IBM and ServiceNow did not launch better chatbots this week. They launched the boring parts that make AI agents useful: identity, tool access, workflow context, logging, handoff paths, and teams that sit close enough to the business to ship something real.
That is the right signal to watch.
IBM made Enterprise Advantage generally available on AWS, built around Amazon Bedrock, Bedrock AgentCore, a Context Studio with a knowledge graph, an MCP gateway for secure tool access, lifecycle management, observability, and CloudWatch/AIOps links. ServiceNow and Accenture announced forward-deployed engineering teams that build agent workflows inside customer environments on the ServiceNow AI Platform, with more than 300 pre-built agent skills and AI Control Tower sitting over the top.
Strip out the vendor polish and the message is plain: agents are not a model problem anymore. They are an operations problem.
We see the same thing in smaller systems. A prototype agent can read a PDF, draft an email, or look up an order status in a day. That feels impressive for about five minutes. Then the real questions arrive. Which customer records can it see? Which actions can it take without approval? Where does it log the reason for a decision? What happens when the API fails halfway through a workflow? Who owns the change when the finance process moves from spreadsheet to agent-assisted queue?
Those questions decide whether the project survives contact with the business.
IBM's numbers underline the gap. Its January 2026 study of 2,007 senior executives found that 79% expect AI to add significant revenue by 2030, but only 24% have a clear view of where that revenue will come from. Executives also expect AI investment to rise about 150% by 2030, while 68% worry their AI work will fail because it is not tied into core business activity.
That tracks. Most failed AI pilots are not killed by weak models. They die because nobody connected the agent to the messy part of the business with enough care.
Mid-market companies should copy the architecture, not the budget. Pick one workflow where the current process is slow, repetitive, and measurable. Give the agent narrow permissions. Put every tool call behind identity and audit logs. Store the context it needs in a shape the system can retrieve reliably. Add a human escalation path before the first production user touches it. Measure cycle time, error rate, and handoffs, not demo applause.
The MCP gateway matters here because it gives teams a cleaner way to expose tools without handing an agent the keys to everything. So does observability. If an agent changes a record in Xero, updates a CRM field, or drafts a support response, the business needs to know what happened, why it happened, and who can unwind it.
The next practical edge will come from the companies that choose the right two or three workflows, build the control layer first, and let autonomy grow only where the logs prove it should.
Leave a Reply