AI agents & delegation
Video · 5 min read
The short answer
Watch: How is working with AI changing from going back and forth with a chatbot to assigning work to a team of agents?
Watch the explanation, then explore the practical implications below.
What changes when you delegate a piece of work?
In a conversational workflow, a person often directs each step: ask a question, inspect the response, add context and ask again. In an agent workflow, the system may carry out more intermediate steps using the tools and information made available to it.
The amount of autonomy depends on the implementation. An agent is not automatically authorised to contact customers, change records or commit the business. Define those responsibilities explicitly rather than assuming that a broad task description settles them.
Write a brief that can survive without constant correction
Explain the purpose of the work and the decision it will support. Specify the relevant scope, approved sources, constraints, output format and completion criteria. Include situations where the system must stop or ask for clarification.
A weak brief such as “research our competitors” leaves the system to guess what matters. A stronger brief names the competitors or market, the customer segment, the decision being considered and the evidence required. It asks the system to separate observations from inference and to flag gaps it cannot resolve.
An example: preparing a competitor comparison
An illustrative task might be to compare the publicly stated service offerings of three named competitors for an upcoming positioning discussion. The output should cite source pages, record when they were accessed, distinguish published statements from assumptions and identify questions that remain unanswered.
The task ends with a reviewed comparison and options for discussion. It does not include contacting competitors, submitting forms or changing the company’s offer. Those are separate actions with different responsibilities. A useful brief makes that boundary visible from the start.
Use several agents only where the work benefits
Research, analysis and drafting can be separated, but adding agents also creates handovers and opportunities for assumptions to spread. The same unsupported claim repeated by several agents is not independent confirmation.
Give each part a clear input and output. Keep source evidence attached when work passes between steps, and identify who checks the final result. Begin with the simplest workflow that meets the need; use multiple agents when the division of work has a clear practical benefit.
Plan for partial results and exceptions
Define what happens when a source is unavailable, information conflicts or the time allowed is exhausted. The system should return a useful account of what it completed, what remains uncertain and what requires a decision, rather than present an incomplete task as finished.
Where the workflow can take actions, record what was attempted and what succeeded. Before retrying a failed step, establish whether it already took effect. Appropriate access limits and approval controls should support the brief, especially where an action would be hard to reverse.
Review the outcome against the original purpose
Check factual support, completeness and relevance to the decision. Assess the net effort, including briefing, review and corrections. A long output or an impressive demonstration does not establish that delegation improved the work.
At CSM, we would start with a real recurring task and a named business owner. We would agree the brief, evidence, approval points and acceptance criteria, then test representative and difficult cases. The aim is less coordination in the middle of the process, supported by clearer direction at the start and meaningful judgement at the end.
Questions for your leadership team
What complete piece of work are we delegating?
Define the result and the point at which the task is finished.
What may the system read and do?
List approved sources, tools and actions, with explicit limits on external commitments.
When must it return to a person?
Specify missing information, conflicting evidence and decisions outside its authority.
Does splitting the work across agents help?
Check whether the division improves quality or speed enough to justify additional handovers.
How will we judge the final result?
Review evidence, relevance and total effort against the original business purpose.
Full video transcript
The full video transcript is included below for reference.
I think the biggest change is that you’re starting to assign work rather than just ask questions.
With a chatbot, you tend to stay beside it. You give it an instruction, look at the answer, correct something and ask again. That can be very useful, but you’re still driving every step of the process.
With an agent, you can give it a more complete task. Here’s what I need done. Here’s the information you can use. Here’s what a good result looks like. Here are the things you need to come back to me about. Then it can work through more of that on its own.
The idea is about being able to bring that capacity into existence when you need it. You have a piece of work that needs research, analysis and a draft recommendation, and you can organise AI around those parts of the job.
But the quality of the brief becomes much more important. If you’re sitting beside someone, you can correct a misunderstanding after five minutes. If you send them away with a vague instruction, they can do a great deal of work in the wrong direction. The same problem applies here.
So “look into our competitors” is a weak brief. What decision are you trying to make? Which competitors matter? What sources are acceptable? Do you want a recommendation, or do you want options that you’ll decide between?
And you need to be clear about where the task ends. Preparing an email and sending an email are different responsibilities. Recommending a change and making that change are different responsibilities.
I think that’s the practical shift. You spend more thought at the beginning deciding what you’re asking for and more judgement at the end deciding whether the result is useful. You may do less of the work in between, but that makes the clarity of those two moments more important.
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.