What decision this helps ALTEN make
A lightweight operating loop for ALTEN's Asia AI delivery: one intake, one triage, one scorecard, one rolled-up view the central AI team in France reads across all five markets without losing local texture.
The loop runs once per client mandate. Each scorecard rolls up to France. Cross-market priorities feed back into the next intake.
After triage, every mandate routes to one of three calls
Pilot now
Data ready, shape fits, regulatory gate clear. Scope to one line for 90 days. Re-score at day 60.
Collect data first
The right lens exists but the data isn't ready. Convert to a 6-week data-collection programme, then re-triage.
Not yet
Wrong-answer cost is above embarrassing with no human in the loop, or the chatbot has nothing to ground on. Refuse, and say so.
Worked example
Korea automotive · predictive maintenance pilot
Consider a Korean Tier-1 automotive OEM whose maintenance organisation is approaching a generational handover, with its most senior engineers within a few years of retirement. An Ambassador running this framework lands on a two-thread pilot.
Thread A is a predictive-AI model on the line's vibration and temperature time series for asset-health forecasting. Labelled history already exists, so the ROI demonstration lands fast.
Thread B is an autonomous-AI changeover orchestrator trained off a machine-teaching session with the senior maintenance engineer before they leave. The session captures the tribal "if it's a hot day, turn the knob down" rules into a deterministic decision tree, which the autonomous layer can replan against in real time.
Generative AI sits only at the operator-facing surface, providing a natural-language interface to questions like "why did you change the setpoint?" That split (generative on the surface, deterministic underneath) is the durable architecture for any pilot where the cost of a wrong answer is above embarrassing.
Built from publicly knowable signal: the ALTEN Italia ball-bearing predictive-maintenance precedent (Azure ML, ~2% scrap reduction) and the Mistral / Prisme.ai stack. This is not a current ALTEN engagement; it extends the same playbook from quality prediction in Italy to asset-health prediction in Korea, under Korea's PIPA regime.
The management artifact
The pilot scorecard
The scorecard is what the Ambassador captures after the call. One row per candidate pilot, filled in before any kickoff is committed, and rolled up to France in the same format across all five markets.
| Field | What to capture | Worked example (Korea automotive, Thread A) |
|---|---|---|
| Country / vertical | One row per pilot | Korea / Automotive (Tier-1 OEM) |
| Pilot shape | Primary shape; hybrid notes separately | Predictive (primary) plus Generative on the operator-facing surface only |
| Data readiness | Labelled? Volume? Cleanliness? | Vibration plus temperature time series, 3+ years of history (illustrative), plant-side telemetry cross-labelled by maintenance log |
| Wrong-answer cost | Embarrassing, expensive, or safety-critical | Missed failure: line stoppage (hours), low direct safety risk. False positive: unnecessary swap (around one shift). |
| Human-in-loop requirement | Advisory, approval, override, none | Approval, with the maintenance lead signing off before any swap. |
| Deployment mode | Cloud, edge, hybrid | Hybrid: cloud training using ALTEN-aligned AI stack, with edge inference for latency and sovereignty. |
| Regulatory gate | PIPL, APPI, DPDP, or other | In-Korea PIPA review; APPI if cross-listed with Japan; no PII so primarily an export-control review. |
| KPI | Adoption, cycle time, scrap, yield, downtime, cost. Pick the 1 or 2 that decide the pilot. | Unplanned downtime hours (primary); maintenance cost per asset (secondary). |
| Scale decision | Kill, collect data, pilot, scale | Pilot on one line for 90 days, with the scorecard re-evaluated at day 60. (Illustrative.) |
The scorecard is also the kill-switch. If data readiness is empty and the regulatory gate is unresolved, the only honest scale decision is collect data, because the pilot doesn't exist yet. And given model drift, even a "scale" decision is never permanent; the same row gets re-scored quarterly on the line.
Asia Manager's role
Why I think this is the Asia Manager's job. Ambassadors can run the same intake locally. What sits with the Asia Manager is the consistency of the scorecard, the country overrides, the escalation path to France, and the call between pilot, collect data, or stop.
Cross-regional play
Will this scorecard travel? The same Ambassador conversation lands differently in each Asia market. Localisation is a layer, not a separate strategy.
| Country | What the Ambassador adapts | Regulatory gate | First pilot fit |
|---|---|---|---|
| India | Developer-led workshops; English-first materials; local-language support where buyer teams need it | DPDP Act | Predictive and computer-vision pilots compound fastest |
| Vietnam | Greenfield deployment; vendor-translated materials | Local privacy/security review; confirm client data-hosting requirements | Analytical and generative pilots without deep data lineage |
| Korea | Hierarchy-respecting approval; working pilot before broader rollout; Korean-language plant materials | PIPA | Predictive (see worked example above) |
| Japan | Slow-consensus sign-off; quality-first criteria; Japanese-language plant materials | APPI | Operational shapes earn trust faster than generative |
| China | Sovereign-cloud or edge deployment; Chinese-language materials | PIPL | Deployment mode decides before the shape does |
Cross-regional collaboration follows a simple rule: one country pilots, the scorecard row travels with the win, the next country adapts it to its own gates. The central AI team in France reads a regional view from the same scorecard rolled up across all five markets.
