Which processes should we give to AI agents first?

High-volume work that still needs some interpretation, with a named owner, a current documented process, decisions you can describe with clear limits, and actions you can stop or reverse. Start there, deliver it well, and use what you learn to choose the next one.

What makes a process a poor fit for an AI agent today?

No clear owner, documentation that no longer matches the work, decisions nobody can describe, or consequences that are hard to detect or reverse. Each of these is fixable. Until it is fixed, a human-led, technology-assisted or workflow-orchestrated mode may serve the process better.

Does every process need to move towards agentic AI?

No. There are six operating modes, from human-led to agentic AI with governed autonomy, and nothing requires a process to progress through them. A well-performing human-led process can be the right answer. One process can also combine several modes.

Can part of a process use an AI agent?

Yes, and that is often where agents start. A workflow can coordinate a case, automation can retrieve records, an agent can investigate and a person can approve the outcome. Match each step to the mode that fits it.

Is This Process Suitable for AI Agents?

Q: Is this process suitable for AI agents?

Check three things. Does someone own it today, by name? Does the documented version still match how the work runs? Do you know which steps carry a decision, who is accountable for each and within what limit? Three yeses make a strong first candidate. A no tells you what to put right first.

A process suits an AI agent when someone owns it by name, the documented version still matches how the work runs, and every decision step has an accountable person and a clear limit. Three yeses make a strong first candidate. A no is useful too: it tells you what to put right first, and which operating mode fits the process today.

Start from the process you have

A pilot team needs a process to hand to an agent. The newest version they can find is a Visio file dated 2023, and the person who drew it has moved to another role.

This is normal. Processes change faster than their documentation, and the people who know the current version are busy running it. The knowledge exists. It just needs pulling together before an agent can work from it.

So before you ask whether the agent is good enough, ask about the process.

Three checks before any agent pilot

1. Does someone own it today, by name? An agent needs someone who can say what the process is for, approve changes to it and answer for its outcomes. A team or a shared mailbox cannot do that. A named owner can.

2. Does the documented version still match how the work runs? An agent follows what it is given. If the procedure describes last year’s process, the agent will run last year’s process, confidently. Check the date, the source and whether the people who do the work recognise it. Where they do not, establish how the work runs today before anything else.

3. Do you know which steps carry a decision, and who is accountable for each? Agents are most useful where work needs interpretation. That makes the decision points the part to get right. For each one, name the accountable person and the limit within which a decision can be made without them.

Three yeses and you have a strong first candidate. One or two noes and you have a short, specific action list, usually quicker to work through than rebuilding a pilot that started on the wrong version.

Readiness and suitability are different decisions

Your organisation can be well prepared for agents and a given process can still be a poor fit today. The reverse is also true: one well-owned, well-documented process can be ready before the wider organisation is.

Keep the two decisions apart. Organisation readiness tells you how fast you can scale. Process suitability tells you where to start. We cover the first in where agentic AI will pay off first.

Match each process to the right operating mode

Suitability is not a yes or no for agents. It is a choice between ways of running the work. On our Process Operating Fit page we describe six operating modes:

Human-led
Technology-assisted
Workflow-orchestrated
Task-automated
Agentic AI with human supervision
Agentic AI with governed autonomy

These are options to evaluate and combine. There is no requirement to progress through them. A single process often uses several: a workflow coordinates a complaint, automation retrieves legacy records, an agent investigates the case and a person approves compensation.

What decides the right mode for each step is practical: the outcome the process must deliver, its volume, how much it varies, the full cost of running it, what happens when something goes wrong, the reliability of its information and systems, the people who will operate it, and how it recovers when something is unavailable.

Where the evidence comes from

You will already hold much of what these checks need: process models, procedures, control evidence and the knowledge of the people who run the work. We help you bring it together and see how far you can trust it, including who owns each part, how current it is and where two sources disagree.

Where the documentation runs out, discovery with the people who do the work and process mining on system event logs show how it runs today. That gives you an agent built on the current process rather than a remembered one.

The problems behind these checks are set out in more depth on finding which processes are safe to hand to AI agents.

What to do next

Pick one process you are considering for an agent. Run the three checks with its owner and two people who do the work. Write down where the answers differ, because that is where the action is.

Once a process passes, the next question is how much the agent should be allowed to do on its own. That is a decision you make step by step, and we set it out in how to set safe autonomy for AI agents.

How to Tell Whether a Process Suits an AI Agent