Data before AI — why decision quality starts with data discipline.

On 26 February 2025, Gartner published a prediction that has been quoted in most discussions of AI adoption since: through 2026, organisations will abandon 60% of AI projects that are not supported by AI-ready data. A second figure travelled far less. In a survey of 248 data management leaders conducted in the third quarter of 2024, 63% either did not have, or were unsure whether they had, the data management practices that AI requires.

Neither number is about models, vendors or budgets. Both describe the material an AI system works on — a proposition measured for fifteen years, whose practices are already written into Swiss and European law for reasons unconnected to AI.

“AI-ready” describes a judgement about one use, not a state a company reaches

Gartner’s own framing is narrower than the phrase suggests. Data is AI-ready when it is aligned to a specific use case, qualified against what that case requires across its life cycle, and governed so that outputs can be traced and trusted — maintained as a continuing practice rather than completed once.

No general clean-up therefore makes a company AI-ready. Readiness is assessed one decision at a time, which changes the question a management team should ask. “Is our data clean?” has no answer. “Does this decision have the data it requires, and how would we know if that stopped being true?” has one.

The productivity evidence for data discipline predates the current cycle by fifteen years

Erik Brynjolfsson, Lorin Hitt and Heekyung Kim, presenting at ICIS in 2011, examined survey data on business practices and IT investment at 179 large publicly traded firms. Those that had adopted data-driven decision-making showed output and productivity 5–6% higher than their other investments and their information technology use would predict. The association also appeared in asset utilisation, return on equity and market value, and instrumental-variable estimates indicated it was not driven by reverse causality.

A second study reached inside the firm. Working with the US Census Bureau, Brynjolfsson and Kristina McElheran fielded the Management and Organizational Practices Survey in 2010 as a supplement to the Annual Survey of Manufactures, across a representative sample of roughly 50,000 establishments. Greater use of data-driven decision-making was associated with a productivity increase of 3% or more, and the improvement appeared only after plants reported adopting it — consistent with causation running from practice to performance.

What counted as adoption there is a management specification, not a technological one: the top two categories for both the availability and the use of data, ten or more tracked key performance indicators, and both long-term and short-term targets. By 2010 the share of plants meeting that bar had nearly tripled to about 30%, concentrated where complements existed — IT investment, educated staff, multi-unit structure.

Most organisations have never measured the quality of the records they decide on

Tadhg Nagle, Thomas Redman and David Sammon assembled 75 data-quality measurements across organisations and processes over two years, publishing the result in the Harvard Business Review in September 2017. On average, 47% of newly created records contained at least one critical error — one material to the work the record supports. Against the loosest standard the authors applied, 97 or more of 100 records free of such errors, only 3% of the measurements were acceptable.

The method behind those numbers is the useful part, and it is deliberately crude. Take the last 100 records a team produced. List the ten to fifteen attributes that matter for the work. Ask two or three people who use those records to mark any containing an obvious error, and count the records with none. The exercise costs an afternoon and produces a number that can be tracked each quarter.

The relevance to AI is direct. A model reads those records faster than any person and does not pause at an implausible price or a duplicated customer; the error rate does not fall on the way in. Measured, it becomes a figure management can move.

In Switzerland and the EU, the first data inventory is already a legal document

Swiss companies have held the raw material for that work since the revised Federal Act on Data Protection entered into force on 1 September 2023. Article 12 requires controllers and processors to keep a record of processing activities stating, as a minimum, the controller’s identity, the purpose, the categories of data subjects, of personal data and of recipients, the retention period, the security measures and any disclosure abroad. Article 6(5) obliges anyone processing personal data to satisfy themselves that it is accurate, and to correct or delete what is incorrect or incomplete.

Article 24 of the Data Protection Ordinance exempts undertakings with fewer than 250 employees on 1 January of the year, unless they process a large volume of sensitive personal data or carry out high-risk profiling. Most Swiss SMEs are therefore exempt from writing the record — not from needing it. It answers the questions that open any AI project: what do we hold, why, who receives it, how long do we keep it, and does it leave the country.

European law adds a date and a definition. The Data Act, Regulation (EU) 2023/2854, in application since 12 September 2025, obliges providers of data processing services to remove obstacles to switching; Article 29 allows only cost-based switching charges during the transition and prohibits them entirely from 12 January 2027, data egress included. Portability then becomes an option rather than a negotiation — but only for a company that knows what it holds and in what format.

The definition sits in Article 10 of the AI Act, Regulation (EU) 2024/1689: data sets relevant, sufficiently representative, as free of errors and as complete as possible, and appropriate in their statistical properties. It binds high-risk systems only, a category most SME applications fall outside — but as a specification for a firm buying a forecasting tool it is usable today.

The European and Swiss figures point to practice, not to access

Analytical capability is no longer the binding constraint. In Eurostat’s 2025 survey of ICT usage, 33.0% of EU enterprises with ten or more persons employed carried out data analytics with their own staff and a further 13.9% through an external provider — 39.9% in total, against 20.0% using any AI technology. The size gradient is steep: 27.9% of small enterprises analysed data in-house against 78.8% of large ones. Counting both routes, national rates ranged from 60.0% in Denmark to 24.5% in Poland.

The Swiss picture is consistent. UBS, surveying some 2,500 firms with the research institute Intervista and publishing on 19 May 2026, found AI in use at six Swiss companies in ten, but few using it systematically. It attributes the difference to the level of digitalisation: firms with developed IT infrastructure and structured data deploy AI considerably further, and use concentrates in ICT, financial services and pharmaceuticals.

Read together, the two datasets repeat the 2011 finding: firms are separated less by the tools they can buy than by whether the data those tools consume is defined, owned and current.

Six decisions available before the next budget cycle

• Name the three numbers that drive the weekly management decision, and for each the single system in which it is authoritative. Where two disagree, decide which wins.
• Run the measurement on the process that matters most — last 100 records, ten to fifteen attributes, share free of critical errors — and repeat it quarterly.
• Give each critical data entity a named owner — customer, article, price, stock, quotation. Ownership of a field, not of a system.
• Build the record of processing activities even if Article 24 of the Ordinance exempts you. It is the cheapest data map a company will ever produce.
• Set down in writing what may be entered into an external AI system, and who authorises an exception.
• Define freshness per entity — how old a figure may be before it stops being used — and check the export formats of your main systems before 12 January 2027.

Gartner’s prediction window closes at the end of this year. Whether 60% proves accurate matters less than whether a company can say which of its own initiatives stalled, and for want of what — a question that rests on fifteen years of measurement rather than on a forecast, and that a leadership team can open with a hundred records and an afternoon.