Back to EdgeBits
Analytics · Assistant · Plain English

Ask your plant in English

July 2026 · 5 min read

For twenty years, the answer to "what happened on shift B" has been a PowerPoint.

Sun 22:00
Shift B loses two hours
Line 3, feed pump
Mon
Analyst pulls data
Historian + SCADA + ERP
Tue
Excel cleaning
Reformat, join, chart
Wed 09:30
Deck lands
Shift D already repeated it

The data was never missing. It was in the historian, the SCADA, and the ERP. What was missing was a query language a plant manager could speak. So the query language became a human, and the human ran on weekly cadence.

English is finally a good query language for factories. It stopped being a demo trick sometime in 2025.

What actually shifted

A standard protocol for connecting language models to real tools (MCP), and models good enough at multi-step reasoning to use them without falling over. Together they mean an assistant can read live tags, pull historical aggregates, walk the audit chain, and cite each source it used — without you writing bespoke agent code for every question.

The assistant page walks the surfaces the model gets: readings, deploy history, and the fasten audit chain, each exposed as a tool with a role check on every call.

Three roles, one assistant

Same tools. Different scope. The scope check happens at the tool layer, not in the UI — an operator cannot ask about the plant next door even by accident.

Owner / MD
all plants · roll-up
Which of my five plants had the highest energy cost per unit last month?
Plant manager
their plant only
How is line 3 today compared to last Tuesday same shift?
Shift engineer
their line or area
Show me pump current just before the 3:47 AM drop.

What a real thread looks like

Owner asks a cross-plant question. Plant manager drills into their own plant on the same thread. The assistant weaves live and historical data and cites what it used.

Ask your plant
Owner · all plants
Which plant had the worst OEE last week?
Assistant
Pune. 68% average, down from 74% the week before. Line 3 feed pump: 6 outages, clustered around Tuesday 03:47–05:20. Recommended fix: raise the connector's reconnect timeout from 30 s to 120 s so the recovery is automatic.
Pune plant manager · drill-down
Show me pump current just before the first drop.
Assistant
Current dropped from 24 A to 0 A at 03:47:12. Two minutes prior, the OPC-UA subscription showed a keep-alive gap. Network layer, not the pump. Fasten reference a1b2c3…

Every claim in an answer points at a specific reading, a specific window, or a specific fasten event. When someone asks "prove that", the trail is right there.

Cloud or self-hosted, your choice

The assistant speaks the same protocol either way. You pick the model, and you can change your mind later without changing anything else in the stack.

Fastest to start

Cloud AI

Claude, GPT, or similar hosted models
  • Battle-tested reasoning
  • Priced per query
  • Data leaves the site only on ask
Data stays with you

Self-hosted AI

On a server you own
  • Nothing leaves your environment
  • Fixed cost, no per-query billing
  • Right choice for regulated industries

The choice usually turns on data-localisation obligations. If your plant is under DPDP-style rules or your customer contracts prohibit sending raw data offshore, self-hosted is the answer and the assistant experience is identical.

Where this replaces work

The obvious win is the shift-review meeting. The less obvious win is the tail of small queries that never made it to the meeting at all — because the friction of formulating them in SQL was too high.

The tail becomes the trunk
Over months, the queries the plant did not know it was asking get answered by default.

The weekly deck shrinks to a summary of what the plant already agreed on. The plant manager's calendar frees up.

Wire the assistant to one line

30-minute call. We connect the assistant to a slice of your existing data and let you ask it questions live. You keep the model choice.

See how the assistant works