Why "We Have AI" Stopped Being a Selling Point | Agentic AI vs. Chatbots

The short version: Facilities leaders are no longer impressed by the AI label. They are asking what the AI actually does, where it gets its information, and whether it executes or just suggests. What separates a chatbot from a true agentic AI runs all the way down to the architecture, and that gap shows up in every work order, every vendor dispatch, and every dollar spent on R&M. This post breaks down what separates the categories, what questions to ask, and what production-ready agentic AI looks like in a multi-site FM operation.

There was a moment at Connex FM this year that said everything about where the industry has landed on artificial intelligence.

At the 2025 executive roundtable, facilitators polled a room full of facilities leaders: “How many of you are personally using AI?” The hands were sparse, at maybe 15 to 20 percent; while the majority were likely AI-curious at best. 

Now, fast forward to 2026, and it’s an entirely different type of energy. According to Fexa’s Shannon Anderson and Beth Mooney, who were on the Connex floor all week, the operators showing up this year were not just browsing, but were very actively interrogating. 

We talked about this in full for the first episode of The State of FM, and the dynamic they described was impossible to ignore.

What Were Operators Actually Asking?

Shannon put it plainly: “They have now fast forwarded to this: [operators] were saying that they needed, and were skeptical, of somebody just saying they have AI. And they were coming to us with really intelligent questions around it.”

Operators spent the past year getting educated, and the pressure driving that education is not coming from IT teams or trade publications, but from boardrooms and investor calls. Companies watching their competitors capture OpEx efficiencies are being pushed hard to find the same. 

As Shannon put it: “Our investors are getting FOMO. We cannot run that same play. We need to find somewhere in our EBITDA where we can apply AI in a way that gets more out of our R&M spend.”

The urgency behind that shift is rooted in business performance – EBITDA targets, OpEx pressure, and competitive exposure – all with AI as the lever that leadership is pointing to. 

This shift has completely changed what facilities leaders need to know about the AI tools that vendors are selling them.

Has “AI” Lost Its Meaning as a Differentiator?

When every vendor at a trade show carries an AI message, the word stops doing any useful work. Beth Mooney, who leads product and customer marketing at Fexa, described the dynamic exactly: “People are really now understanding that there are different versions and different ways that AI can solve problems.”

That understanding matters because the gap between versions is enormous. Most tools calling themselves AI fall into one of two categories: chatbots or rules engines. Chatbots generate responses based on general training data, give you a suggested answer, and leave the execution to your team. Rules engines follow decision trees, look capable in a demo, but fall apart the moment a situation arrives that was not scripted in advance. Neither of them knows your business nor acts on your behalf. They both require a human to review, decide, and do something with the output.

There is a third category, increasingly called agentic AI, and the operational gap between it and the other two goes deeper than features. Its architecture differs entirely, and so do the outcomes.

What Happens When AI Does Not Actually Know Your Operation?

Here is a scenario that plays out across multi-site retail and restaurant operations every week: A store manager submits a work order: “It’s hot.” That is the entire submission, and oftentimes there’s no asset ID, no unit model, and no description of when it started or what they already tried.

A chatbot-style AI, even one embedded in a CMMS, will generate a generic response. It will pull troubleshooting guidance from its training data based on the words “hot” and “HVAC.” The guidance will be accurate in the abstract and useless in practice, because it does not know whether this location has a rooftop unit or a split system, what refrigerant type it uses, whether the warranty is still active, or which vendor your contract requires for this category of repair.

The work order gets routed as “HVAC General Maintenance.” The wrong vendor gets dispatched. Required fields are blank, so the escalation protocol never fires. The vendor arrives unprepared, the SLA clock runs, and because nobody called your warranty vendor, the repair that could have been covered at no cost has now become a full invoice.

Multiply this across hundreds of locations, and you have a systemic cost problem. 

That is what happens at scale when AI guesses instead of acts.

What Does Agentic AI Actually Do?

Agentic AI operates differently at every step of that same scenario.