Pilot shapes
The first classification an Ambassador makes, before scope, KPI, or deployment mode.
| Shape | What it does | Where it lands at ALTEN | Repeatability |
|---|---|---|---|
| Generative | LLM text and image generation; natural-language interfaces | Engineering productivity (code fixing, doc summarisation, RFP drafting), Ambassador enablement | Non-repeatable; can hallucinate |
| Analytical | Pattern detection in historical operational data | Quality root-cause, yield analysis, MES log mining | Repeatable and auditable on fixed inputs |
| Predictive | Forecasting on operational time series | Aircraft part-breakage cost (ALTEN Digilab Toulouse and Airbus), predictive maintenance, asset health | Repeatable, model-bounded |
| Computer vision | Visual inspection, safety, asset health | Surface defect detection, EHS perimeter monitoring, weld and seam QA | Repeatable on calibrated inputs |
| Autonomous (DRL) | Deep reinforcement learning; long-horizon strategy under constraints | Real-time process control, changeover orchestration, repeatable plant-floor decisions | Repeatable, with explicit return-control when out of training distribution |
Grounding: the shape taxonomy follows ALTEN's own published categories. The AI in Manufacturing materials cluster "machine learning, computer vision and predictive analytics" as the industrial set; the A3 programme reports generative AI productivity gains of +25% to +37% across software development, test generation, and knowledge management; and the Prisme.ai partnership covers the autonomous (DRL) layer for critical sectors. The verticals in the matrix below are ALTEN's own service lines.
The 5-question triage — picking a shape
Run this before any pilot scope is written.
- Is the output consumed by a human reader, or by a machine or control loop? Human reader, generative may fit. Machine, generative is almost always wrong.
- What is the cost of a wrong answer once? Embarrassing email, ruined batch, or injured operator. Anything above embarrassing disqualifies non-repeatable AI as the primary shape.
- Does the same input need to produce the same output every time? Yes, autonomous, predictive, or CV. No, generative is in-scope.
- Is the value in the natural-language surface or in the underlying strategy? Surface only, generative. Strategy, autonomous, and the surface (if any) is a thin generative wrapper.
- Does the client have the data the shape needs? Predictive needs labelled history. CV needs image volume and label quality. Autonomous needs a simulator and a machine-teaching session with an expert before they retire.
If the answers point to a shape the client does not have data for, name it what it actually is: a data-collection programme. Pilots come later.
The shape matrix
Every (vertical, shape) pair is a possible pilot. The two highlighted cells in Automotive are the worked Korean example above.
When to refuse a pilot
Refusal is mechanical when the scorecard says so. Refuse, or re-scope, when:
- The pilot is generative-AI-led in a process that requires repeatable output, with no human in the loop.
- The client wants a chatbot grounded on PDFs but has no plan to build the IT, ET, and OT knowledge graph underneath. Without that grounding, the chatbot will hallucinate by design.
- The "AI" being proposed would replace an expert whose tribal knowledge has not yet been captured. Capture first, automate second.
What to propose instead — the hybrid pattern
When the Ambassador refuses a generative-only pilot, the constructive next move is the hybrid. Generative AI at the human-machine surface, with one of the operational shapes underneath where the decision actually gets made. The natural-language layer is a thin wrapper over a strategy that lives in the predictive, computer-vision, or autonomous engine beneath it.
That split is the shape of the Mistral AI plus Prisme.ai plus NVIDIA stack ALTEN has already partnered for, and it is what the Ambassador can offer when the client's first idea is genAI-shaped but the use case isn't.
What this becomes
The framework on this page is the static form. Phase 2 ships it as a Claude or Prisme.ai multi-LLM skill the Ambassador runs pre-call, so the triage happens in minutes and the scorecard is pre-filled before the conversation starts.
- Input contract
- Client mandate (free text), country, vertical, known data assets, and regulatory posture.
- Tools the skill calls
classify_use_case(mandate → shape),check_data_readiness,lookup_regulatory_gate(PIPL / APPI / DPDP / PIPA / light),score_wrong_answer_cost,recommend_deployment_mode(cloud / edge / hybrid), andemit_scorecard_row.- Verification gates
- Cannot return "pilot" or "scale" when data readiness is empty or the regulatory gate is unresolved. Cannot return "generative" as primary shape when wrong-answer cost is above embarrassing and there is no human in the loop.
- Output
- A filled scorecard row with an audit trail of which triage questions scored which way, in the same format every Ambassador uses across every country.
- Rollout
- One country pilots the skill (Korea, given the worked example above), then ports to Japan, India, China, and Vietnam with country-specific regulatory and adoption-pattern overrides. The central AI team in France reads the rolled-up audit trail across all five markets.