Data science pays

Part 1 of 4 in our series on data science, the mid-market and agentic AI progress for data science.

Most of the AI in business today stops at analytics: it reports what happened. Data science starts there. It predicts what happens next, explains what causes it, prescribes which decision is best, and then puts that decision into operation. That is where the money is. Firms at the frontier of data-driven decision making run at 4–8% higher revenue productivity (Brynjolfsson & McElheran, 2019).

This post looks at where data science comes from, why the people who do it well are scarce, and what it has measurably been worth to the companies that use it.

Where data science comes from

The term was coined in 2008 by practitioners at Facebook and LinkedIn (Patil, 2011). Data science sits at the intersection of statistics, programming and business expertise (Conway, 2010). In practice it blends with the neighbouring decision disciplines: machine learning, econometrics and operations research. Its answers come from experiments and continuous measurement. That is the "science" in the name.

Why senior data scientists are scarce

Experienced data scientists are highly qualified and hard to find. One of the explai founders (Dirk) sponsored the applied science job function company-wide while at Zalando. It was the highest-paying job function there, with a 20% salary premium over software engineering, the second highest.

The job market took off with a report projecting a global shortage (McKinsey Global Institute, 2011) and Harvard Business Review calling data scientist the "sexiest job of the 21st century" (Davenport & Patil, 2012). The pandemic accelerated it, when businesses had to fast-forward their digital transformation (McKinsey & Company, 2020).

Between 2022 and 2025 the market overcorrected, with a wave of entry-level bootcamps (Raine, 2023; Handshake, 2025). The result is today's skill gap: an oversupply of entry-level candidates with superficial maths, engineering and business skills, and even stronger demand for seniors, because experienced data scientists are well suited to evaluate and safeguard AI systems (Lightcast, 2025).

What data science is worth

For large companies the premium paid off, across industries and functions. The list includes only field experiments and peer-reviewed studies.

Pricing and revenue

• Price optimisation: online flash-sale fashion, Rue La La, US. +9.7% revenue on the styles treated (Ferreira et al., 2016).
• Personalised pricing: B2B SaaS, ZipRecruiter, US. +19% profit over the optimal uniform price. The +86% also quoted from this paper is measured against the status-quo price, and is mostly the value of discovering that the product was under-priced (Dubé & Misra, 2023).
• Revenue management on a forecast: ferry operator Molslinjen, Denmark. $2.6–3.2M saved a year ($5M cumulative to December 2023), fewer delayed departures and 3% less fuel and emissions, from machine-learning forecasts of passengers and vehicles up to a year ahead, in production since 2020 (Pinson et al., 2025).

Assortment and allocation

• Inventory allocation: fast fashion, Zara, Spain, across roughly half the store network. +3–4% sales, worth $233M in revenue and $28M in net income in 2007 (Caro & Gallien, 2010; Caro et al., 2010).
• Assortment: auto-parts retail, US. +5.8% and +3.6% sales in the two categories where the recommendations were implemented, against typical annual comparable-store growth (Fisher & Vaidyanathan, 2014).
• Demand forecasting: retail, Walmart data, 42,840 series. The top machine-learning method was 22.4% more accurate than the best statistical benchmark, and 3% more accurate at SKU-store level. Only 7.5% of 2,666 teams beat that benchmark (Makridakis et al., 2022).

Customers

• Retention targeting: telecom, a membership organisation and European digital TV, randomised. Targeting the customers who respond to a campaign, rather than those most likely to leave, cuts churn by a further 4.1 and 8.7 percentage points for the same spend (Ascarza, 2018). Profit-based targeting is worth about 4% of firm profit from a single campaign (Lemmens & Gupta, 2020). Targeting the highest-risk customers instead can backfire: at one carrier, that campaign raised churn from 6.4% to 10.0% (Ascarza et al., 2016).
• Conversion and trial design: SaaS, 337,724 users. +6.8% subscriptions against a 30-day trial for everyone. A uniform 7-day trial alone gave +5.6%, and personalising the trial with a causal forest did worse than that simple rule (Yoganarasimhan et al., 2023).
• Recommendations: streaming, Netflix. Personalisation and recommendations are worth over $1bn a year and several points of churn (Gomez-Uribe & Hunt, 2015). Even on the strictest causal reading, where at least 75% of recommendation-driven activity at Amazon would have happened anyway, this is worth hundreds of millions at that scale (Sharma et al., 2015).

Risk, fraud and money owed

• Credit scoring: e-commerce credit, Germany, 270,399 purchases. Adding a digital footprint to the credit bureau score raises the AUC from 68.3% to 73.6%. That is the honest size of a machine-learning gain in credit risk. The 10–25% found in consulting decks has no primary source (Berg et al., 2020).
• Fraud and anti-money-laundering triage: Spar Nord bank, Denmark. 33% fewer false positives while keeping 98.8% of true positives (Jensen & Iosifidis, 2023).
• Collections and recovery: a medium-sized Dutch collection agency, 7,839 delinquent borrowers. Repayment rose from 43.1% to 53.2% (+23.4% relative) with fewer calls, when an algorithm chose whom to call instead of the collection officers (Wang & Zhou, 2024).

People

• Labour planning: a US specialty retailer, 168 stores, six months. +4.5% revenue and +$7.4M annual profit after labour cost. The stores had been under-staffed: the analysis found money that the finance function had cut (Fisher et al., 2021).
• Hiring: 15 US service firms, around 300,000 low-skill hires. Job testing increases completed tenure by just over 25%, and managers who hire against the test's recommendation end up with worse hires (Hoffman et al., 2018). Combining the same candidate data by formula rather than by judgement predicts more than 50% better (Grove et al., 2000; Kuncel et al., 2013).

Operations and assets

• Routing and network planning: rail, Deutsche Bahn. €74M saved a year and 34,000 tonnes of CO₂ from optimised rolling-stock rotation, compared with manual planning (Borndörfer et al., 2021).
• Disruption recovery: airline, Swiss / Lufthansa Group. €12M and 14,000 tonnes of CO₂ saved in daily use, with €30M a year projected group-wide (Davies et al., 2026).
• Process control: Tata Steel, India, continuous annealing, live since January 2023. Output in the premium quality band rose from 30% to 50%, with 8% less fuel per tonne, 13,000 tonnes of reprocessing avoided a year, and savings of $2.5M and 10,000 tonnes of CO₂ (Jagnade et al., 2025).
• Predictive maintenance: offshore wind, UK. Up to 8% lower operations and maintenance cost and 11% less lost production (Turnbull & Carroll, 2021).

Mid-sized firms are rare in this list. Molslinjen is one, and it took an outside partner for its gains to be measured at all.

From project to capability

Every result above measures the first good model against a manual or naive baseline. Treating data science as a capability rather than a one-off project keeps the gains coming. Dirk saw this first-hand at Zalando:

• Pricing: 12 markets. About 6% higher profit at equal sales and revenue, across 23 A/B tests in the 2023–24 sale campaigns (Birr et al., 2026).
• Ranking: a transformer model against the previous production ranker. +4.04% engagement and +0.86% revenue on browse, and +0.70% and +0.17% on search (Celikik et al., 2024).

These are improvements on top of models that were already good, in a company that had already invested in data science for years.

Next in this series

The 4–8% productivity gap is not limited to the companies in this list. Earlier cross-sector estimates land in the same band (Brynjolfsson et al., 2011). Yet most of the economy does not capture it. In part 2 we look at why the mid-market misses out, and what that costs.