The proof of concept
I score first. Then you show me the sales.
A back-test on files you already have, at no cost and no obligation. It ends with a number you measured yourself and a decision you make on your own figures.
Seven steps. Mutual confidentiality, your history, a model built and locked, a blind test, the result, the value, the decision. About nine weeks from the first call to a live test file. The typical run is a US direct response marketer with a few years of mail history, and the pattern has held on files the model had never seen.
Seven steps
How it works
- 01Mutual confidentialityOne call and a short agreement. Scope, mutual NDA, a term that says the scores are for judging the test and not for use until we have an engagement, and a plain statement that nothing in it is a price quote. You pay nothing and owe nothing.
- 02You share the historyThe files you already have. What you mailed or contacted in past years and the orders it produced, with a field layout. That's all. Files travel encrypted to your own key through a share you control, and no name ever appears on a shared screen.
- 03I build the modelOn the older years only, from your own buyers, strengthened with data licensed for your program. Within your own file the same person mailed under five spellings of one address becomes one person. The result is a ten-band score built on revenue per person, not response alone. Factors and cut points are locked and written down before I see the test year.
- 04I test it blindYou send the newest file. I score it cold on the locked model, keyed to your record number, and return it with a fingerprint on the record. Only then does the sale file move. We both tie orders to pieces, exact matches only, and compare. Nothing can be tuned after the sales arrive, and the sequence proves it.
- 05You see the resultThe ranking has to survive re-runs under changed rules, on people the model never saw, inside single drops, with the largest orders capped. Aggregate tables with no names on them go to two separate AI models asked to break it. Every figure arrives with the table behind it. Where a difference is noise, I say so.
- 06You put a value on itProfit per thousand at every depth, at your own cost per piece and your own customer value. I rank. You set the cut line, because where to stop is your economics, not mine.
- 07You decideIf the score ranks your buyers on a file it never saw, we talk about an engagement. If it doesn't, you'll know, with the numbers, and so will I. Either way, if you don't go forward every copy is deleted, aggregated or not, and you get that in writing.
Nine weeks
The timeline
| When | What |
|---|---|
| Weeks 1 and 2 | Call, agreement signed, encrypted channel set up, history received, first audit findings back within the week. |
| Weeks 3 and 4 | Model built, locked and written down. Build years returned scored. Test file scored blind, fingerprinted, delivered. |
| Weeks 5 and 6 | Sale file received. Orders tied to pieces. Re-runs and reviews. Findings walkthrough, every number traced to its table. |
| Weeks 7 to 9 | Value at your own figures. Written evidence package, what was promised on the first call beside what was delivered. Decision. Live test file scored and returned. |
The lines
What I won't do
- Tune anything after the sales arrive. The fingerprints and the dated rules make that checkable.
- Fuzzy match an order to a piece, or let a result ride on a loose match.
- Compare your customer file against your prospect file. They are different populations and are read within themselves.
- Suppress a record for lack of a score. Unscored records are their own decision group with their own economics.
- Treat absence as proof. A record that doesn't appear in the licensed data is a review item, never a hard exclusion.
- Publish the recipe. The variables and segments stay mine. Under NDA you see every data source and the license terms that cover your program.
- Keep your data. If you don't go forward, every copy is deleted and you get that in writing.
After the decision
Why clients stay
The proof of concept is the engagement in miniature. The same loop runs every cycle afterward. Build, lock, test blind, read, decide, on the next drop, the next channel, the next year's file. Files change and models drift, and a score that was checked once is not a score you can trust twice. That's why it takes a few weeks to start, and why the clients who start tend to stay. The process is the product.
It also sets how mistakes are handled, because there will be one eventually. When I find it, and I have, the correction goes to you in writing with new fingerprints and a plain account of what changed. Nothing is edited in place.
Start one
Email me with the type of business you run, the mail or campaign history you have, and a rough sense of annual volume. I'll reply with the short agreement and we'll set the first call.
ben@kristalytics.comMQ&K ADV INC dba Kristalytics