When the manager types “it’s hot,” an agentic work order intake tool does not return a suggestion. It guides the intake in real time, asking the right clarifying questions based on the assets registered to that specific location. It applies your business rules as the work order is built: required fields, trade assignments, NTE limits, vendor routing logic, site-specific protocols. By the time the work order is submitted, the agent has identified the right vendor, the compliance fields are complete, and the escalation path is intact.

The agent does not just suggest the correct route, but executes the intake in a way that makes every downstream step possible. That is the functional definition of agentic: the AI acts on your behalf, within the guardrails you have configured, rather than waiting for a human to act on its recommendation.

The other distinction worth understanding is where the agent draws its knowledge. A bolt-on AI pulls from the open internet, or from a shared training corpus that knows facilities management in the general sense. An agent built into your CMMS pulls from your data: your equipment records, your vendor contracts, your configured SOPs, your trade routing rules, your permission structure. It does not guess at your NTE threshold because it already knows it. It does not route to the wrong vendor because it has already been told which vendor handles which trade at which location.

This distinction is what Fexa means when describing FexaAI as workflow-intelligent. The agents are not smart because they were trained on more data. They are precise because they operate inside the workflows your team already configured. Your business rules are the intelligence. The agent is the mechanism that applies them at scale.

What Should You Ask Any AI Vendor Before You Buy?

The operators Shannon described at Connex FM were arriving at the booth with sharper questions than a year ago. 

Some of those questions are worth having on hand whenever an AI evaluation is underway.

1. Can the AI handle a problem it has never seen before, or only pre-built scenarios? If a vendor’s AI only functions within a defined decision tree, it is a rules engine with a modern interface. Push on what happens when a work order arrives outside the script.
2. Does the AI live inside the workflow your team already uses, or is it a separate tool they have to open separately? Adoption is the single most reliable proxy for whether AI is actually working. Tools that sit outside the daily workflow get used inconsistently. The Work Order Agent at Fexa reached 70% to 80% organic adoption without a mandate or a training program. Teams chose to use it because it made their existing workflow faster, not because they were required to add a new step.
3. Does the AI pull only from sources you have approved? If an agent is drawing from the open internet to answer a question about your equipment or your vendors, the output cannot be traced or trusted. Every answer should link back to a source your team can verify.
4. Is your data kept completely separate from other customers? If a platform uses your operational data to improve a shared model, the improvements your operation generates could benefit competitors using the same system. Get the data isolation policy in writing.

Where Does FexaAI Stand Today?

Fexa launched the Work Order Agent into general availability earlier this year, making it the first production-ready agentic AI tool purpose-built for multi-site facilities management. The results across the live customer base include approximately 300 hours recovered per FM team monthly, work orders completed five days faster, and roughly 30 percent fewer triage escalations reaching a human queue.

The Answers Agent, now rolling out to select customers, extends the same architecture to operational queries. Facilities managers, store managers, and regional directors can ask questions in plain language and receive responses drawn directly from Fexa data: work order status, vendor ETAs, open invoice details, assignment information. The agent does not synthesize or interpret. It retrieves. Every answer is scoped to the requesting user’s permissions, tied to a source record, and accurate because it reflects what is actually in the system.

Together, the Work Order Agent and the Answers Agent represent the first multi-agent platform in facilities management where multiple agents share the same operational context. The same asset data, permission structure, vendor records, and workflow configurations that make the Work Order Agent precise also power the Answers Agent’s responses. Two agents, one shared intelligence layer. The same context that makes one precise makes the other accurate.

What Does the Bar Look Like Now?

Beth Mooney put it well: “A year ago I felt like I was in the ocean with binoculars looking back to the beach. Now people are jumping in.”

The question is not whether AI belongs in facilities management. That debate is over. The question now is which AI is actually doing what it claims, which AI understands the difference between an R-22 refrigerant event and a general HVAC complaint, and which AI is built to act versus built to suggest.

Operators are asking harder questions. The answers to those questions should come from proof, not from positioning language. The time is now, because as Beth Mooney puts it: “there’s no longer an excuse to not understand what’s in your facility.”

The tools exist. The data is there. The only remaining question is whether the AI you are evaluating knows how to use it.