Last week we launched something I’ve wanted to build for years: you can now ask about your global device fleet in plain language, inside ChatGPT, Claude, or Gemini, and get real answers from real operational data. I want to use this issue not to sell it, but to explain the part of the problem that took the longest — because it’s the part every company connecting AI to its systems this year is about to run into.
The hard part of AI in IT operations isn’t the interface — it’s whether the record underneath is true. Every enterprise software category is shipping an assistant this year, which means asking will be a normal way to operate business systems within about twelve months and the interface will stop being a differentiator almost immediately. What’s left is the unglamorous question: is the data the assistant reads accurate? For IT asset data the answer is usually no — only 45% of organizations can see all their devices in a single view, and nearly 1 in 3 devices goes unaccounted for at offboarding. Fix global IT asset management first, then connect the AI. The sequence is the whole argument.
The signal: the interface question is already settled
Every enterprise software category is shipping an AI layer right now. Connectors, assistants, copilots. Within a year, asking will be a normal way to operate business systems, and the interface question will be settled.
Which means the interface stops being the differentiator almost immediately. What’s left is a much less exciting question: is the data underneath true?
For IT asset data specifically, the answer is usually no. Research this year found only 45% of organizations can see all their devices in a single view, and one company running proper discovery found 60x more devices than its records showed. Nearly 1 in 3 devices goes unaccounted for at offboarding (Gartner).
The problem: a wrong record is passive; a wrong answer is not
A wrong spreadsheet is passively wrong. Someone opens it, squints, moves on. Bad data sitting still is a slow leak.
An AI assistant reading a wrong record is actively wrong. Ask it “which devices are unaccounted for?” and it won’t hedge, won’t caveat, won’t mention that the record was last reconciled in March. It answers — fluently, instantly, in the same confident tone it uses when it’s right. Then someone makes a procurement decision, an audit response, or a security assessment on that answer.
The model isn’t malfunctioning. It’s faithfully reporting what the system told it. The system was wrong, and the model had no way to know.
So AI does something to a data-quality problem that no dashboard ever did: it removes the friction that used to expose it. The slowness of manual work was, accidentally, a quality control — someone chasing a device discovered the record was stale. Remove that friction without fixing the record underneath and you don’t get efficiency.
You get confident error at machine speed.
The operator takeaway: three questions before you connect an assistant
The sequence matters, and most of the market has it backwards. Fix the record, then connect the AI. Three questions to ask before pointing an assistant at your IT data:
1. How does this record get updated?
If the answer involves a person remembering, it drifts — and the AI will report the drift with total confidence. Records generated as a byproduct of the work (the deployment writes itself in; the recovery closes itself out) can’t drift far, because nobody has to remember anything.
2. When was it last checked against physical reality?
Not report-against-report. Rows against actual devices. Your last ten offboardings are a free audit: how many produced a recovered, wiped, documented device?
3. Can it tell you how confident to be?
The most valuable thing an AI layer over operational data can do is expose its own uncertainty — this order hasn’t moved in 11 days; this device’s last confirmed location is 60 days old. An assistant that can’t distinguish fresh truth from stale record is a liability dressed as a productivity tool.
One GroWrk lens
Here’s why we were willing to connect an AI to our platform, and where it currently stops.
We didn’t bolt AI onto a database someone maintains. The record is produced by the work: a device is sourced and delivered — that writes the record. It’s recovered and wiped — that closes it. A destruction certificate is issued — that’s attached. There’s no separate act of bookkeeping to forget, which is the only reason the answers are worth trusting.
And the honest limits, because this issue’s whole argument depends on saying them: today the AI / MCP layer does a bounded set of things well — reading fleet, employee, and order information, and triggering standard workflows. Deeper blocker and SLA visibility, broader workflow actions with role-based permissions and confirmations on sensitive steps — those are next, not now.
The honest version of an AI-native pitch isn’t “ask it anything.” It’s “here’s why the answers are trustworthy, and here’s exactly where it currently stops.”
One stat
60x. The gap one company found between its device record and its actual devices.
Now imagine an assistant answering fleet questions from that record. The model would be working perfectly. Every answer would be wrong.
If you’re connecting AI to your IT stack this year — and you probably are — the highest-leverage work isn’t the integration. It’s making sure the record it reads is generated by reality. So: how does your asset record get updated today, by the work, or by a person remembering?
If you want to see a record that’s generated by the work — and an AI layer that’s honest about where it stops — we would like to show you.
Book a demo →
