US$3.1bn on AI data. 63% unready.

Money is moving to the data layer of AI. Gartner's September 2026 spending forecast puts spending on AI data at US$3.1bn in 2026, up from US$0.8bn in 2025, and at US$6.5bn in 2027. Spending on AI platforms for data science and machine learning grows from US$19.4bn in 2025 to US$26.4bn in 2026 and US$35.6bn in 2027. Data teams are being asked to feed AI, not just to report on the business.

Most organisations are not ready for that job. That is the finding a data leader should start from.

Most data is not AI-ready

Gartner put the gap in two numbers in its February 2025 release on AI-ready data. 63% of organisations either do not have, or are unsure whether they have, the right data management practices for AI. And Gartner predicts that through 2026, organisations will abandon 60% of AI projects that are not supported by AI-ready data.

60%. The share of AI projects without AI-ready data that Gartner expects organisations to abandon through 2026.

"AI-ready" is not the same as "clean". A data warehouse built for monthly reporting can be accurate and still unfit for an AI system that needs current, well-described, permissioned data at the moment of a decision. Definitions that live in the heads of analysts, metrics that differ between departments, and access rules that were never written down all break when a model or an agent starts asking questions.

The Global Market Study 2026/27 names the business problem of this area as data readiness and trust in numbers. What helps, it says, is a data-readiness check and a semantic layer before scaling: one agreed definition per business metric, documented where both people and machines can read it.

Wide use, thin returns

The return question sits on top of the readiness question. Use of AI is broad: in official EU statistics, 20% of enterprises with ten or more staff used AI in 2025 (Eurostat, as reported in ETHORITY's AI growth facts). But in McKinsey's August 2026 survey, only 37% of organisations attribute any EBIT effect to AI.

Analytics is where that gap becomes visible, because data teams are usually the ones asked to measure it. An AI forecasting feature that lowers forecast error but is never used in a planning decision has produced a better number, not a better result. The ETHORITY evidence synthesis Saved hours are not AI value separates three levels that are often merged: task improvement, released capacity and a business outcome. Only the third pays.

37% of surveyed organisations attribute any EBIT effect to AI (McKinsey, Aug 2026).

The data under the models

Data & Analytics buyers face a second data question: what the models they use were trained on. The Good AI Prism keeps registry cards for twelve widely used public datasets, read from their cards and licences, and they show how varied the terms are.

Exhibit: licence terms on twelve profiled datasets

Source: Good AI Prism Registry, dataset cards; selection of 8 of 12.

Across the twelve cards, at least seven different licence regimes appear, from public-domain dedication to research-only use. The ImageNet registry card records "research only". The The Stack v2 card records licences per file, which means compliance depends on which files a model actually used. For an analytics team building its own models or fine-tuning one, these terms decide what may go into production.

What the Prism shows about analytics offerings

The Good AI Prism places 22 AI offerings and one provider record (Qlik) in Data & Analytics, from AI features in business-intelligence platforms to forecasting, data preparation and natural-language query tools. Each is mapped from the provider's own pages, and the decision facts a buyer should collect are the same as in the Knowledge & Search topic area: pricing unit, hosting, training terms, integrations and certifications, each with source and date.

Three questions from the profiled cohorts transfer directly to analytics tools. Where is the data processed, especially when a natural-language query sends table contents to a hosted model? Is customer data used to train the provider's models? And which billing unit applies: seats, queries, compute or credits? In the twelve Software Development cards, for example, at least five billing units appear. Analytics tools that charge per query or per compute unit make volume the main cost driver.

Measure decisions, not dashboards

The study recommends two outcome measures for this area: decision latency and forecast error. Both tie analytics to a business result. Decision latency asks how long it takes from a question to an action. Forecast error asks whether the numbers that drive planning got better. A new dashboard, a new model or a new chat interface on top of the warehouse is only progress if one of the two moves.

Use case. A demand forecast. A retailer adds an AI forecasting feature to its planning tool. The team records today's forecast error per product group and the days between a forecast and a purchasing decision. After three months it compares both, and it checks how often planners actually used the AI forecast. If error fell but planners overrode the forecast, the next investment is in trust and process, not in a better model. If both measures improved, the case for scale is documented.

Usage-based pricing raises the stakes. The Global Market Study cites Gartner, via Deloitte, expecting at least 40% of SaaS spend to be usage-, agent- or outcome-based by 2030. Natural-language analytics invites more questions. More questions mean more compute. Without a cost per decision supported, a successful rollout can look like a budget problem.

What to do Monday

1. Pick the five metrics your executives use most. Check whether each has one written definition, one owner and one source. Where it does not, fix that before connecting any AI tool.
2. Run an AI-readiness check on one data domain. Test currency, documentation, access rules and lineage for the data a planned AI use case will read.
3. Record the training-data terms of the models you use. Note licence and restrictions, especially where you fine-tune or build your own.
4. Ask analytics suppliers three questions in writing. Where is query data processed? Is it used for training? Which billing unit applies at your volume?
5. Set two outcome measures before the pilot. Baseline decision latency and forecast error for one planning process, and decide in advance what improvement justifies scale. The AI Readiness & Maturity Check starts exactly there.

The Good AI Prism keeps every registry fact sourced and dated, so that a data strategy can start from documented terms rather than from a vendor demo.