Past the keynote demos, six workloads are absorbing the budget. None of them are agents writing novels. All of them are about shaving cost out of a cost-of-goods line.
Frontier model capability is no longer the binding constraint inside the average Fortune 500. The binding constraints are data plumbing, change management, and a procurement team that learned its templates in the SaaS era. The good news is that the workloads now in production are unglamorous and durable — they pay back inside a quarter and they don't require a CEO keynote to be defended at the next budget review.
We surveyed 140 enterprise programmes shipped in H1 2026. Six workload families account for 82% of recurring AI spend. Here is what is actually on the shop floor — what it does, who is buying it, and what it has replaced.
The agent reads the CRM, opens the knowledge base, drafts the reply, and — increasingly — issues the refund. NPS up, headcount flat, attrition down.
Pair-programming was the wedge. Autonomous repo-level agents are now closing low-risk tickets overnight, with a human reviewer signing off in the morning.
Vision models triage photos, language models draft the adjudication, a human approves anything above $5k. The float improves before the premium does.
The model talks to the technician, books the slot, raises the PO and closes the ticket. The dispatcher becomes an exception-handler.
Reply rates from AI-personalised outbound are now indistinguishable from human; cost per meeting booked is one tenth.
Boring, but the largest aggregate productivity win. RAG over Sharepoint / Confluence / Slack is finally consistent enough to trust.