A team of AI agents or one more employee?
A team of agents absorbs volume, repetition and opening hours. It carries no judgement, no liability and no relationship. These are two answers to two different problems.
The comparison comes up in almost every committee we meet, and it is badly framed. An agent is not a cheaper colleague: it does not do the same job. This page sets out what each one does, and what neither will ever do in place of the other.
The six differences, from the most structural to the most circumstantial
Ranked by how much they change the decision. The first two are often enough to settle it.
-
Liability cannot be delegated to an agent
An employee carries professional liability, and so does the organisation. An agent does not: its actions remain those of the organisation operating it, and human supervision is what makes that operation tenable.
This is the difference most often forgotten in committee, and it is the one that determines which use cases are possible at all.
-
Judgement on new cases
An agent handles what resembles what it has seen and what it was configured to do. Faced with an unprecedented situation it does not decide: it hands over, or it gets it wrong if it was never taught to hand over.
Hence the rule we apply: every agent has a named human fallback, and a list of what it never does alone.
-
Volume, repetition and opening hours
This is where the agent is clearly superior: it handles the midday peak like the 3 a.m. lull, with no queue, no absence and no drop in quality from fatigue.
Trade-off: that regularity also applies to its mistakes. A configuration fault repeats identically across the whole volume, where a person would have flagged the anomaly.
-
Time to service
A hire requires a search, a notice period and a ramp-up. A team of agents requires scoping, connection and user training — then it is operational on its perimeter.
Trade-off: the agent is operational on ITS perimeter only. A colleague absorbs neighbouring tasks without anyone having to plan for them.
-
Relationships, internal and external
Negotiating, calming an unhappy customer, understanding what someone did not say, representing the organisation: this is an employee's own ground, and we do not claim to replace them there.
An agent can prepare those exchanges — retrieve the history, draft a first reply — and that is often where it frees up the most time.
-
Cost structure
A post carries pay, employer contributions, equipment, training and a turnover risk. A team of agents carries a subscription, a commissioning and token consumption rebilled at cost.
We publish no costed comparison between the two: it would depend on a collective agreement, a grade and an activity volume we do not know. Agent prices are public; the cost of a post is known to you alone.
What each one does, and does not do
| Type of task | Team of AI agents | One more employee |
|---|---|---|
| High-volume repetitive processing | The agent's own ground | Possible, but costly and demotivating |
| Availability outside working hours | Continuous | Limited by employment law and organisation |
| New or ambiguous case | Hands over to a person | Decides and takes responsibility |
| Professional liability | Carried by the organisation | Carried by the person and the organisation |
| Difficult customer relationship | Prepares, does not conduct | Conducts |
| Unplanned neighbouring tasks | Outside the configured perimeter | Absorbed naturally |
| Sudden surge in load | Immediate | Hiring or overtime |
Read line by line, the comparison stops being a budget trade-off: the two columns only overlap on part of the table. In most organisations we work with, the team of agents does not replace a post — it gives existing posts back the time they spent on the top rows.
What is included and what is rebilled at cost
| Item | Included in the published price | Rebilled at cost |
|---|---|---|
| Scoping, connection, training | Yes | No |
| Human supervision and monitoring | Yes | No |
| Agent updates | Yes | No |
| Model token consumption | No | Yes, at observed cost |
| Work outside the perimeter | No | By quotation, never without written agreement |
The only variable item is token consumption, rebilled with no margin. It is the first thing investment committees ask about, and vagueness on this point is what causes the most AI projects to fail.
When you should hire, and not install an agent
These cases come up often. In each of them we advise against the agent — saying so here saves everyone a pointless audit.
- The work is essentially relational: negotiation, management, representation, supporting people.
- Volume is low but every case is singular: there is no regularity on which to configure an agent.
- The task carries named professional liability, particularly in a regulated profession.
- Your data is not usable as it stands: an agent configured on disorderly information will reproduce that disorder, at scale.