Know When the Workflow Should Stop Being Agentic

The situation

Once a skill library exists, it becomes tempting to connect everything: research the account, decide the offer, draft the message, send it, update the CRM, schedule the meeting, and follow up automatically. Technically, the pieces may all be possible. That does not mean the entire chain should run without intervention.

The naive solution

If the agent can perform a step reliably most of the time, automate the next step as well until the whole workflow is autonomous.

The hidden constraint

Risk compounds across a chain. A small error in research can become a false claim in a draft, then a customer-facing message, then a commercial commitment. Some mistakes are cheap to reverse and others are not. The correct autonomy boundary depends on consequence, not on whether the model is capable of generating the next action.

The decision

Define autonomy deliberately per transition. Low-risk reversible actions may run automatically. High-impact actions can require approval, additional verification, deterministic policy checks, or a human decision.

The implementation

I like to label skill outputs as analysis, draft, recommendation, approved action, or executed action. That makes the transition visible. A skill may prepare a WhatsApp response but return it as a draft. A separate approval event then allows a send capability to act. Financial, legal, security-sensitive, or reputation-sensitive operations can use tighter rules than internal summarisation.

The tradeoff

Approval gates reduce automation and can slow throughput. That is not automatically a product failure. In many business workflows, the valuable AI capability is removing 80% of preparation work while leaving the final 20% decision with the accountable person.

The evidence I would look for

I would expect the system to record who or what authorised consequential actions, to stop cleanly when confidence or evidence is insufficient, and to avoid framing every escalation as an agent failure. Sometimes escalation is the designed successful outcome.

The reusable lesson

The goal is appropriate autonomy, not maximum autonomy. A good skill knows not only how to act but when its responsibility ends.

The consultant's questions

• Which transitions in this workflow create financial, legal, security, or reputational consequences?
• Where is a human currently adding judgement rather than merely copying information between systems?
• What is the cheapest point at which a bad assumption can be caught?
• Does the system treat escalation as a supported outcome or as an exception?