Buyer's guide · July 2026
Nearly every CMMS and equipment platform now advertises AI. Some of it is working capability you can verify in a demo; some of it is a label on features that existed for years. This guide is the test we'd want applied to any vendor — including us.
Real answers from real records, read-only by design, and nothing runs until you approve it.
Real product screens. Figures shown are illustrative demo data.
01 · Everybody's software has AI now. Here is what ours will not do. It will not touch your data. It will not invent an answer. And it will not act without you.
02 · Ask it a question and it reads your actual records, then answers with the machine, the number, and where the number came from.
03 · It is read-only by design. It can look. It can never touch. Your data cannot be edited by a model, because the model was never given a pen.
04 · When it drafts a plan, the steps, the parts, the warranty check, nothing runs until a human ticks the box. The AI does the digging. You stay the boss.
05 · That is what AI in maintenance should look like. Intelligence you can check. Torgix.
Read this first
This guide is published by Torgix, which sells software in this category — you should weigh it accordingly. We have tried to earn your trust the only way a vendor can: every general claim here is stated so you can test it against any vendor, every claim about Torgix is verifiable in a free trial on your own data, and the section on what AI cannot reliably do applies to us too.
As of mid-2026, most major vendors in equipment and maintenance software — including Tractian, Clue, Fiix, MaintainX, Tenna, and Torgix — market AI features. The point of this guide is not that one vendor's AI is real and the rest are not; it is that the same four questions separate substance from labeling wherever you point them.
The test
Ask these in every demo. They take five minutes and they are hard to answer with a slide.
Ask the AI a question about the demo fleet and then ask the vendor to show where the answer came from. A grounded AI computes from records — work orders, costs, manuals, usage logs — and can point to its sources. A generic chatbot generates plausible text. The difference decides whether you can repeat an answer in a budget meeting.
Ask exactly which records the AI can create or change with no person approving. For maintenance data that drives safety-critical work and audits, the defensible answer in 2026 is “nothing” — AI drafts, humans approve. If the vendor cannot enumerate the AI's write permissions, nobody has drawn the line.
Get AI pricing in writing: is it included, a premium tier, per seat, or metered by usage? All models exist in this category. A feature that is free in the pilot and metered at renewal changes the economics of relying on it daily.
A trial on your own fleet is the only benchmark that matters. Photo triage on your machines, manual questions on your models, anomaly scans on your fuel data. A vendor confident in its AI will let the product take the exam.
Working capability, 2026
Six capability classes are real, shipping, and verifiable in demos today. For each: what works, and the red flag that suggests labeling rather than capability.
| Capability class | What actually works | Marketing red flag |
|---|---|---|
| Photo understanding | Triage an issue from photos and a description; read printed part numbers and nameplates; review inspection photos. Modern vision models do this well. | Demos that only ever use the vendor's three sample photos. Ask to shoot your own. |
| Document-grounded answers | Answering spec and procedure questions from OEM manuals on file, with sources shown. | Manual answers with no manual on file and no label saying so. Ask what happens when the document is missing. |
| Drafting for approval | Work orders, PM schedules, and plans proposed from history and context, then edited and approved by a person. | “Autonomous” maintenance claims. Autonomy is not the feature; a good reviewable draft is. |
| Anomaly detection | Statistical screens on fuel burn, cost, and downtime patterns, narrated in plain language. The math finds it; the AI explains it. | “AI found savings” with no explanation of the screen behind it. Ask what a clean scan reports. |
| Plain-language Q&A | Questions like “what did this machine cost last quarter?” computed from the customer's records on demand. | Canned demo questions only. Ask something specific to the demo data that isn't on the script. |
| Import mapping | Reading messy spreadsheet columns and proposing field mappings, flagged for review before import. | “One-click migration” with no review step. Ask to see what happens to an ambiguous column. |
The honest limits
These limits apply to every vendor, Torgix included. A vendor who names them is telling you the truth; one who claims to have escaped them deserves your hardest questions.
Forecasting that a specific component will fail on a specific date generally requires condition monitoring — vibration, oil analysis, sensor trends — that most mixed fleets do not collect. Pattern-based warnings from usage, cost, and downtime data are real and useful; component-level failure prediction from work-order history alone is, for most fleets, marketing ahead of the product.
A language model answering from general knowledge will sometimes produce confident, wrong, plausible-sounding answers — the worst failure mode in a maintenance context. This is why grounding and explicit uncertainty labels matter more than model choice. Any AI answer that cannot cite its source should be treated as an estimate, whoever the vendor is.
AI triage ranks probable causes; the mechanic confirms the diagnosis. AI drafts the work order; the manager sets the priority. The systems that work in the field treat AI as a fast first pass for a person to finish — not a decision-maker. Staffing plans built on AI replacing technicians are ahead of the technology.
Our answers to our own test
Your records. Briefs, scans, analytics, and answers are computed from your own data; answers grounded in your documents carry a grounded badge, and anything unverified is explicitly labeled an estimate.
Nothing. All fifteen capabilities follow one rule, enforced in the product: the AI proposes, a person approves — work orders, PM schedules, issues, and imports alike.
Nothing beyond the plan. All fifteen AI capabilities are included at every tier, starting at $99/month: no AI tier, no per-seat AI pricing, no metered tokens. That is a claim we have not seen another vendor in this category make, and we make it in writing.
Yes — the free trial runs every AI capability on your real data from day one, no credit card required. Bring your own photos, your own manuals, and your own messy spreadsheets. See all fifteen capabilities.
Common questions
Watch the explainer
Real product screens. Figures shown are illustrative demo data. Prices approximate as of August 2026.
01 · Every maintenance software vendor says AI now. Some of it is real and useful. Some of it is a chatbot bolted to a brochure. This video is made by Torgix, which builds AI into its platform, so we are not neutral. But we will give you the same test we would want used on us.
02 · Here is what AI in maintenance genuinely does well today. It reads your actual records and answers questions with the machine, the number, and the source. It drafts. Repair plans, morning briefings, cost summaries. It never gets tired of the paperwork nobody likes. Watch it work on a live fleet. The question, the answer, the receipts.
03 · Here is the marketing version. Fully autonomous maintenance. AI that predicts every failure before it happens. Be skeptical of both claims. Prediction needs history and sensors your fleet may not have, and autonomy is exactly what you do not want between a guess and a work order. The vendors doing real work say what the AI reads and where a human signs off. The brochures say magic.
04 · Four questions for any demo. One, is the AI grounded? Ask it about a specific machine and make the vendor show where the answer came from. Two, can it write to your data, and if so, what stops it? The safe answer is read only analysis, with drafts a human approves. Watch for the approval gate on screen. Three, is the AI metered? Some vendors bill AI by usage credits, so ask what a month of real use costs. Four, does it degrade honestly? Ask it something it cannot know, and see whether it says so or makes something up.
05 · For the record, here is how Torgix answers its own test. The AI reads your records and cites them. It is read only by design, drafts plans, and nothing runs until a human ticks the box. It is included in the price, not metered. And when it cannot answer safely, it says so. The full written breakdown of real versus marketing is linked below, and the trial needs no credit card. Torgix.
Every AI capability, live on your data, in a free trial. No credit card required.