is now on nearly every CMMS and asset-intelligence slide deck sold into HTM. Most of it is rules-based automation with a new label. This guide is the checklist that separates the tools worth your budget from the ones counting on you not asking the hard questions.
What you get from this guide: a printable checklist to take into any vendor demo, organized around the five questions that decide whether an AI-enabled HTM tool is worth buying. Not a product ranking. A framework you can use against any vendor.
No vendor paid for inclusion in this guide. We don’t rank tools here. We give you the questions that make vendors uncomfortable, because those are the questions that protect your budget and your next survey.
- Verify the AI before you pay for it. Ask what the model was trained on and what it does with no data. If the vendor can't answer, you're buying automation at an AI price.
- Integration is the whole game. An AI layer that can't read live data from your CMMS, EHR, and security platform is a silo with a dashboard.
- Compliance owns the output. An AI-suggested PM interval has to be defensible under TJC and your AEM program. A black box is not survey-ready.
- Read the data-ownership clause first. If you can't export your records cleanly on exit, you're locked in regardless of how good the tool looks today.
- Price the whole thing. First-year cost includes migration, integration, and training, often several times the license fee.
Start with the problem, not the AI
The most expensive mistake in this category is buying intelligence to sit on top of a broken process. If your PM completion data is unreliable, your inventory is inaccurate, or your work-order discipline is loose, an AI tool will not fix any of that. It will optimize on bad inputs and give you confident, wrong answers faster.
Automation scales whatever you already do. If the underlying process is sound, AI can extend it. If it isn’t, you’re paying to scale the dysfunction.
So the first question isn’t “which AI tool?” It’s “what specific, measurable problem am I trying to solve, and is my data clean enough for a tool to solve it?” Write that down before the first demo. If you can’t name the problem in one sentence, you’re not ready to buy.
Question 1: Is it actually AI, or automation with a label?
This is the question vendors least want you to ask precisely. There’s nothing wrong with rules-based automation. Constraint-based PM scheduling is genuinely useful. The problem is paying an AI premium for it.
The distinction matters because the two technologies behave differently and fail differently. Machine learning is trained on examples and degrades gracefully on inputs it hasn’t seen. Rules-based automation runs logic a programmer wrote and breaks, or returns nothing, when reality falls outside those rules.
The data problem specific to medical equipment
Predictive maintenance works in industrial settings because the data supports it: hundreds of identical machines running continuously generate thousands of failure events a year. Biomedical equipment doesn’t generate data at that scale.
A fleet of 200 ventilators might produce a handful of documented failures over five years. That’s far below what most supervised models need to predict anything reliably.
That’s not a reason to dismiss AI in HTM. It’s a reason to be specific about which claims are credible. ML for image-based asset identification or natural-language work-order classification is plausible today. ML for failure prediction on low-failure-rate devices usually is not, unless the vendor has an unusually large longitudinal dataset.
Ask these in the demo
- What dataset was this model trained on, and how large is it?
- What’s the minimum amount of our data the model needs before its output is trustworthy?
- What does the tool do with a device type it has never seen?
- Can you show a false-positive / false-negative rate on healthcare equipment, not general industrial data?
- Is this prediction or is it a rule someone configured?
If the answers are vague, the AI claim is marketing. We flag this distinction on vendor profiles in the HTMwire directory, whether a tool uses genuine machine learning or relabeled automation, so you can check before you ever sit through the demo.
Question 2: Does it integrate with the systems you actually run?
An AI tool is only as good as the data it can reach. If it can’t pull live data from your CMMS, your EHR, your security platform, and your test equipment, it’s a separate system your team has to feed by hand. Double data entry kills adoption faster than any missing feature.
“Integration” is a word vendors use loosely. Make them define it. A one-time CSV import is not an integration. A read-only API that updates nightly is different from a live bidirectional connection.
Ask these in the demo
- Do you have a live connection to our CMMS, or do we import and export files?
- Show me a real customer with a live Epic or Oracle Health (Cerner) connection.
- Can you ingest CVE and device-vulnerability data from our security platform (Claroty Medigate, Ordr, Cylera, Asimily)?
- Does test-equipment data (for example Fluke) flow in automatically, or is it manual?
- When your tool updates a record, does it write back to our CMMS, or only display it in your dashboard?
A tool that can’t write back to your system of record means your CMMS and your AI tool will drift out of sync. Your team will stop trusting both.
Question 3: Does the AI’s output survive a survey?
Here’s the question most demos skip entirely: if this tool recommends extending a PM interval or deprioritizing a device, can you defend that decision to a surveyor?
