Foundation
Clean, owned data. One leading system per data domain, maintained mandatory fields, clear ownership.
// The target picture
Where the journey goes and why not simply to the right
Every AI plan can be placed in two dimensions. How autonomously does the system work? And how reliable is it while doing so? Both dimensions form a quadrant.
Most companies want to move right, towards more autonomy. The most common finding, though, sits in the bottom right: a high ambition for autonomy on a weak foundation. That is the risk zone. This is where the projects are born that never reach operations.
Our target picture is reliable autonomy. Processes that run on their own, and do so reliably enough that nobody has to watch them constantly. The way there leads upwards first, through reliability. Hence our principle: reliability before autonomy.
Reliability
Autonomy
Solid base
Target zoneReliable autonomy
Unused potential
Most common findingRisk zone
// The four levels
Every AI use case rests on four levels
Chatbot, forecast or ordering agent: underneath every use case sit the same four levels. You can skip levels. But not without consequence. The weakest required level limits how reliable the whole thing can become.
Enablement
Data becomes usable. Pipelines, interfaces and data models that systems can access.
Activation
Knowledge becomes retrievable. Documentation, rules and context an AI can understand and use.
Autonomy
Agents act. Tasks run on their own, with control, a log and clear limits.
The level that limits your use case today is what we call the bottleneck. Finding it is the purpose of the production readiness audit.
// The reliability score
How evidence becomes a number
At the end of an audit there is a reliability score from 0 to 100. It says how reliable your use case can become today. This is how it comes about, in four steps.
The matrix
We do not rate the whole company, we rate a matrix. Six retail data domains (product master data, customer, transaction, stock, suppliers, online behaviour) times four levels. That makes 24 cells. Every cell gets a maturity value from 0 to 100.
The required cells
Your use case does not need all 24 cells. A replenishment agent needs the product master and stock, for example, but not the online domain. What gets rated is what the use case really needs.
The bottleneck sets the score
The reliability score is the lowest value of all required cells. Not the average. A chain is as strong as its weakest link. The weakest cell is your bottleneck.
Actual against target
Every degree of automation needs a minimum level of reliability. The distance between your actual score and that target is the gap the action plan closes.
| Intended degree of automation | Target score |
|---|---|
| Assistance (AI suggests, a person decides every action) | 50 |
| Recommendation (AI recommends, a person reviews regularly) | 60 |
| Autonomous with exception approval (AI acts, a person steps in on exceptions) | 65 |
| Fully autonomous (AI acts without ongoing control) | 75 |
Adjustments: high cost of error +10, regulatory requirements +5, low cost of error −5.
Maturity bands
0 to 20Ad hoc
21 to 40Initial
41 to 60Established
61 to 80Managed
81 to 100Optimised
We communicate every value as a band with a confidence note, for example “37 to 47, confidence high”. Never as a falsely precise single number. A measurement based on three weeks of evidence is good enough for an investment decision. It is not made for a decimal place.
// The measurement standard
The rules we rate against
Five rules make sure two assessors with the same evidence reach the same result. They apply to every audit without exception.
Evidenced or capped
Every criterion needs a verifiable artefact. A data quality report, interface documentation, an operations log. Without evidence the rating stays in the initial band, whatever is reported.
Fixed point anchors
Every criterion is rated on a scale from 0 to 4. Every score has a fixed, public description. The cell value is 100 × sum of points ÷ (number of criteria × 4).
No agent points without an agent
The autonomy level is capped at 40 as long as no agent runs in production. Concepts and pilots do not count as operations.
Four-eyes review
Every finding is checked by a second person before it leaves the house. Differences between assessors are documented and resolved.
No value without a band and confidence
Every value we communicate names the maturity band and the confidence (high, medium, low). Confidence depends on the quality and completeness of the evidence.
We publish the full measurement standard with all criteria and point anchors as a document. So you can recalculate every number of a finding yourself.
Download the standard V1 (PDF)
// Terms
Six terms that appear in every finding
- Production readiness
- The ability of your data and system landscape to run an AI use case reliably in daily business. Not in a test, but in operations.
- Reliability score
- An evidence-based number from 0 to 100. It says how reliable your use case can become today. Set by the weakest required cell.
- Bottleneck
- The one cell in the matrix that limits your use case today. The bottleneck shows where the next investment belongs.
- Data domain
- A coherent data area of retail. We distinguish six: product master data, customer, transaction, stock, suppliers, online behaviour.
- Maturity band
- One of five sections of the scale from 0 to 100: ad hoc, initial, established, managed, optimised. We always communicate values as a band, never as a single number.
- Confidence
- How solid the evidence behind a rating is: high, medium or low. It is stated with every value.