From Pilot to Production
A six-step method to move supply chain AI out of the demo and into the decisions planners make every day: start from the decisions that carry value, prove each use case in live operations, and let automation grow only as fast as trust is earned.
Built for planning and operations leaders whose AI pilots never reached daily use.
- 6 stepsfrom decision map to scale
- 8–10 weeksto a prioritized AI roadmap
- 12 weeksper pilot wave, in live operations
Most AI pilots work. Few reach production.
The model performs in the demo, then stalls in the handover to a planning team that runs on exceptions, conflicting targets and spreadsheets nobody switches off. The cause is rarely the algorithm. It is a use case chosen for its technology rather than its value, data that was never fit for the decision, and no agreement on who decides when the planner and the model disagree. Our method deals with all three.
It is how our founder brought AI into planning at a global consumer goods multinational, where more than 40% of routine planning decisions now run without manual intervention. We describe the failure pattern in Why AI Pilots Die on the Shop Floor. In our strategy method, AI is usually one program of the masterplan; this method is how that program is delivered.
- 01
Decisions
Map the recurring decisions and the value at stake.
- 02
Portfolio
Score use cases and choose where to start.
- 03
Data
Make the data each decision needs fit for use.
- 04
Prove
Run pilots in live operations, against a baseline.
- 05
People
Agree decision rights, overrides and new roles.
- 06
Scale
Roll out in waves, monitor models, track value.
Automation is earned, one rung at a time
Every use case has a target rung on the ladder. It climbs only when its accuracy and its override rate show that people can trust it.
Insight
The model shows what is happening and why. People decide and act.
Recommend
The model proposes a decision. People accept, adjust or reject it, and every override is logged.
Act with approval
The model prepares the action. A person approves it, often in bulk.
Act within guardrails
The model acts on its own inside agreed limits and escalates the exceptions.
Decisions that touch product quality or safety may never climb beyond the second rung, and that is by design.
What happens at each step
Decisions first
We start from the decisions, not the technology. With planning, operations and finance we list the recurring decisions in the supply chain, such as forecasts, replenishment, allocation, scheduling and supplier orders, and estimate the value at stake in each: service, cost, cash and waste. The question “what can AI do for us?” becomes “which decisions cost us most when they are late or wrong?”
Key questionsWhich decisions are made most often? Where does a poor decision cost most?
- A decision inventory with owners, frequency and value at stake
- A performance baseline for each decision
Use-case portfolio
Each candidate use case is scored on value, feasibility, data readiness and risk. We start with three to five that can show results within a quarter, and set for each the rung of the autonomy ladder it should reach.
Key questionsWhere can AI prove value fastest? How much autonomy is realistic?
- A scored portfolio of use cases
- First-wave use cases with business case and target rung
- A build, buy or partner choice for each
Minimum viable data
Instead of a multi-year data program, we make the data fit for the chosen decisions: one shared data model for products, locations and partners, quality rules at the source, and a refresh rate matched to the decision. Most tactical decisions need reliable data every few hours, not every second.
Key questionsWhat data does each decision really need? What can we trust today?
- Data requirements for each use case
- A shared data model and quality rules
- Named data owners
Prove it live
Pilots run inside real planning work, with the planners who will use them, and against a measured baseline. We usually start with predictive models (what is likely to happen) and add prescriptive logic (what to do about it) once the predictions are trusted.
Key questionsDoes it beat the current way of working? Will the team use it?
- Working pilots in live operations
- Results against the baseline
- A go, adjust or stop decision for each use case
People and machine
We agree who decides when the planner and the model disagree, log every override and review the patterns, and redesign roles so planners spend their time on exceptions and scenarios rather than routine adjustments. Planners and data specialists work as one team, and AI is presented as a way to grow capabilities, not to cut headcount, because that is what makes adoption last.
Key questionsWho owns each decision now? What does the planner’s job become?
- Decision rights for each use case
- Override and escalation rules
- Role profiles and a training plan
Scale and govern
Proven use cases roll out in waves across sites and product lines. Models are monitored for accuracy and drift, value is tracked monthly in the existing planning review, and funding is released in stages as results come in, rather than through one large business case up front.
Key questionsWhat scales next? How do we keep the models reliable over time?
- A roll-out roadmap in waves
- Model monitoring with clear ownership
- A value scoreboard reviewed monthly
A mid-size medical devices manufacturer
A European manufacturer of surgical instruments and single-use devices, with about €300 million in revenue, two plants and an external sterilization partner. It sells to hospitals directly and through distributors, keeps consignment stock in operating rooms, and manages about 12,000 products with shelf lives and lot traceability.
Service to hospitals is protected with high inventory, and expired stock is written off every year. An earlier AI forecasting pilot performed well in testing but was never used by the planning team.
In a ten-week assessment, the team mapped its decisions, chose four use cases and agreed the rules for people and models. The first pilot wave ran over the following quarter. The exhibits below show how the method takes shape.
An illustrative example for a typical mid-size medical devices manufacturer. Names, figures and timings are indicative and do not describe any specific client.
Where the decisions are
| Decision | How often | Value at stake | Data readiness today |
|---|---|---|---|
| Hospital demand forecast, by product and account | Weekly | ||
| Consignment stock levels in hospitals | Weekly | ||
| Distribution center replenishment | Daily | ||
| Sterilization slot planning | Daily | ||
| Production scheduling | Daily | ||
| Expiry risk and stock rotation | Weekly | ||
| Supplier order timing for critical components | Weekly | ||
| Tender volume estimates | Monthly |
Choosing where to start
The autonomy ladder, today and at month 18
First pilot wave: results against the baseline
| Use case | Measure | Baseline | After 12 weeks |
|---|---|---|---|
| Hospital demand forecast | Forecast error (weighted MAPE) | 38% | 29% |
| Expiry risk and stock rotation | Stock expiring before use, % of inventory value per year | 2.4% | 1.6% |
| Distribution center replenishment | Orders released without manual change | 15% | 62% |
| Sterilization slot planning | Booked sterilization slots actually used | 81% | 90% |
An 18-month roadmap in three waves
One scoreboard, reviewed monthly
| Layer | Measures |
|---|---|
| Adoption | Share of decisions taken without manual change; planner override rate and the reasons behind it |
| Model quality | Forecast error and bias; accuracy drift for each model; data quality at the source |
| Business value | Service level to hospitals; inventory and cash; expired stock written off; sterilization capacity used |
| Capability | Use cases on each rung of the autonomy ladder; planners trained; time spent on exceptions instead of routine work |
Four deliverables, from roadmap to running AI
The assessment and roadmap take 8 to 10 weeks; each pilot wave then runs about 12 weeks in live operations.
AI roadmap and use-case portfolio
The decision inventory, scored use cases, target rungs on the autonomy ladder and a roadmap in waves, with business cases.
Data fit for the first wave
A shared data model, quality rules and named owners for the data the first use cases need.
Working pilots with results
Pilots running in live operations, measured against the baseline, with a go, adjust or stop decision for each.
Operating model for people and machines
Decision rights, override rules, role profiles, training and the monthly scoreboard that keeps AI in production.
Is your supply chain AI in production, or still in a pilot?
Connect directly with our team to discuss which decisions would pay back first. Or begin with the three-minute Supply Chain & AI Readiness Assessment.