groundup.ai
11/18/2026Nine criteria for evaluating a Cognitive Maintenance platform, with a scoring rubric, vendor red flags and questions to ask before you buy or run a pilot.
To evaluate a Cognitive Maintenance platform, test whether it diagnoses faults (not just detects anomalies), explains its reasoning, delivers value without years of your own failure data, fits your existing work-order process, and proves results on your assets in a short pilot. Nine criteria cover it. The checklist below scores each one.
Every vendor demo looks good.
The dashboard is clean. The anomaly is caught. The presenter clicks through to a chart showing a bearing trending toward failure, and someone in the room nods.
Then the platform goes live on your plant, and the questions start. Why did it flag that pump? What does "anomaly score 0.82" mean? Who is meant to act on it? Why did nobody see the compressor trip coming?
The gap between the demo and the ops floor is where most maintenance technology purchases go wrong. This checklist is built to close it;, for maintenance and reliability leaders shortlisting vendors, and for the procurement teams who need a defensible way to compare them.
A Cognitive Maintenance platform is industrial AI that reasons about machine data, diagnoses the likely fault, explains why, and guides maintenance teams on what to do next, rather than only generating predictive alerts.
Predictive maintenance answers "is something changing?" That was a major step forward. But an alert still leaves the hardest work β interpreting the signal, naming the fault, deciding urgency β to an analyst or the technician on shift. When that capacity runs out, alerts pile up and trust drains away.
A cognitive platform takes on that interpretation. It answers three further questions:
That's the standard this checklist holds every vendor to.
Score every shortlisted vendor against these nine criteria.
Does the platform name the fault β bearing defect, misalignment, cavitation, seal distress β or does it only flag an anomaly? Ask to see real alerts from a live customer site, not a demo dataset.
Can a technician see why the system reached its diagnosis: which signals moved, which fault signature they match, how confident it is? If the answer is a black-box score, expect trust to erode by the third false alarm.
Does each finding come with a next step and a planning window β "replace drive-end bearing within 3β4 weeks" β that a planner can schedule? Detection without a recommendation pushes the hardest work back onto your team.
How long until the first useful diagnosis on your assets? Platforms built on large cross-industry fault libraries can diagnose from the early weeks. Platforms that learn only from your history may need a year or more of failures before they're useful.
Which signals does it read β vibration only, or sound, temperature, current and process data too? Some faults show up in sound or heat before vibration. More signals in context means earlier detection and fewer false alarms.
Two identical motors can behave very differently depending on duty and load. Does the platform learn each asset's normal individually, and account for variable speed and changing process conditions?
Do diagnoses flow into your CMMS or work-order process automatically, with the evidence attached? Do technician findings flow back so the model improves? A platform that lives in a separate dashboard becomes a second inbox.
Can sensors be fitted without stopping production? Does it work at remote or poorly connected sites? Where is data stored and processed, and does that meet your IT, OT security and data-residency requirements?
Will the vendor run a time-boxed pilot on your critical assets, against a baseline you agree upfront? And who supports the rollout: a ticket queue, or engineers who come to site and work alongside your crew?
Rate each criterion 0β3, weight the four cognitive criteria double, and require a minimum score on diagnosis and workflow fit before price is discussed.
Criterion | Weight | 0 β Absent | 1 β Basic | 2 β Strong | 3 β Proven on site |
1. Diagnosis | Γ2 | Anomaly score only | Generic fault category | Named fault per asset | Named faults confirmed at teardown on customer sites |
2. Explainability | Γ2 | Black box | Chart only | Evidence + confidence | Technicians verify reasoning unaided |
3. Action + urgency | Γ2 | None | Generic advice | Specific action | Action + planning window used in work orders |
4. Time to value | Γ2 | 12+ months of own data | 6β12 months | First diagnoses in weeks | Diagnoses in weeks on comparable sites |
5. Multi-signal | Γ1 | One signal | Two signals | Three or more | Signals fused to cut false alarms |
6. Per-asset baseline | Γ1 | Fixed thresholds | Fleet baseline | Per-asset baseline | Per-asset, load- and speed-aware |
7. Workflow fit | Γ1 | Dashboard only | Email or export | One-way CMMS link | Two-way CMMS loop |
8. Deployment + security | Γ1 | Line stop needed | Partial fit | Meets requirements | Proven at sites like yours |
9. Proof + support | Γ1 | No pilot | Remote pilot | Baseline-backed pilot | Pilot with on-site engineers |
π [Download the Excel checklist here]
Maximum score: 39. A useful bar: no vendor goes to commercial negotiation scoring below 2 on criterion 1 or criterion 7, however good the rest looks. A platform that can't diagnose, or can't get a diagnosis into your work orders, won't pay back.
Watch for anything that only works on the vendor's data, in the vendor's room, with the vendor's analyst driving.
Ten questions, designed to be put in an RFP or asked in the room β each one maps to a checklist criterion.
We built this checklist to hold ourselves to it. Here's where Groundup.ai stands. Score us the same way you'd score anyone else.
#CognitiveMaintenance is the category Groundup.ai was built for: machine intelligence that reasons, diagnoses and guides maintenance teams, rather than only generating alerts.
The best way to test any of this is on your own machines.
Predictive maintenance detects that a machine's condition is changing and raises an alert. Cognitive Maintenance goes further: it diagnoses the likely fault, explains the evidence and recommends the next action, so teams can act without interpreting raw data.
Root-cause fault diagnosis, explainable reasoning, recommended actions with urgency, a large fault library for fast time to value, multi-signal sensing, per-asset baselining, CMMS integration, secure deployment and a pilot on your own assets.
A desk evaluation against a checklist can take two to four weeks. A pilot on critical assets typically runs about 90 days, which is long enough to confirm diagnoses and compare against a baseline.
Not necessarily. Platforms built on cross-industry fault libraries can diagnose common faults on new deployments. Your own history still helps the model refine its view of each asset over time.
Use the same weighted scoring rubric for every vendor, ask the same questions, and insist on evidence from live customer sites rather than demo data.
Choosing on dashboard quality and sensor count instead of diagnosis quality and workflow fit. A platform that can't get a trusted diagnosis into your work orders rarely pays back.
A checklist tells you what to ask. A pilot tells you the answer.
Book a free 30-minute Asset Assessment and score us against this checklist on your own machines.
Join Our Wall Of β€οΈ Just Share And Tag Us!
We spotlight leaders driving innovation & performance, giving you visibility while shaping the future of AI
Related Posts