The next SaaS user will not be a human or an agent. It will be a human-agent team.

Two shifts are converging. AI is changing how work gets done, while software is moving from selling access to tools towards delivering outcomes. A practical workflow made that convergence clear.

Much of the productivity we get from agents comes from no longer having to operate every application ourselves. The agent can move between systems and complete the workflow. But this introduces another problem: the work can become a black box, while long-running, data-heavy tasks can be surprisingly brittle.

Intelligence was not the constraint

A monthly Stripe-to-Xero revenue-recognition process is a useful example.

The agent understood the accounting task. That was not the problem. The problem was execution: processing a large transaction dataset, generating scripts, calling multiple tools and maintaining state throughout a long-running job.

Eventually, it reached its execution limit before completing the work.

The issue was not intelligence. It was asking a probabilistic agent to repeatedly perform work better handled by deterministic software.

Put the repeatable work in software

The solution was a small web application built for that workflow.

It imports the transaction data, applies the approved revenue-recognition logic, maintains the deferred-revenue schedule and preserves the work from month to month. The agent uses the application to run the process. The accountable human uses it to inspect the calculations, review exceptions and approve the result.

The application has two users with different needs:

  • The agent needs structured inputs, deterministic logic and persistent state.
  • The human needs visibility, controls and an audit trail.

That is a useful clue to where SaaS is heading.

Design for the division of work

The shift is not from software for humans to software for agents. It is from software designed for a human operator to software designed for a human-agent team.

Agents will operate more of the workflow. Humans will define the rules, handle exceptions and remain accountable for the result. The product's value moves from providing access to a tool towards delivering a reliable, reviewable outcome.

This does not mean AI will not displace work. It already does. But the more useful question may be less, "Will a human or an agent do this job?" and more:

How should the work be divided between the agent, the software and the human accountable for the outcome?

Agent-native does not have to mean human-free.