The Joint Commission’s EC.02.04.01 requires a documented maintenance strategy for the equipment on your inventory. If you run an Alternative Equipment Management (AEM) program, ANSI/AAMI EQ103:2024 requires that any deviation from manufacturer-recommended maintenance rest on a documented, evidence-based rationale. An AI recommendation is not a rationale on its own.
That means a black-box model that can’t explain why it suggested a change is a compliance liability, not an asset. You need to see the basis for the recommendation, capture it, and be able to produce it on demand. CMS Conditions of Participation make the same documentation expectations legally binding for Medicare-certified facilities (CMS).
Ask these in the demo
- When the tool recommends a maintenance change, does it show the evidence behind it?
- Can we export that rationale into our AEM documentation per EQ103?
- How long does it take to produce EC.02.04.01 documentation from this system?
- Does an AI suggestion auto-apply, or does a human approve and document it first?
The right posture: AI suggests, a qualified human decides and documents. Any tool that quietly changes intervals without a documented, defensible trail is creating survey risk on your behalf.
Question 4: Who owns the data, and can you leave?
Data portability is the single biggest source of lock-in in this category, and it’s worst with AI tools because your operational history is often what makes the model useful. Vendors know that. Read the contract before the demo wins you over.
There are three traps. First, your exit data comes back only in a proprietary format that’s expensive to migrate. Second, the vendor licenses your own operational data back to you rather than granting ownership. Third, specific to AI: your device records are used to train a shared model across the vendor’s customers, with no opt-out.
Ask these in the demo
- If we terminate, what format is our data returned in, and how long does it take?
- Do we own our work-order, PM, and device records outright, or are they licensed?
- Does our data train your shared AI model? Can we opt out?
- Is there a fee to export our full history?
Get the answers in writing. A vendor that owns your data, or your only clean copy of it, owns your renewal negotiation.
Question 5: What does it actually cost?
The license fee is the smallest part of the number. For AI-enabled tools, the real cost includes data migration, integration build-out, training, and frequently a managed-services or “AI tier” upcharge layered on the base platform.
Data cleansing is usually the hidden line item. ECRI has long flagged data cleansing and normalization as the make-or-break factor in getting value from a CMMS (ECRI); an AI tool raises the stakes, because a model trained on dirty data produces confident garbage. Budget weeks of cleanup, not days.
Ask these in the demo
- What’s the all-in first-year cost including migration, integration, and training?
- Is the AI capability included, or is it a separate tier or per-device upcharge?
- What’s the typical year-over-year increase at renewal?
- Who does the data cleansing (your team or ours), and what does that cost?
If a vendor will only quote the license fee, you don’t have a price. You have a starting point for a much larger number.
Run a real pilot before you commit
Demos are staged on clean data. Your environment is not clean. Before signing, run a proof-of-concept on your actual inventory, with your actual records, not the vendor’s polished sample set.
Structure it so it can fail honestly:
- Define success first. Pick two or three measurable outcomes (PM completion rate, mean time to repair, documentation hours saved).
- Set a baseline. Pull your current numbers so you can measure against something real.
- Use your data. Require the vendor to run on a representative slice of your real inventory, including the messy device categories.
- Fix the window. Run it long enough to see real results, with a clear end date and decision point.
- Name the decision-makers. Decide in advance who signs off and what threshold they need to see.
A pilot that can’t move your chosen metrics against a real baseline has told you something valuable before you spent the budget.
How to verify a vendor’s claims
Treat every vendor-reported number as a hypothesis, not a fact. The HTMwire standard: who published this, and what were they selling?
- Ask for a reference customer with a department like yours, and talk to them without the vendor on the call.
- Separate independently verifiable claims from vendor-reported ones, and weight them differently.
- Cross-check security and recall claims against FDA’s own guidance and postmarket expectations (FDA Medical Device Cybersecurity).
- Check whether AAMI, ECRI, or a Tier-1 body has actually weighed in on the capability, or whether the only source is the vendor’s own white paper (AAMI).
The bottom line
An AI-enabled HTM tool is worth buying when it solves a problem you’ve already defined, runs on data clean enough to trust, integrates with the systems you actually use, produces output you can defend in a survey, leaves your data portable, and earns its full cost, not just its license fee.
Settle those five questions in the demo, not after the contract. Bring the checklist. Make the vendor answer live. The tools worth your money will welcome the questions. The ones counting on you not to ask are the ones to walk away from.
Sources
Tier 1: Institutional
Related HTMwire guides
This guide is maintained by the HTMwire editorial team. We review and update it quarterly. Last reviewed: June 2026.
No vendor paid for inclusion or favorable coverage. Read our evaluation methodology for details.
