Managed Agents, Not Assistants

The pitch around enterprise artificial intelligence has changed in the last year: vendors are no longer selling a chat box that helps an employee write an email, but a managed agent that can open systems, follow rules, and finish work with logs attached. OpenAI now describes Frontier as a platform for “deploying secure, production-ready AI agents” tied to systems of record, while Salesforce says Agentforce is for “digital labor,” not just assistance. (openai.com) (salesforce.com) That shift starts with what companies learned from the first wave of copilots in 2023 and 2024. A copilot could draft text inside one screen, but a finance team still needed a person to click through approvals in an enterprise resource planning system, update a customer record, and hand off exceptions to a manager. (openai.com) (microsoft.com) A managed agent is closer to a junior worker with a badge than to an autocomplete bar. Microsoft says autonomous agents in Copilot Studio can run in the background from triggers, make decisions, and execute tasks without waiting for a user prompt, but only inside instructions and guardrails an organization defines. (microsoft.com) Once software can act across systems, the buying criteria change fast. Microsoft’s governance documents put data policies, authentication, connector controls, and maker audit logs at the center, which is a sign that information technology teams are judging these tools less like chat apps and more like enterprise software that can move money, records, or customer messages. (microsoft.com 1) (microsoft.com 2) That is why “reliability” now means something more concrete than a model giving a good answer in a demo. OpenAI’s agent tooling emphasizes observability, tracing, evaluations, and versioning, because a company needs to inspect each step the agent took the way it would inspect a failed payment run or a broken application programming interface call. (openai.com 1) (openai.com 2) The audit trail has become part of the product, not an extra feature. Salesforce engineering says its Agentforce audit systems support 500 enterprise customers and process 20 million model interactions each month, which shows how much enterprise demand has moved from “can it answer?” to “can we reconstruct exactly what it did?” (salesforce.com) Escalation paths are getting the same treatment. Salesforce’s audit trail tools are built so organizations can review masked data, toxicity scores, and usage records, and Microsoft’s governance guidance focuses on runtime protection status and administrator visibility, which means the real product promise is supervised automation with a human fallback when the agent hits an edge case. (salesforce.com) (microsoft.com) The other reason this framing is spreading is that agents can swallow older point solutions. If one agent can read a ticket, pull account history, update a customer relationship management record, trigger a workflow through MuleSoft, and draft the reply, a buyer may need fewer separate tools for search, routing, scripting, and handoff. (salesforce.com 1) (salesforce.com 2) But every system an agent touches adds another governance problem. Microsoft’s cloud adoption guidance for agents focuses on boundaries around what data agents can use, where they can operate, and how long they can retain information, because a tool that spans five systems also spans five sets of permissions, retention rules, and compliance risks. (microsoft.com) That is why the language has moved from assistants to managed agents. The winners in enterprise artificial intelligence are increasingly the vendors that can show permissions, logs, tests, and escalation controls around cross-system work, not just the vendors with the smoothest demo of a chatbot typing fast in one window. (openai.com) (microsoft.com)