Ex-DraftKings Staff Say Its AI Ranked Bettors by Expected Loss, Not by Harm Risk

TL;DR

The New York Times reported on September 19 that DraftKings built a machine-learning model in 2023 that scored customers by how much money each one was likely to lose after receiving a free bet, and that a model to flag customers developing a gambling problem, drawing on the same data platform, was stalled or killed. Six former employees described the loss-targeting work and four described the shelved harm work. You do not have to take their word for the payoff: DraftKings' own 2026 Investor Day deck says AI "Automated and Personalized $400M of promotional spend in 2025" and claims a 1,300 basis-point lift in net revenue margin on promotional wagers.

What the model scored

The 2023 model took three inputs that any sportsbook already has in its warehouse: how often you play, what your account balance looks like, and your loss-to-wager ratio. It emitted one number per customer. Higher score, more expected loss per promotional dollar spent on you.

That is a propensity model with a revenue label on it. Jayden Butts, a former data analyst quoted by the Times, described the assignment in the flattest possible terms: "We are looking for traits and features that we can target that indicate a good investment."

The reason this is a story rather than a marketing-ops footnote is what sits next to it. According to four other former employees, an effort to score the same customers for gambling-harm risk, using the same signals and the same infrastructure, did not ship.

DraftKings put a number on the upside itself

Here is the part that does not depend on anonymous sourcing. On March 2, DraftKings held a virtual investor day, and slide 39 of its own deck is titled "Customer analytics power personalization." Under the heading SMART PROMOTIONS it reads: "Automated and Personalized $400M of promotional spend in 2025 through AI." Under TAILORED RECOMMENDATIONS, tagged LTV, it claims a "+1,300 bps increase in Sportsbook Net Revenue Margin on promotional wagers."

1,300 basis points is 13 percentage points. The deck does not give a baseline or a comparison window for that figure, and it applies to promotional wagers specifically, not to the whole book. Treat it as a company claim, not an audited line item.

The audited line items point the same direction. DraftKings' FY2025 Form 10-K reports Sportsbook Net Revenue Margin of 7.1%, up from 6.0% in 2024 and 5.6% in 2023, and attributes the rise in revenue per payer to "increased Sportsbook hold percentage and improved promotional reinvestment."

Scale check: revenue of $6,054.5M in 2025 on an average 4.0 million monthly unique payers, $125 of revenue per payer, and $1,379.9M of sales and marketing expense. Promotional credits are a separate lever from that marketing line, which is exactly why a model that aims them better shows up in margin rather than in opex.

The two boxes in the same diagram

Slide 37 of the deck draws DraftKings' in-house stack as eight boxes. Two of them are "MACHINE LEARNING & AI" and "RESPONSIBLE GAMING." They sit in the same picture, which makes the Times' reporting easier to picture: the wiring between those two boxes was apparently optional.

DraftKings is not quiet about responsible gaming. The deck claims a #1 industry ranking based on research by Eilers & Krejcik Gaming, plus FY2025 figures of 58 million-plus responsible-gaming center visits, 52%-plus tool usage, and 44 million-plus responsible-gaming messages sent. The 10-K opens its responsible gaming section with "Responsible Gaming is fundamental to DraftKings' mission."

Read the deck's own responsible-gaming timeline, though, and the entries are principles, funding programs, awards, a headquarters center, a budget-builder tool, and, in 2023, the announcement of a first Chief Responsible Gaming Officer. Every item is a policy, a disclosure, or a self-service widget. Not one of them is a predictive model. The predictive model, per the reporting, went to the other side of the flywheel.

Why a builder should care

Strip out the sportsbook and this is the most common architecture decision in applied ML. You assemble a feature store describing user behavior, then you choose a target variable. Expected revenue and expected harm are frequently computable from the identical inputs, because both are downstream of the same escalating engagement pattern.

A smoke detector and a barbecue timer can hang off the same thermometer. The sensor has no opinion about which alarm you wire it to; that choice is made in a product meeting, not in the model.

Three concrete implications if you own a scoring system:

• Label choice is the policy. Once the features exist, "we could not detect that" stops being a technical statement. It becomes a statement about which target you funded.
• Your own marketing is discovery material. DraftKings' investor deck quantified the upside of promo targeting six months before a newspaper described the mechanism. Slides written for shareholders read very differently once a reporter has the internal docs.
• Self-service tools are not a risk model. Budget builders and messaging campaigns put the detection burden on the user. A propensity score does not. Shipping only the former while holding the data for the latter is a defensible product choice right up to the moment someone writes it down.

Caveats worth keeping straight

The internal-model reporting rests on one outlet's investigation, sourced to documents and interviews we have not seen. DraftKings disputes the characterization, saying it "rejects any implication that its marketing practices are unfair or improperly targets customers" and that promotions are "directed toward customers who demonstrate sustained, engaged use of our platform, not toward customers based on their losses."

Also note that "shelved" is not "never existed," and several secondary write-ups of the Times piece garbled DraftKings' own numbers, reporting the 1,300 bps promo-margin lift as a flat "13%" and inflating the deck's 25% chatbot containment figure to 100%. The deck is public and 74 pages long. Read the slides, not the recaps.

Key Takeaways

• Former DraftKings employees say a 2023 model scored customers by expected loss per promotional dollar, using play frequency, account balance, and loss-to-wager ratio.
• Four former employees say a gambling-harm risk model built on the same data platform was stalled or killed.
• DraftKings' own 2026 investor deck says AI automated and personalized $400M of promotional spend in 2025 and claims a 1,300 basis-point margin lift on promotional wagers.
• The audited 10-K shows Sportsbook Net Revenue Margin rising 5.6% to 6.0% to 7.1% across 2023-2025, credited partly to "improved promotional reinvestment."
• Once a feature store exists, whether you score for revenue or for harm is a funding decision, not a modeling constraint.
• DraftKings denies improperly targeting customers; the internal-model claims come from a single investigation and remain contested.

Sources: The New York Times, DraftKings 2026 Investor Day presentation, DraftKings FY2025 Form 10-K (SEC), iGaming Republic, Gadget Review via Yahoo, Techmeme