#Automate2026 opens Monday at McCormick Place. Fifty thousand people, a thousand exhibitors, and a show floor that takes two days to walk. Somewhere in the middle of it I'm giving a talk, and since most of you will be deciding between sessions and cold brew, I'll tell you exactly what's in it so you can make an honest call.
The title is "From Chatbots to Agents: How AI Is Redefining Sales and Service in Industrial Automation." That sounds like every other AI session on the schedule. It isn't, and here's the specific reason.
Most AI talks at a show like this either sell you a product or wave their hands about transformation. This one opens with a chatbot getting a real question wrong, in a way that costs somebody between ten and seventy-five thousand dollars.
A customer needs to replace a discontinued NOSHOK 800 Series pressure transmitter. They specify NPN switch output. The bot confidently quotes the PTI15 equivalent with PNP output. PNP and NPN are not interchangeable. The sensor powers up, reads fine on its own display, and never talks to the PLC. The customer escalates, the distributor eats the return, the chatbot gets disabled, and everyone goes back to phone and email. The bot didn't malfunction. It did exactly what it was built to do. That's the problem.
The talk is built around a question I now hear in almost every customer conversation: when are the hallucinations going to be fixed? My answer is that they aren't, because they were never a bug. When an LLM picks a Siemens replacement CPU, it isn't checking your customer's rack. It's picking the most statistically associated part number from its training data. I'll show you four real Siemens CPUs, all valid products, where the model assigns the highest probability to the one that doesn't fit. A better model picks the wrong one more confidently. More training data won't help either, because your customer's installed base was never in the training set and never will be.
Then the part I actually care about. Coding agents largely solved this for software, and the reason is worth understanding because it tells you what industrial AI is missing. A coding agent can't lie to the compiler. The code runs or it doesn't, the tests pass or they fail, and the agent loops on that feedback until it works. Industrial product recommendations have no compiler. You can't unit-test a part crossover. The answer ships, and you find out three days later when it arrives at the dock wrong.
So the middle of the talk is about how you build a feedback loop when there's nothing to compile against. I walk through the actual evolution the field went through to get here: prompt engineering in 2022, context engineering in 2024, and the harness era we're in now. I'll explain why RAG alone tops out around 75 to 85 percent on industrial product queries, using a Rittal cabinet example where five real SKUs differ by a single digit and look identical to a vector search. I'll show the summarization trap, where an agent compresses a long conversation, silently drops the detail that the customer is running an existing S7-300, and then cheerfully reports the job done with a configuration that won't physically fit the rack. None of these are hypothetical. The screenshots are real, pulled from generic AI tools over the last twelve months.
The constructive half is the harness: the layer between the model and reality that decides what the model can touch, grounds every claim against structured data, and closes the loop. For a coding agent that's the compiler and the test suite. For an industrial agent it's your knowledge graph, your live catalog, and your customer's actual system. Same pattern, different ground truth. I'll show what we built, why every answer it produces carries a citation and a confidence tier, and how MCP makes that harness portable across web chat, voice, your ERP, and your sales team's Slack without rebuilding it four times.
And because none of this matters without proof, I'll show what it looks like live. You'll leave with three things to take back to your team, and they're not about us. Stop asking which model a vendor uses, because the model commoditizes. Ask what grounded data the answer is checked against and what closes the loop. Demand a grounding layer, where any vendor should be able to tell you where an answer came from, what rule produced it, and how confident it is. And build for portability, because lock-in is over and the agent that works in your portal today should work in your CRM tomorrow.
If you sell, spec, or support industrial products and you've watched an AI demo look brilliant and then disappoint in production, this is the session that will tell you why, and what to ask for instead.
Tuesday, June 23, 10:15 to 11:00am, room S404D. McCormick Place, Chicago. Come find the session, and if the timing doesn't work, come find me on the floor. I'll spend most of my time at one of our partner's booth: The Imaging Source, booth #3491. I'd rather walk you through a real wrong-part disaster than hand you one more slide about transformation.