Most manufacturers don’t have an AI problem. They have a timing problem. The data that would have alerted them to a problem is already in their historians, MES, quality systems, and SAP systems. It’s just sitting in silos. People only find out what’s happened once the line has stopped. The impact of a line stop is immediate: lost output, idle crews, late deliveries. Unplanned downtime is just a constant drain on every month’s production.

The industry has run a number of AI pilots that can answer questions. A whole lot fewer have made it to the plant floor, where the answer has to become a work order, a revised plan, or a supplier order, and someone has to be behind it.

That gap is what we set out to close. Our AI Lab has built eight working agentic AI use cases in manufacturing and automotive on ALTi AIOS™ (Altimetrik’s AI Operating System) on Google Cloud. These cover the plant floor, the supply chain, warranty, and field service.

How the agents work

All 66 agents across the eight use cases work in exactly the same way. They monitor the systems the plant is already running, catch problems early, and prepare the fix (this could be a work order, a revised production plan, a backup supplier, or a claim decision). A named individual has to approve it before the agent goes and does anything. They then check afterward that the fix has in fact worked.

Agents prepare the fix, a named person approves, then agents act and check

Watchexisting systems around the clock
Detectproblems early with evidence
Preparethe fix or plan for review
Approvea named person signs off
Act and checkagents act, then verify it worked
On top of the SAP, MES, quality, dealer and procurement systems already in place



Two options, really. One is we do nothing at all. The agents work on top of the SAP, MES, quality, dealer, and procurement systems, so there’s no rip-and-replace program before you actually see value.

The other is that people remain accountable. Every recommendation is traceable back to the record it came from, and the decisions that matter have a person’s name on them.

Eight use cases, from the plant floor to the field

Each one is a working system with a live demo. The results below come from those demos on sample plant data, so they show what the agents can do, not a customer outcome.

1. Predictive Maintenance Advisor

Most alarms only go off once a reading crosses a fixed threshold, so slow degradation goes unnoticed until the machine actually breaks. The agents learn what ‘normal’ is for each machine and can then predict a failure seven days in advance, tell you what the likely cause is, and produce the work order. The planner approves it, and the agents then raise it in SAP, alert the technician, and check the repair. In the demo, they picked up and fixed most failures before they stopped the line.

  • How it works: An orchestrator coordinates ten agents. Detection learns the normal for each machine per product model, so a switch from sedans to SUVs isn’t going to be picked up as a fault. Prediction gives you a 7-day failure risk per part, and a correlation agent will link the wear to the rework that you subsequently find at quality.
  • Where people decide: The planner either approves or rejects each work order with a reason, and the most urgent work will never close on its own.
  • It learns: After each repair, a feedback agent records what was found, and a retrained model only goes live if it tests no worse. Over time, technicians are confirming more of its findings.
  • Connects to: Plant historians, MES, Computerized Maintenance Management System (CMMS), and ERP via Open Platform Communications Unified Architecture (OPC UA), Message Queuing Telemetry Transport (MQTT), Plant Information (PI), or SAP, with work orders written to SAP PM or IBM Maximo.

2. RootIQ: Process Root Cause Analysis

When a batch goes wrong, you often get a whole raft of readings moving at the same time, and the biggest alarm is just a symptom. RootIQ monitors every batch, uses multiple methods to rank likely causes against each other, and then checks each against the history of incidents. An analyst signs off on the report. In the demo, the real root cause is invariably one of the top few suggestions, with no analyst required.

  • How it works: Five agents operate one pipeline. The statistical and causal methods run in code, score each batch and rank the causes, and then Gemini provides the explanation in plain language.
  • Why it can be trusted: The data is checked prior to analysis, a human selects the method, and every sentence in the report will cite a stored record. Any text that fails a check is simply dropped.
  • Where the decision lies: A supervisor looks at each alert, an analyst signs off on each cause, and the process manager raises the follow-up task.
  • Getting started: A pilot on one production line takes about 10 weeks, including the blind test against the plant’s own past incidents.

3. Shift Handover and Line Performance

Handovers still occur on whiteboards and in people’s memory, so open repairs and quality holds get lost between shifts. The agents will explain where the production time was lost, will follow up on every open item, and will draft the handover 30 minutes prior to shift end, with every fact traceable back to its source record. In the demo, every minute of lost time was accounted for.

  • How it works: Eight agents share one view of the shift. They refresh the MES, downtime, quality, and maintenance data every 30 minutes and calculate line efficiency using Overall Equipment Effectiveness (OEE) and the six standard loss categories in code, compared with the last seven shifts.
  • Follow-through: Open tickets and quality holds are chased every 30 minutes, and the owners are kept up to date when anything changes. A closed fix is tested on later shifts and re-opened with evidence if the stop comes back.
  • Where the people decide: The outgoing supervisor edits and signs the handover, and the incoming supervisor then accepts it item by item.
  • Links to RootIQ: When the same loss appears in four of the seven shifts, the agents will propose a root cause analysis for an engineer to approve.

