
  • Expiry: the time limit or change that forces a fresh decision
  • Recovery: the person and method for stopping, reversing or containing the action

Think of those six fields as the chain of custody for a decision. Break one link and you no longer know whether the authority that entered the workflow is the authority that reached the outside world.

The NCSC's guidance makes the same problem practical. It recommends least privilege, tightly bounded scope, named responsibility, active monitoring, approval gates for higher-consequence actions and the ability to halt autonomous activity. NIST's AI Risk Management Framework puts governance, context, measurement and management into one operating model rather than pretending the model is the whole system.

The principle is simple: authority must be specific, inspectable and revocable.

The strongest objection is speed

The obvious objection is that this creates friction. Bind every action to six fields and useful automation could become a queue of forms, meetings and people protecting their backs. That would be poor control. An internal summary, a reversible classification or a draft for review shouldn't carry the same burden as a price change, customer communication, supplier decision, production deployment or movement of money.

Scale the control to four things:

  • Reach: how many people, records or systems can be affected
  • Consequence: what customer, financial, legal or operational harm could follow
  • Reversibility: how quickly and completely the result can be undone
  • Uncertainty: how much important evidence remains unresolved

Low-consequence work can use bounded autonomy, sampling and after-the-event review. Higher-consequence work needs narrower access, segregation of duties, or SoD, explicit evidence thresholds, named approval and a tested kill switch. The aim isn't to make AI slower. It is to stop speed from smuggling authority past the people who are supposed to own the decision.

What to change this week

Choose one live AI-enabled workflow that can affect a customer, money, reputation or operations. Don't begin with a policy covering every imaginable use case. That is how you end up wallpapering the boardroom while the pipe is still leaking. Take one recent or proposed action and follow it through observe, recommend, approve and act.

Ask:

  • Did the executed action match the exact object that was approved?
  • What evidence threshold had to be met?
  • Who owned the commercial consequence, not just the system configuration?
  • What change would have expired the approval?
  • Could an authorised person stop it before the consequence left the organisation?
  • Could the action be reversed while reversal was still useful?

If one answer is missing, narrow the workflow before expanding its authority. Your AI may have valid credentials, a human approval and a pristine audit log. That still doesn't prove it had permission to make this decision. Before it acts, make the system prove which authority it is using, who granted it and whether that authority still exists.

Share

About the author

Chris Wheeler

Chris Wheeler

Chief systems operator, C. Misselwitz & C. Wheeler letsrocc GbR

Chris Wheeler is Chief Systems Officer and co-founder of letsrocc, a senior-only commercial diagnosis and intervention practice. He has spent more than three decades across ecommerce, marketing, product, growth, SEO, technology and operations, working agency-side and client-side and carrying responsibility for P&L and execution. He is brought in when marketing activity and dashboards look convincing but revenue, margin or customer behaviour tell a different story. Through the Marketing MRI, he examines the commercial system behind the numbers, isolates the constraint and helps leadership decide what to fix, stop or sequence next. His view is simple: marketing should earn its place in the P&L.

View profile →

More from Chris Wheeler

Become a Voice

What matters. What doesn’t. What others are missing. What needs challenging. And what comes next.

Voices is where operators, thinkers, researchers, founders, creatives and challengers make sense of what is happening in marketing right now.

Become a Voice →
The Rewrite

Stay close to the conversation.

The conversations and intelligence worth knowing about.

More from Voices

All Voices →
A close-up of a white grid of repeating oval pixels with blue edges and iridescent streaks, representing the measurement complexity behind CTV advertising.

Decide what counts as proof before your CTV campaign launches

Connected TV (CTV) advertising is often measured in two misleading ways. Last-click reporting undervalues its impact because viewers typically respond later on another device. View-through reporting can overstate it by attributing conversions to ads that may not have caused them. The solution is to agree what counts as proof before the campaign launches. Lomit Patel, Growth & marketing leader
A man in a grey jacket crosses a narrow woven-grass rope bridge over a rocky gorge, representing the risk of optimising for short-term gain.

When AI Optimizes for Conversion, Customer Lifetime Value Pays the Price

AI can predict when consumers are most likely to buy. But when conversion algorithms exploit moments of vulnerability, they can undermine the customer lifetime value and brand trust they are designed to grow. Lomit Patel, Growth & marketing leader
Pathsways going in different directions

The question your buyer asks decides whether your asset is the answer

Your brand asset can be accurate, sitting on a domain you control, and completely correct, and still not be the answer anyone gets. The gap between those two states has become a measurable thing, and it is not where most teams are looking. Ipalibo Da-Wariboko, Founder & CEO, CULT+MATH
A blurred black-and-white tunnel converging on a bright white light at its centre, representing the risk of reconstructed customer identity from combined data.

Are You Giving Away Your Customers’ Identities?

Removing an identifier does not necessarily remove the identity. Identity is not always something you hand over; it can also be an output, produced when separate pieces of data are connected. Dela Quist, Strategic advisor, writer & speaker, DQ Consulting