Can I trust AI in a CMMS? Only if it's governed and auditable.
SMMS's AI assistant, Elisa, operates under a fail-closed governance model: every mutating tool call is checked against a catalog-driven policy before it runs, every governed action is recorded to an audit ledger before execution, and audit-prep AI assistance strips any citation that doesn't resolve to a real source record. AI capability rollout is staged and enabled per tenant, off by default.
What can Elisa do today?
Elisa's capability rollout is staged and per-tenant — not every feature below is on for every tenant, and none of it is a general-purpose chatbot wrapped around your data. The genuinely supported list, as shipped:
Drafts PMs and maintenance plans
Elisa can draft preventive-maintenance plans and propose updates to existing ones. Every mutating draft runs through the same governed action policy below — nothing schedules itself onto your calendar without a human accepting it.
Opens and assigns corrective work orders
Work-order creation and assignment are two of the eight YELLOW mutating tools covered by the catalog-driven policy: permission-checked, work-type-access-checked, and idempotent against duplicate execution.
Parts availability, LOTO lookups, repair sequencing
Read-only lookups a technician can ask for mid-repair: checking spare-parts availability, retrieving the lockout/tagout sequence for an asset, and generating an ordered repair walk-through. Every one of these fails closed to "no procedure on file — do not fabricate" rather than inventing steps when nothing is on record.
Tightens RCFA write-ups — rewrite-only
Elisa rewrites the wording of an RCFA narrative a human investigator already entered. It does not conduct investigations, select root causes, or fill in evidence on its own — every finding on the page is the investigator’s.
How RCFA works in SMMSHow is AI governed in SMMS?
This is the differentiator: rather than a general permission check bolted on afterward, every mutating AI tool call runs through the same gate stack, and an unrecognized state denies the action instead of defaulting to allow.
All 8 mutating AI tools (create/assign work order, create/update maintenance plan, adopt/publish/archive maintenance plan, generate OEM maintenance plans) are routed through a single governed path: a database catalog resolves each tool’s mutation_mode before it can run.
An unresolved catalog row, an unlisted agent, or an unrecognized mutation_mode denies the action — there is no default-allow path. A mode of forbidden, NULL, or anything the policy doesn’t recognize is a deny, not a pass-through.
Every governed action reserves an execution record before it runs (insert-before-execute), so a crash mid-action can’t silently double-execute or vanish — a repeat request either replays the recorded result or is refused, never re-run blind.
Audit-prep AI assistance validates every citation it emits against the exact source-record set that was in its own prompt — hallucinated references are stripped and recorded, never silently kept, and an output left with zero valid citations is refused before it is even stored as a draft.
A governed action, end to end
A technician at Meridian Fabrication Works (a synthetic demo tenant) asks Elisa to open a corrective work order for a failed conveyor bearing. Elisa resolves "create work order" against the policy catalog: the tool is active and approved, the service-layer flag is on, the technician's account holds the same work_orders:create permission the REST route itself would require, and the catalog's mutation mode allows the action to proceed. An execution record is reserved before the handler runs, so if the request is retried (a flaky connection, a double-tap), the second attempt replays the first result instead of creating a duplicate work order. The work order that lands in the queue is identical to one a human would have created by hand through the UI — same permission check, same validation, same audit trail — the only difference is who typed the request.
Limitations
SMMS is pre-revenue with no deployed customer sites — every capability claim on this page is about what the product does today, not about adoption numbers or testimonials that don't exist yet to cite honestly.
- Does not conduct RCFAs — rewrites narrative text a human investigator already entered; it does not select root causes or determine findings.
- No autonomous action outside the audited tool layer — every mutating action is gated by the catalog-driven policy above; there is no separate, ungoverned code path for the AI to act through.
- No wake-word or always-listening voice engine ships today — Elisa’s voice interface is push-to-talk in the mobile app.
- External-document retrieval is gated, not open — Elisa answers from your tenant’s own scoped maintenance data and connected knowledge sources, not an unrestricted web crawl.
Frequently asked questions
What can Elisa (SMMS’s AI assistant) do today?
Elisa can draft PMs and maintenance plans, open and assign corrective work orders, look up parts availability and LOTO procedures, generate repair sequences, and rewrite RCFA narrative text under a strict no-hallucinations policy. AI features are enabled per tenant and fail closed by default — an unentitled tenant simply doesn’t see the surface.
How is AI governed in SMMS — what stops it from doing something it shouldn’t?
Every mutating tool call resolves against a database-driven policy catalog before it runs: an unresolved catalog row, an unlisted agent, or an unrecognized mutation mode denies the action rather than defaulting to allow. Actions are reserved (recorded) before execution for idempotency, and permission checks reuse the exact same authorization code the REST routes use — nothing is re-implemented in a separate, less-checked path for AI calls.
Does Elisa conduct RCFAs or FMEA analysis for me?
No. Elisa rewrites RCFA narrative text a human investigator already entered — it does not conduct investigations, select root causes, or determine an FMEA worksheet’s Severity/Occurrence/Detection scores. Every finding and every score on the page is the analyst’s, not the model’s.
Can I turn AI off for my tenant?
Yes. AI capability is gated per tenant behind feature flags that are off by default; an unentitled tenant sees no AI surface at all, and admins can decline the rollout.
Is every AI action logged?
Every governed mutating action is recorded to an execution ledger before it runs and its result is recorded after — the audit trail exists independent of, and prior to, any human review of the action.
Does SMMS train AI models on my data?
SMMS’s Privacy Policy discloses that when you use Elisa, your message text — and your voice audio and its transcription, if you use voice — is sent to OpenAI to generate Elisa’s response (see /privacy for the exact vendor-disclosure wording). The current policy does not make a separate statement about model training either way; treat that as an open question to confirm with your implementation contact rather than an assumption in either direction, and check /privacy for the current, canonical wording before relying on this answer.
Related pages
See governed AI on your own maintenance data.
Request a demo and we'll walk through the policy catalog and audit trail on a tenant shaped like your fleet — or take the free RMI QuickScan first to see where your program stands today.