4. Shop Floor Knowledge Copilot

Operators can easily lose a significant amount of a shift just searching for the right procedure, and general chatbots will often guess or reword the safety steps. This copilot answers at the machine (by voice or text) in the operator’s own language and only draws from the approved procedures. The safety steps are quoted exactly as they appear in the procedure, and it will pass on any questions it can’t answer to an expert. In the demo, the vast majority of the guided jobs were done correctly the first time.

  • How it works: There are 14 agents in total. 7 work with the people in real time (answering questions, turning the alarm codes into the correct fix, checking photos of the key steps, etc.). 6 work in the background proposing changes, and there’s an orchestrator that coordinates all of them.
  • Why it’s safe on a live line: Answers come only from current approved procedures (with the source page shown). The code enforces the rule that safety steps are quoted exactly as in the procedure, not left to the model.
  • Captures know-how: An operator’s spoken note becomes a stoppage, a rework, or a handover record. An expert’s answer becomes a new procedure only after it’s been approved by two people.
  • Connects to: Document stores such as SharePoint, plus SAP Plant Maintenance (PM), MES, and quality systems. In the demo, it passed every safety and accuracy test.

5. Production Scheduling and Changeover

When a tester fails, or a parts lot is held, the planners will often go in and rebuild the plan by hand whilst the line is waiting. That agent takes the production target, considers delivery, changeover time, and cost, and replans in minutes when the floor changes. It will deal with the small slips itself and will raise the issue of the bigger ones with the planner. In the demo, it halved the changeover time.

  • How it works: One planning agent using ten skills (demand, capacity, scarce materials, pilot builds, deviations). Google OR-Tools does the sequencing, and then Gemini or the plant’s own rules make the call. Every step is recorded.
  • The correction loop: A deviation monitor will compare actual output against the plan. If a line is more than 30 minutes behind, the agent will replan from that point and run the slow line at its actual speed until it’s repaired.
  • Guardrails: The agent only acts on its own for small fixes within the prescribed limits of time, capacity, just-in-time deliveries, and overtime. Any bigger changes go on hold for the planner, and every published plan version is kept.
  • In the demo: Every line-side just-in-time delivery got through on time, and overtime during a parts hold almost disappeared.

6. Supplier Risk and Sourcing

Risk alerts typically name the supplier rather than the part, the line, or the number of days of stock at risk, and finding a back-up can take weeks. The agents will check each alert against SAP, Ariba, quality records, and contracts to establish what is at risk and when, what the alternatives are, and then produce a costed plan for the buyer to approve. In the demo, they gave weeks’ notice of a supplier’s shipment hold.

  • How it works: An orchestrator runs 13 agents. Each supplier/part pair is scored on the basis of likelihood x exposure, combining the financial, delivery, quality, and event signals with the part’s share, stock cover, and lead time. All the scores, dates, and money are calculated in code.
  • Sees the sub-tier: A sub-tier agent will follow the parts lists and the customs records so an event at a chip maker can be traced back to the modules that were built from its chips.
  • Know the stop date: For each vehicle program, the agents will work out the day the line would stop and then put forward costed options (additional buffer stock, alternative supplier, split order, Request for Quotation (RFQ), or supplier notice).
  • Where the decision lies: The buyer within their own limit and the Head of Procurement above that. Nothing gets ordered until someone actually says yes, and then the order/sourcing event is fed into SAP or Ariba.

7. Warranty Claims Intelligence

Dealer claims can sit for days before they’re settled, and fraud and errors creep up on the warranty costs. The agents will check every claim against the vehicle data and the policy, settle the straightforward claims in seconds, and pass on the exceptions to an adjuster. They’ll also pick up on field failures early and re-file the claim with the supplier. In the demo, most of the claims were settled with no manual intervention at all, and the field issues were picked up well in advance.

  • How it works: Fifteen agents cover the entire warranty lifecycle. The intake, policy, and anomaly agents all make the decision on each claim, looking at fault codes, component readings, mileage, and repair history to check for rollbacks, duplicates, inflated labor, and parts that the vehicle never needed.
  • Early warning: An early warning agent is watching the plant tests, the dealer hotline, the fault codes, and the claims all together. A root cause agent will then go on to trace the failure back to a supplier lot and build date.
  • Recovers costs: A recovery agent applies the supplier contract share, cap, and time window claim by claim and files the chargeback with an evidence pack.
  • Gets ahead of the repairs: A vehicle with a fault code is identified prior to it even getting to the dealer, so the parts are already in place. In the demo, there were no bad claims paid, and every exception did get to an adjuster with a stated reason.

8. Field Service Assistant

A lot of service calls still end up requiring a return visit, and the technicians frequently lose signal in the basements and machine rooms. The assistant creates an offline job pack for each ticket, walks the technician through the repair step by step with the exact part highlighted in the image, and will keep going even if the network goes away. Guided work results in fewer return visits and less rework.

