AI agents & decision-making
Video · 5 min read
The short answer
Watch: Are AI agents already making strategic decisions on behalf of leaders?
Watch the explanation, then explore the practical implications below.
Can an operational task contain strategic choices?
Yes. A task can sound like routine execution while embedding a choice about what matters to the business. Selecting which enquiries deserve attention, which evidence appears in a report or which option is explored first can influence subsequent decisions.
The video highlights this issue without implying that every agent has independent authority over company strategy. The actual authority depends on how the workflow is configured, which tools it can use and where people approve its work. Inspect those details rather than relying on the label “agent”.
A practical example: prioritising customer enquiries
Suppose an agent is asked to rank new enquiries. “Best opportunity” could mean the largest potential contract, the highest expected margin, the strongest fit with delivery capacity or the type of customer the business wants more of. These priorities can produce different rankings.
If the criteria are unspecified, a convincing explanation may conceal an assumption you would reject. Agree the criteria with the commercial owner, define the information supporting each one and decide what happens when evidence is missing. Avoid treating an estimated probability or margin as an established fact.
Surface assumptions before they shape commitments
Ask the system to show the criteria it used, cite the supplied evidence and separate missing information from negative evidence. An enquiry with no recorded budget should not automatically be treated as a poor prospect unless that is an explicitly agreed rule.
Request alternatives and identify which assumptions would change the recommendation. Inspect the actual inputs and outputs as well as the explanation. A well-written account of a decision does not by itself prove that the underlying comparison was complete or correct.
Separate preparation, recommendation and action
An agent could prepare a shortlist for review without contacting customers. It could draft follow-up questions without sending them. It could act within a limited, approved workflow while escalating exceptions. These are different levels of delegation and should be designed deliberately.
Specify the actions it may take, the records it may access and the conditions requiring approval. Enforce those boundaries through permissions and workflow controls where possible. A sentence asking the model to be careful should not be the only control on a consequential action.
Test how the workflow behaves at the boundaries
Include cases where the largest opportunity has poor fit, a smaller enquiry has strategic value, information is incomplete or two records conflict. Check whether the system follows the agreed criteria and whether the reviewer can see why a case was escalated.
Also test interruptions and failed connections. Determine what has already happened before retrying an action, and make exceptions visible to an owner. A workflow should not create duplicate commitments or conceal a failed step behind an apparently complete summary.
Review outcomes and maintain the criteria
Record recommendations, overrides and relevant outcomes. Examine why people changed the ranking or rejected an action. Repeated disagreement can reveal unclear priorities, missing evidence or a rule that no longer reflects the business.
Review work that was not escalated too. Strategic priorities can change, so the person responsible for the workflow should approve updates to its criteria and limits. Do not allow every conversational correction to become a permanent company policy automatically.
How CSM would approach an agent workflow
We would start by mapping the choices inside the task with the business owner. Together, we would agree what a good result means, which evidence supports it and which actions require human authority. The first scope should be narrow enough to test against representative work.
We would then build review, monitoring and recovery into the process. The opportunity is to reduce preparation and coordination while keeping consequential assumptions visible. Leaders should be able to understand what has been delegated and intervene before an unsupported recommendation becomes a business commitment.
Questions for your leadership team
What choices are hidden inside this task?
List how the agent selects evidence, ranks options and decides what to omit.
Have we defined what “good” means?
Agree the criteria, their relative importance and the evidence required to apply them.
What can the agent recommend and what can it execute?
Make the distinction explicit in the workflow and its permissions.
What happens when the evidence is missing or conflicts?
Specify the escalation route and what must be resolved before action.
Can we trace and challenge the result?
Keep the source evidence, recommendation, approval and action available for proportionate review.
Full video transcript
The full video transcript is included below for reference.
If you give an agent a task that involves several steps, it has to make choices along the way. What information should it look at? Which option should it explore?
What does it consider important enough to include or leave out? Some of those are routine choices. Some can have strategic consequences.
So I think we need to be careful about saying the AI is “just executing” as if nothing meaningful is being decided. Take a task like recommending which customer enquiries deserve attention first. To do that, the system needs some view of what makes an enquiry valuable.
Is it size? Probability of winning? Margin?
Whether it fits the business you’re trying to build? If you haven’t made that clear, it may still produce a convincing recommendation. But there’s a judgement inside that recommendation that you haven’t necessarily agreed with.
That doesn’t mean you shouldn’t use it. It means you should decide which parts of the decision you’re delegating. You can ask it to bring you several options, explain the reasoning behind each and identify what would have to be true for the recommendation to make sense.
You can ask it to challenge its preferred option. You can tell it that the final choice belongs to you. For other work, you might be comfortable letting it act within clear limits.
The important thing is that those limits reflect the consequences of getting it wrong. I think one of the most useful habits is asking, “What did you assume in order to reach this answer?” Because the answer can look very good while being built around a priority you would never have chosen. The process should let you see and challenge those assumptions before they become commitments the business has to live with.
Put these questions to work in your business
Bring a real decision or workflow to a conversation with CSM. We can explore the evidence, responsibilities and practical steps needed to make AI useful for your team.