Veikkaus is buying a retention engine in public, and 75 competitors can read the specification

Veikkaus is procuring a machine-learning system. The procurement notice is public, because Veikkaus is state-owned and its purchasing is subject to public procurement rules.

The notice states the objectives: customer retention, player lifetime value, revenue. The system is to deliver AI-based experiences such as bonuses and other actions based on an individual player's behaviour and situation, so that a message, offer or bonus can be targeted at a player to bring them back or keep them playing.

Harm prevention and responsible gambling are not listed among the objectives.

Virve Marionneau, associate professor and gambling researcher at the University of Helsinki, told Yle that reconciling commercial objectives with harm prevention is the largest open question in the package. She also made the point that makes this difficult: a small share of players produces a large share of revenue, so the commercially valuable player and the player whose gambling should be constrained are frequently the same person.

Marionneau is careful to say it is not known whether Veikkaus intends to use the system in that way. Neither do we.

The part the Finnish market should notice

Every one of the 75 licence applicants can read this document.

Veikkaus is a state-owned company. Its procurement is published. The incumbent's personalisation and retention roadmap, including what it is specifying, what capabilities it expects a vendor to supply and roughly when, is therefore a public artefact in the middle of a competitive market opening.

None of the applicants has any comparable obligation. They are private companies; their technology choices surface only when a supplier announces a deal.

So the information flows one way. The operator with the largest Finnish player database, the strongest brand recognition and a legacy of long sponsorship agreements is also the only one whose build plan can be read by its competitors. Whether that is a disadvantage or simply an unusual feature of this market is a question for the people writing competitive strategy, but it is real and it is asymmetric.

The bonus problem

A personalisation engine built to deliver individually targeted bonuses is being specified into a legal regime that sharply restricts bonuses.

Under the Gambling Act, bonus play money and items such as free scratchcards are permitted within limits. Other bonus offers are prohibited. The Act also requires marketing to be moderate in content, visibility and frequency, and prohibits targeting anyone on the national self-exclusion register.

From 1 July 2027 Veikkaus is a licensee like the others, including for its competitive-market business. Whatever is procured now runs under those rules then.

That leaves a narrow design question with wide consequences, and it applies to every operator building the same capability rather than only to Veikkaus: what does an engine trained to maximise retention actually do when the lever it was built to pull is mostly unavailable? The plausible answer is that it pulls the remaining ones harder, which is message frequency and timing. The rules anticipate that, and the regulator has already shown this year that it enforces marketing rules against individuals and companies that break them.

One model, two uses

The uncomfortable technical fact underneath this is that the model which identifies a player drifting toward higher-risk behaviour and the model which identifies a player worth a retention offer are built from the same signals: volume, frequency, session length, deposit pattern, time of day, chasing behaviour after losses.

In most regulated markets, using data and modelling to detect gambling harm is either a legal requirement or a strong international norm. Finland's Act builds mandatory identification, player-set deposit limits, a company-level transfer limit and a year of transaction history into every licensed account, which is unusually good raw material for exactly this kind of detection.

So the same procurement can produce a harm-detection system or a retention system, depending on what the objective function is set to. That choice is not visible from outside the company, which is Marionneau's second point: risk-detection algorithms built by commercial operators are difficult for anyone external to evaluate.

This is the supervisory problem the Licensing and Supervision Agency inherits in July 2027. Not whether operators use machine learning, which they all will, but whether a supervisor can tell what a given model is optimising for.

What an applicant should take from this

• Specify harm detection as an objective in writing, not as a downstream use of a retention system. The difference will be legible in a procurement file or a vendor contract, and those documents can be requested.
• Assume model governance becomes a supervisory topic. Nothing in the Act currently requires a licensee to explain its models, but Denmark has already added an independent audit function over anti-money laundering controls, and the direction of European supervision is toward evidencing controls rather than asserting them.
• Check what the bonus engine does when bonuses are restricted. If the answer is more messages, that is a marketing compliance exposure rather than a product feature.

What we do not know

• Whether Veikkaus intends to use the system for individually targeted bonus offers after July 2027, or how it will reconcile that with the Act's bonus restrictions.
• Whether harm prevention appears as a requirement elsewhere in the procurement documentation. We have only the stated objectives as reported.
• Which vendor, if any, has been selected.
• Whether the Licensing and Supervision Agency intends to supervise model design at all.

We have asked Veikkaus no questions for this article and have relied on the public reporting cited below. We will update it if the company sets out its position.