What's Working,
What's Still a Pitch
agent use cases launched — 2 worked well, 3 needed heavy supervision, 2 were quietly shelved
That internal debrief is a more honest picture of enterprise AI agents in 2026 than anything in a vendor announcement or case study.
and expanded
marginal efficiency gain
after production errors
Code generation and review in software development
GitHub Copilot Workspace and similar tools take a task description, explore a codebase, write changes across multiple files, and run tests. The productivity gains are real — for teams treating output as a strong first draft, not finished code.
Customer support triage — not resolution, triage
An agent that reads an inbound request, works out what it's about, grabs relevant customer context, and hands it to the right person with a summary — that's table stakes in high-volume support now.
Research and synthesis across large document sets
Given a goal — "find all termination clauses across 200 contracts" — an agent works through it systematically. Law firms and consulting practices are the heaviest early adopters.
The failure mode is a sequence, not a single wrong answer
An inventory agent missed a supplier change and placed a large, incorrect order autonomously. The oversight that would have caught it had been removed to "let the agent prove itself." The lesson: human checkpoints aren't training wheels.
The bottom line
The two successful use cases had one thing in common: failure modes were visible quickly, the cost of an error was low, and the feedback loop was tight. That's not a coincidence — it's a design principle for every agentic deployment.