From chat to delegated work: how to measure AI

What changed in October 2026

OpenAI reported on 14 October that Codex and ChatGPT Work account for more than half of output tokens in its sample of EU ChatGPT Enterprise and Business customers. Weekly active users of these tools among EU small and medium businesses grew more than eightfold since May. This is one provider’s data rather than a measure of the whole market, but the direction is clear: organisations are moving from AI advice towards delegated, multi-step work.

Microsoft describes the same shift at business-process level. Agents can retrieve CRM context, draft a response, create a case, operate outside business hours and hand a difficult situation to a person. This is no longer a standalone chat: it is a governed route across data, rules, actions and accountable employees.

Why licences and message counts do not prove value

Active users measure reach, not outcomes. A team can generate more text without shortening a process, improving decision quality or reducing a queue. A high automation rate is equally misleading if errors require expensive rework.

Measurement must attach to one defined process: receiving a request, checking a document, qualifying an enquiry, preparing a report or updating CRM. Record the baseline duration, quality and human cost first, then compare the same types of cases.

Five metrics for a delegated AI workflow

A minimum management dashboard should show operational outcomes and risk boundaries rather than model activity.

• Completion rate: the share of cases the agent brings to the defined outcome without repeated manual work
• Outcome quality: the share of answers or actions that pass the control criteria on the first review
• Human handoff rate: which cases the agent escalates and whether it does so before a risky action
• Cycle time: elapsed time from request to outcome, including approvals and corrections
• Cost per acceptable outcome: model, infrastructure, human review and rework cost per completed case

Set autonomy boundaries before launch

Divide work into three levels. The agent may complete low-risk actions automatically, prepare consequential actions for approval, and immediately hand exceptions, uncertainty and sensitive cases to a person. These rules must exist in access controls and workflow routing, not only in a model instruction.

For a business, that may mean handing off complaints, contract changes or unusual discounts. For a hromada or public authority, it may mean requests containing personal data, questions outside an approved knowledge base or decisions with legal effect. The audit trail should retain sources, actions, approvals, outcomes and the reason for escalation.

A practical 30-day plan

In week one, choose one repeatable process and record its baseline. In week two, document data sources, an accountable owner, quality criteria and actions that always require approval. In week three, run a constrained pilot on historical or de-identified cases and test common errors and escalation paths.

In week four, expose the pilot to a small stream of live requests under human supervision. Decide whether to scale using the five metrics rather than the impression created by a demo. If results are stable, add data sources or actions one at a time and retain a benchmark set for regression testing.

Practical takeaway

AI maturity is not the number of chats; it is the organisation’s ability to delegate a complete unit of work safely. Start with one process, a shared outcome definition and five metrics. Data, permissions, escalation and auditability belong in the solution from day one.