Economics
The comparison is not agent versus employee. It is agent-plus-human versus the current path.
Companies frame this as replacement. The useful math is which steps a person should stop doing, and whether the new mix is cheaper, faster, or better.
14–26%
productivity gains in structured support and software work. Gains are smaller where the job is judgment.
Stanford HAI, AI Index 20261
How I see it
Agent vs. human economics
Humans stay on judgment, authority, relationships, and rare exceptions. Agents take gathering, routine interpretation, system updates, and follow-up. The economic question is whether that split reduces cost per case and cycle time without raising error cost.
A loaded employee at $90K who spends 30% of the year on a workflow is $27K of surface area. An agent that removes half of that, plus review time, has to beat the leftover human cost and the run cost.
Quality can move either way. An agent that is consistent on the common path and escalates the rest can beat a tired queue. An agent that is silently wrong is more expensive than a slower person.
Do not use this analysis to announce a headcount program. Use it to decide whether the workflow is worth redesigning.
Common mistakes
What teams usually get wrong.
Replacement as the headline
That framing creates resistance and usually overstates what the first agent can do.
Comparing an agent to a fully loaded FTE 1:1
The first agent rarely owns 100% of a role. Compare on the case, not the job title.
Ignoring error cost
A cheaper wrong case is not a savings.
A useful diagnostic
Five questions before you fund the work.
What share of the role is this workflow, really?
If it is 15%, you are not replacing a person. You are changing a shift of their week.What is cost per case today?
That is the comparison unit.What is the error cost of a miss?
If it is high, keep more human time in the model.How much human time remains after the agent?
Review and exceptions are still people.Is the goal capacity, cost, speed, or quality?
Pick the primary one. The mix of agent and human follows it.
Economic model
Mixed pathagent run cost + remaining human time + error cost ≤ current human path
If this does not hold on conservative inputs, keep the current path or redesign the workflow without an agent.
Three credible paths
How far should you go?
Do not force one solution. Choose the path the economics, the risk, and the organization can support.
Give people time back
Use the agent to create capacity inside the same team.
Volume is growing and you do not want headcount to grow with it.
Savings are real only if the time is used or the overtime stops.
Reduce the cost of a queue
Target a shared service or back-office queue with a unit-cost baseline.
The work is already measured as items and cycle time.
Requires a clean done state.
Keep the work human
If judgment and error cost dominate, improve the process without an agent.
The economics do not clear.
This is often the right answer and feels like a non-event.
When this is the wrong next step
Do not fund an agent here.
- The conversation is only about reducing headcount.
- You cannot get a cost-per-case number.
- The role is mostly relationship work dressed up as a process.
A useful next step
Bring one workflow. Get guided into production.
We guide the implementation, go deep on the technical path, and stay hands-on through operations — or tell you when a simpler answer is better.
Discuss an AI opportunity