  • How it works: Across two pipelines. In the cloud, the manuals, procedures, and diagrams are processed once into image-linked steps in a knowledge store. On site, the assistant looks at the asset on the ticket first and only falls back to the wider knowledge base if it can’t find what it needs.
  • Works offline: A job pack and Gemma (Google’s on-device model) are downloaded to the technician’s device so the session continues with full context should the network drop, and then syncs to the ticket later.
  • Adapts to the fault: The next step is determined by what the technician reports, the fault status, and the steps that have already been done rather than being a fixed script.
  • Hands-free: Technicians can ask by voice or text, and every step shows the exact component that’s marked on the diagram. In the demo run, we didn’t need any expert video call at all.

Why one platform matters

It would be easy to build eight separate tools. We chose to build eight use cases on one platform, as the difficult bit about industrial AI isn’t really the first use case. The hard bit is the fifth one, where you’ve got the same controls, same audit trail, and same integrations required across the board.

ALTi AIOS™ provides that common framework. Each use case is connected to the plant historians, MES, SAP and the quality systems via the same governed connectors. Each use case goes through the same approval steps, records every agent action and decision, and is entirely based on the client’s own approved data. The models do the reading, the reasoning, and the drafting. The fixed rules and calculations do the exact bits (production schedules, alarm thresholds, claim amounts, etc.).

The foundation is Google Cloud: Gemini on Vertex AI for reasoning, the Agent Development Kit for building agents, BigQuery for the plant and business data, and Gemma on the device so the Field Service Assistant can keep going offline. For a manufacturer, that means a second or third use case is just reusing what the first one built rather than building it from scratch.

How it’s built

Every layer of  ALTi AIOS™ for Manufacturing is a managed Google Cloud service or an open standard, so each use case is compatible with the client’s security model and the systems they’re already running.

ALTi AIOS for Manufacturing

Five layers on Google Cloud, governed end to end

01

Experiences

Use-case dashboardsGemini Live voiceCommand Center
02

Data and Integrations

Data
FirestoreBigQueryCloud Storage
Integrations
MCP serversIntegration ConnectorsSAP PM and S/4HANAIBM MaximoSenseye and AuguryAriba and CoupaRisk feedsDealer DMS and telematics
03

Orchestration, Agents and Grounding

Orchestration
ADKAgent EngineCloud Run
Agents
66 governed agentsAgent StudioGemini
Grounding
RAG EngineVector SearchDocument AICited evidence registry
04

Governance

Four-eyes approvalGolden evaluation gateEvidence drift guardHash-chained auditModel Armor
05

Substrate

Cloud KMS CMEKVPC Service ControlsIAP


ALTi AIOS for Manufacturing: five layers on Google Cloud, governed end to end

  • Experiences: Each use case has its own dashboard with voice via Gemini Live and a command center across the use cases.
  • Data and integrations: Plant, case, and business data is held in Firestore, BigQuery, and Cloud Storage. The MCP servers and Integration Connectors feed into SAP PM and S/4HANA, IBM Maximo, the various condition monitoring tools (Senseye, Augury, etc.), Ariba and Coupa, the risk feeds, and the dealer and telematics systems.
  • Orchestration, agents and grounding: One orchestrator per use case (built with Google’s Agent Development Kit (ADK) and running on Agent Engine and Cloud Run) to route each task to the correct specialist agent. 66 governed agents (built in Agent Studio and powered by Gemini) do the work. A Retrieval-Augmented Generation (RAG) Engine, Vector Search, and Document AI tie answers to the client’s documents and data, and a cited evidence registry records what each answer was based on.
  • Governance: Four-eyes approval via a golden evaluation gate that new models must pass (a safeguard against drift in the evidence), a tamper-evident audit trail, and Model Armor screening of both prompts and responses.
  • Substrate: Customer-managed encryption keys, Virtual Private Cloud (VPC) Service Controls and Identity-Aware Proxy to keep data and access within the customer’s boundary.

Here’s what the approval step looks like in practice. The example below is simplified pseudocode for the Predictive Maintenance Advisor (to illustrate the pattern rather than actual production code).

# Illustrative only: simplified to show the pattern, not production code.

def handle_failure_risk(asset):
    risk = predict_failure(asset, horizon_days=7)      # the model scores the risk
    if risk.score < policy.alert_threshold:          # fixed policy, not a model guess
        return

    order = draft_work_order(asset, risk.evidence)    # the agent drafts the fix, citing evidence
    decision = request_approval(order, approver="maintenance_planner")

    if decision.approved:
        sap_pm.create_work_order(order)              # only now does anything reach SAP
        notify(order.technician)
        schedule_check(order, after="repair")        # agents check that the fix worked

    audit.record(order, decision)                    # every outcome is logged, approved or not

The model just scores and drafts. Policy sets the threshold, a named planner makes the call, and the audit record is written regardless of whether the order is approved or not.

Because every use case sits on the same layers, adding the next one is mostly configuration rather than a new build.

The same design also ensures that clients aren’t locked in. The assets and components are loosely coupled so they can be ported to OpenAI and Anthropic models as well as to AWS and Azure. A manufacturer can run them in the cloud with the models of their choice, with the same agents, approvals, and audit trail.

To learn how Agentic AI can enhance your business operations, reach out to us at aifirst@altimetrik.com or book a consultation with our AI engineering experts

Here

Contact Us

We'd love to hear from you.
Contact Us