CPQ Prediction Server

What a Prediction Server actually does

Every manufacturer with a CPQ system already has the raw material: years of quotes, each a record of what a real customer chose, and most with a known ending. Won or lost. What almost nobody does is learn from it systematically.

The Prediction Server does that one thing. It trains a model on your historical configurations and then, as a new quote is built in Tacton CPQ, predicts the choices a similar customer made before. The seller sees a suggestion next to each open question. Change the country, the size or an early option, and the suggestions move.

That is the whole product. It is not a new CPQ, not a pricing engine, and not a replacement for a salesperson's judgement. It is a second opinion on every choice, drawn from everyone who configured something similar before.

What 8,000 real quotes taught us

We did not want to sell this on a slide, so we tested it. In 2024 two KTH students, Mats Desaulty and Edvin Fasth, ran a degree project with us on 8,000 quotes generated in real use by a lift manufacturer's CPQ system: customer, installation country, and every technical and aesthetic choice in the configuration.

What came out:

The model beats the obvious guess, clearly. A nearest-neighbour model predicted the customer's actual choice 83 percent of the time on average. A language model reached 84 percent. Always suggesting the most common option, the naive baseline, scored 68 percent. Per feature the model ranged from 72 percent to a perfect 100 percent.

It holds up where it matters. The interesting result was not the easy features. It was that accuracy stayed high for features with eight or eleven options and no dominant favourite, exactly where a human seller has the least intuition and the customer needs the most help.

The data was the problem, not the maths. Roughly 60 percent of the rows were duplicates. The same installation country appeared under several different names. Every feature had to be normalised before a model could learn anything. This is the part of the project that took the time.

Predicting orders needs volume. Trained only on the quotes that actually became orders, the model had far less to learn from and some predictions fell sharply, one feature from 78 percent to 32 percent. The lesson is not that order prediction fails. It is that it needs a bigger pile of orders than one lift manufacturer produces in a few years.

We chose the simpler nearest-neighbour model for implementation. It was within a point of the language model, and a model your own team can understand and retrain is worth more than one that is marginally cleverer.

The full study is the degree project "Recommendation Models in Product Configuration", KTH, 2024.

Three ways to use the prediction

The study also produced a framework we now use to decide what a Prediction Server should do for a given manufacturer. The same model, three different jobs.

Social proof. Show the customer what others configuring a similar product chose. Fastest to deploy, and right for products with many design options and few technical constraints between them. People trust a choice more when they know it is a common one.

Guidance to the right product. Train the model on the configurations that worked and let it steer towards them. This also lets you replace technical questions with questions the customer can actually answer. A haulier does not know which crane hydraulics they need. They know exactly what they are going to lift, and how often.

Priming for orders. Train on the configurations that went to order and nudge every new quote towards the choices that win. This is where the money is, since even a small lift in hit rate lands directly on revenue. It is also the one that needs the most data, for the reason above.

Most manufacturers start with the first and grow into the third as the order history accumulates.

What it needs from you

The model is only as good as the quote data behind it. Three things decide whether a Prediction Server works or disappoints, and all three showed up in the lift study.

Consistent naming. If the same option has been recorded five different ways over the years, the model sees five different things. This is the same discipline we describe on our CPQ integration page: inspect the data close to the source, before it reaches the model.

Outcomes recorded. For the order-priming use, every historical quote needs a known result. If lost quotes were never closed out in the system, the model cannot learn what losing looks like. Sorting this out is usually the first job.

Retraining. Products, prices and markets move. A model trained once and left alone drifts within a year. We retrain quarterly and compare predicted against actual, so you can see the accuracy rather than take it on faith. Retraining is scoped into the project from the start, not bolted on later.

What it will not do

It will not fix a bad sales process. If quotes are wrong because the product rules are wrong, fix the configurator first. The nine mistakes we keep seeing in CPQ projects apply here in full.

It will not conjure order predictions from thin data. A few hundred orders a year across many product families is not enough to learn what wins. Recommendations, yes. Close probabilities, not yet.

It will not replace the conversation with the customer. A suggestion is a prompt to ask why, not a substitute for asking.

How we deliver it

The Prediction Server is part of our business intelligence work for Tacton CPQ customers, and we deliver it as a consulting project rather than a boxed product. Every manufacturer's quote data looks different, so the model is built on yours. We have implemented Tacton since the platform existed and run analysis workshops since 2001, so the project follows the same shape as our other Tacton CPQ implementations: a short workshop to agree scope, data sources and which of the three uses to start with, a fixed price once the scope is clear, and a target of going live within five months. For how we price CPQ projects in general, see our pricing page.

Most of the effort goes into the data, not the model. The lift study is the honest guide here: expect the first weeks to be about duplicates and naming, not algorithms. Once the data is clean, the model is the quick part.

The suggestions live inside Tacton CPQ, so sellers see them where they already work. The reporting side lives in the same BI layer as our CPQ analytics work, because the Prediction Server is one output of that layer, not a system of its own.

Ready to see what your quote history knows?

Book a meeting and bring two numbers: quotes per year, and how many of them have a recorded outcome. That is enough for us to tell you which of the three uses your data can support today.

Frequently asked questions