What a factory risk report looks like — and what the numbers mean
“AI factory risk score” is the kind of phrase that means nothing until someone shows you the working. So here is the working: what goes into a factory report, what each number is actually measuring, and — just as importantly — what a score cannot tell you.
What the score is trying to answer
One question: how likely is this factory to fail to deliver what was ordered, on the terms that were agreed?
It is not a quality rating. It is not a rating of the company’s finances. It is a delivery-risk score, built from what a factory has actually done across transactions rather than from what its company profile says about itself.
The four numbers, and what each one really tells you
Order completion rate
The share of orders the factory completed as agreed. A figure like 98% sounds unremarkable until you translate it: two orders in a hundred did not complete as agreed. On a $200,000 order, “as agreed” is doing a lot of work in that sentence.
The number to be sceptical of is 100%. A factory with genuine volume and no incomplete orders at all is usually a factory with a short history in the data, not a perfect one.
Repeat-purchase rate
How many times the average buyer comes back. This is the most under-rated number on the report, because it is the hardest to game. A buyer who has already been burned does not place a second order, and definitely not a fourth.
A figure like 4.2× means the typical buyer has come back three more times after the first order. That is a stronger signal than any certificate, because it is a revealed preference by people who had money at stake.
Disputes
Recorded disputes across the network. Zero is good. What matters more than the count is the pattern: a factory with a handful of disputes that were resolved is often a safer bet than one with none recorded and very little history.
When we see disputes, the useful question is what they were about. Specification disagreements point at communication and drawings. Delivery disputes point at capacity. Quality disputes point at process control. They are not the same risk.
Transactions analysed
The sample size — and the number that should govern how much weight you give the other three. A score built on 1,240 transactions is a different object to a score built on 12. A high score on thin data is not a high score; it is an absence of evidence.
If a report does not tell you the sample size, that is a problem with the report.
A worked example
Take a report showing 82/100, low risk, 98% completion, 4.2× repeat purchase, 0 disputes, 1,240 transactions analysed.
Read it as: this factory has substantial history in the data, buyers come back repeatedly, almost everything completes, and nothing has escalated to a dispute. That is a good factory to be dealing with.
It is not a statement that your order will be fine. It is a statement about a base rate. The 2% that did not complete as agreed were also placed by buyers looking at a good score.
What a score cannot do
This part matters more than the rest of the article.
- It cannot see your specification. A factory with an excellent record on standard product may have never made the thing you are ordering. Scores are historical; your order may be novel to them.
- It cannot see your drawings. A large share of “quality failures” in imported building materials are specification failures — the factory built exactly what it was told to build.
- It cannot predict a change of circumstance. Ownership changes, key staff leave, a plant relocates, a raw material supplier changes. History is a base rate, not a forecast.
- It is not a substitute for inspection. A score tells you how much you should worry. An inspection tells you what is in the container.
This is why we treat scoring and inspection as two different tools rather than alternatives. The score informs the decision to proceed and on what terms. The inspection checks the specific goods.
How the data gets there
The inputs are the transaction records across the network — orders placed, whether they completed as agreed, whether the buyer returned, delays against the agreed date, and whether anything escalated to a dispute. Where a factory’s performance deteriorates seriously and is not resolved, it can be suspended or removed from the network.
Two limits worth stating plainly. First, the data is about performance within this network — a factory may have a long and excellent history elsewhere that is not visible here. Second, scores are recalculated as new transactions land, so a report is a snapshot with a date on it, not a permanent rating.
How to use a report properly
- Check sample size first. Everything else is conditional on it.
- Read repeat purchase before completion rate. It is harder to game and it reflects buyer judgement, not just process.
- Treat the score as a prior, not a verdict. It should change how much verification you do, not whether you do any.
- Match the score to the order size. A 68 on a $30,000 order and a 68 on a $300,000 order are different decisions.
- Ask what changed. If a score moved, the reason is more informative than the number.
Why we do it at all
Straightforwardly: we are the ones settling the invoice with the factory, so we look at the factory before we do. The scoring exists for our own decision. We give clients the same view of it that we have, because a buyer making a decision with the same information is a better outcome for everyone than a buyer making one without it.
Factory assessment and risk scoring are advisory services provided for information only. They are not a warranty of the goods, a guarantee of factory performance, or a condition of your obligations to Linkwox. Purchasing decisions remain yours. Figures used in the example above are from a sample report and are illustrative.
