Assist or replace? Deciding role by role in an AI transformation.

The wrong question: which jobs will AI replace?

“Which jobs will AI replace?” makes for good headlines and bad decisions. A job is a bundle of tasks. A customer service agent answers routine questions, calms angry customers, spots a product fault in a pattern of complaints and explains a policy exception to a long-standing client. An AI agent can take some of those tasks over today. It should not take all of them.

Asking the question at the level of jobs produces two bad outcomes. Leaders either overreach, putting an agent on work it should only support, or they underreach, deciding a role “can’t be automated” and missing the half of it that can. The useful unit is the task.

The four-question test

For every task in a role, ask:

1. How rule-based is the work? Could a capable new hire do it from a written procedure and a set of examples? The more the answer is yes, the further an agent can go.
2. What does an error cost? A wrong draft that a person reads before it goes out costs a minute. A wrong payment, a wrong legal answer or a wrong decision about a person costs far more. Higher error cost means a person stays closer.
3. Who approves the output? If someone already reviews the work today, that review is the natural place for a human check on the agent. If nobody does, decide whether someone should.
4. How often does the work arrive? A task that arrives fifty times a day repays the effort of building and testing an agent. One that arrives twice a month usually does not.

The answers place each task on one of our three organisational levels: executive work (low volume, high judgement), operational work (departmental flows with a KPI) and task work (high volume, a clear right answer). The 3-tier AI workforce model explains the levels in more depth.

Three patterns: assist, co-work, take over the task

Pattern	What the agent does	Where a person stays	Fits when
Assist	Prepares, drafts, summarises	Decides and signs off everything	Judgement-heavy, low volume, high error cost
Co-work	Handles the routine share end to end, escalates the rest with context	Handles escalations, approves anything that sends or pays	Frequent work, mostly rule-based, some exceptions
Take over the task	Owns the task	Reviews exceptions and a sample	High volume, clear right answer, low error cost

The pattern is set per task and can change. Agents usually start one step more cautious than the test suggests and move up once their tests pass and the owner signs. Every step keeps a human in the loop somewhere; what changes is where.

Worked examples for ten roles

These are typical patterns from the test above, not results from specific clients. Your answers will differ with your volumes, error costs and rules.

Role	Task the agent takes over	Task where it co-works	Task where it only assists	What stays human
Customer service representative	Tagging and routing incoming tickets	Answering routine questions from the help content	Drafting replies to complaints	Angry or high-value customers, policy exceptions
Sales development representative	Researching accounts and contacts	Drafting and sequencing first contact	Preparing call notes	Live conversations, qualification calls
Accounts payable clerk	Extracting invoice data, matching to orders	Flagging mismatches with a proposed fix	—	Approving payments, supplier disputes
Bookkeeper	Categorising routine transactions	Month-end reconciliations	Explaining variances	Judgement on unusual entries
HR coordinator	Scheduling interviews, collecting documents	Answering routine policy questions	Drafting job descriptions	Any decision about a candidate or employee
Marketing coordinator	Resizing and formatting content	Drafting posts from a brief	Campaign analysis	Brand decisions, final sign-off
Paralegal	Extracting key terms from contracts	Comparing clauses with the standard	First read of a new contract	Legal advice, negotiation
Operations planner	Collecting status from systems	Proposing a schedule	Scenario analysis	Trade-offs between customers
Executive assistant	Inbox triage	Drafting routine replies	Meeting briefings	Anything sent in the executive’s name without review
Department head	—	—	Weekly briefing, KPI commentary	Decisions, people, priorities

Notice what the table does not contain: a single role that disappears. In each role a share of tasks moves to an agent, and the person’s job shifts towards approval, exceptions and the work that needs judgement.

What happens to the people: redeploy, upskill, approve

When agents take over tasks, the hours go somewhere. Leaders have three honest options, and most use all three:

• Approve. The person who did the task now checks the agent’s output and handles the exceptions. That is a real job, and it needs to be in their targets.
• Upskill. The same person takes on work that was waiting: proactive account reviews, better documentation, the projects nobody had time for.
• Redeploy. Over time, as volumes grow without new hires, people move to where the company needs them.

Whether headcount changes over time is a management decision, and it should be made openly. How to bring teams along through that change is covered in our guide to AI change management.

Next steps

Run the four questions for the department with the most hours in frequent, rule-based work. If more than half its tasks land on “co-work” or “take over”, it is a strong candidate to go first.

To see how assist-or-replace decisions fit into a company-wide programme with fixed prices per phase, read about our AI transformation services.

---

Key terms in this post: AI agent · human-in-the-loop · bounded autonomy · escalation rule · AI workforce