Manufacturing · 13 linked questions

Grounding the Line: A Maintenance Knowledge Assistant Across 40 Plants

A global industrial manufacturer replaces a legacy chatbot with a Claude-based maintenance and supplier-document assistant. Follow the build across edge-plant connectivity limits, an existing MCP tool layer, cost pressure, and a migration cutover — debugging symptom to architecture cause along the way.

One evolving scenario Questions build on the context Rate your confidence

How to use it · Read the brief, then work top to bottom — the situation evolves as you go. Rate your confidence before submitting; on review, the “ confident & wrong” questions are the misconceptions worth fixing first. A study tool, not the graded exam.

The scenario

Meridian Industrial runs 40 plants worldwide and wants a single assistant that answers maintenance questions ('how do I reset the torque calibration on Line 7?') and supplier-document questions ('what is the cure time in this bonding-agent datasheet?'). Today a rule-based legacy chatbot handles a narrow FAQ; technicians distrust it and call the help desk instead. The pilot must ground answers in real manuals and supplier PDFs, survive spotty plant connectivity, reuse the MCP tool layer already wired to the CMMS and document store, and stay inside a fixed monthly budget — all before a hard cutover retires the legacy bot.

Architecture: Claude agent over an existing MCP tool layer (CMMS lookup, supplier-doc retrieval, work-order write-back); RAG against maintenance manuals + supplier datasheets; regional API gateway per continent.
Existing stack: On-prem CMMS (Maximo-style), on-prem document store behind the plant firewall, a legacy intent-based chatbot, and an internal MCP server exposing three tools. Observability today is request logs only.
Integration: MCP tools bridge cloud Claude to on-prem systems via a per-region connector; supplier PDFs sync nightly; work-order writes must be auditable.
Regions: 40 plants across NA, EU, and APAC; several remote plants have intermittent satellite links with hours-long outages.
Budget: Fixed monthly ceiling; bulk supplier-doc reprocessing must not blow the interactive budget. No scaled 100–1000 score anywhere in scope.
SLA: Interactive answer p95 under 6s when connected; graceful degradation offline; every work-order write traceable end to end.

Keep this context in mind — later questions build on it, and the situation evolves as you go.

0 of 13 answered

P3 · Q1Grounding architecture / retrieval vs. parametric
The team debates how the assistant should answer 'what torque spec does the Line 7 gearbox manual list for the M12 bolts?'. The manuals are internal, revised quarterly, and vary by plant. What is the soundest primary design?
How confident are you?
The situation changes
Beat 1 — A plant reports confidently-wrong answers
Two weeks into the pilot, the APAC flagship plant escalates: the assistant gave a torque value that was confidently stated, fluently worded, and wrong — it belonged to a different gearbox revision. The retrieval logs show the correct passage WAS among the candidates fetched. The answer just did not reflect it. Leadership wants root cause, not a prompt tweak.
P7 · Q2Observability across agent + tool hopsSelect 2
To diagnose the confidently-wrong torque answer, the team currently has only top-level request logs. Which TWO observability capabilities would most directly let them locate where the correct passage was lost?
How confident are you?
P3 · Q3Ops symptom → architecture cause (confident-wrong)
The trace from Q2 confirms it: retrieval returned the correct passage, but an intermediate step compressed the fetched candidates into a short summary to save tokens, and that summary flattened two gearbox revisions into one figure. The model answered faithfully — from bad context. What is the correct root-cause reading and least-risky fix?
How confident are you?
P4 · Q4MCP least-privilege / tool scoping
The MCP work-order tool currently holds a CMMS credential that can create, edit, and DELETE any work order at any plant. The assistant only ever needs to open a new maintenance work order at the technician's own plant. What is the best change before wider rollout?
How confident are you?
P6 · Q5Cost — bulk vs. interactive workloads
Every night, 60,000 newly synced supplier PDFs must be summarized and tagged for the retrieval index. Running them through the same real-time endpoint the technicians use spikes cost and contends with daytime interactive traffic. What is the most cost-appropriate approach?
How confident are you?
P7 · Q6Edge resilience / intermittent connectivity
Remote plants lose their satellite link for hours. During an outage a technician asks the assistant a torque question and the cloud API is unreachable. What is the best-designed behavior?
How confident are you?
The situation changes
Beat 2 — A migration cutover date is set
The steering committee sets a hard date: in six weeks the legacy intent-based chatbot is switched off and the Claude assistant becomes the sole maintenance interface at all 40 plants. The legacy bot, for all its faults, reliably handled a core set of ~300 canned FAQ intents that technicians depend on daily. The committee asks the team how to de-risk the cutover.
P4 · Q7Migration — parity evaluation before cutover
With the cutover date fixed, what is the most important thing to do BEFORE flipping off the legacy chatbot?
How confident are you?
P3 · Q8Legacy / on-prem integration under the parity planSelect 2
The parity eval exposes gaps where the assistant can't reach data still living on-prem behind the plant firewall. Which TWO integration choices best close the gaps without weakening security?
How confident are you?
P6 · Q9Cost — repeated context / caching
Profiling shows the interactive endpoint resends the same large safety-and-procedure preamble (identical for every technician at a plant) on every single turn, dominating token spend. What is the most effective cost lever that preserves behavior?
How confident are you?
The situation changes
Beat 3 — An MCP tool starts mis-routing
Days before cutover, a new symptom appears: supplier-datasheet questions ('what is the cure time for adhesive X?') are being answered by the CMMS work-order tool, which returns irrelevant maintenance tickets. Nothing errors; the wrong tool is simply being chosen. The team recently reworded several tool descriptions to 'sound more helpful.'
P7 · Q10Ops symptom → architecture cause (mis-routing)
Investigating the mis-routing, the team confirms both tools are healthy and returning valid results — the agent is just picking the wrong one for supplier-doc questions. What is the most likely root cause and the right fix?
How confident are you?
P4 · Q11MCP tool design / routing hygieneSelect 2
To harden tool routing before cutover, which TWO practices most directly prevent the kind of mis-routing seen in Beat 3?
How confident are you?
P3 · Q12Improving grounded accuracy the right way
Post-cutover, a plant wants the assistant to be sharper on its own niche extruder line, whose quirks live in a local maintenance wiki. A vendor suggests 'just fine-tune Claude on the wiki.' What is the best-advised path?
How confident are you?
P7 · Q13Post-cutover observability / auditable writes
Cutover is done and the legacy bot is off. An auditor asks the team to prove that a specific work order the assistant created last month was grounded in the correct manual and authorized correctly. What must the architecture provide?
How confident are you?

All case studies