Your semantic layer is a cost control

Ask three analysts in the same company for last quarter’s active customers and you can easily get three numbers. People notice, argue about it in a meeting and settle on one. Ask an AI agent the same question and it picks one of the three and gives you the answer in a confident voice.

In May 2026 Gartner attached a figure to that problem. At its London summit it said that organisations which prioritise semantics in AI-ready data could see agent accuracy improve by up to 80% and costs fall by up to 60% by 2027.¹

How much weight the number deserves

Not a lot, on its own. Both figures carry "up to", the release gives no sample, and they will be in vendor decks by Christmas. I read them as ceilings.

The mechanism is more convincing than the percentages. Rita Sallam described context with semantic coherence as a cost-control and trust strategy, and the release recommends a context layer as core infrastructure, because schema-based data models alone fall short for agents.¹ Anyone who has watched a BI estate grow three versions of "revenue" will recognise the problem.

Where the money actually goes

None of this shows up on the model invoice, which is why finance reviews of AI spending tend to miss it. My own breakdown, which is not Gartner’s, looks like this. Analysts spend time checking numbers before they reach a board pack, and inconsistent definitions slow every check. People re-run questions when an answer looks wrong, and each run costs tokens and patience. Teams paste business rules into prompts to compensate, and those copies drift apart with nobody owning them. And staff quietly stop using a tool after two bad answers while the licence carries on.

A shared definition attacks all four at once, which is why I would treat the semantic layer as a cost line in its own right.

Start with ten terms

A semantic layer sounds like a platform program. The first version is a list. Take the ten business terms that come up most in executive questions (revenue, active customer, churn, margin and so on) and write down four things for each.

1. The definition in plain words, signed off by the business owner.
2. The table or metric that implements it.
3. Who is allowed to change it.
4. The known exceptions, such as a region that reports on a different calendar.

Put the list where people and agents can both read it, and make the agents use it. After a month you will know which terms cause the most arguments, and those are the ones to model properly first. Gartner expects universal semantic layers to be treated as critical infrastructure, next to data platforms and cybersecurity, by 2030.² A list of ten terms is how that infrastructure starts in practice.

Portability is now something you can test

A year ago, asking a vendor about open semantic standards got you a promise. Snowflake announced the Open Semantic Interchange in September 2025 with partners including Salesforce, dbt Labs, ThoughtSpot, Alation and Atlan.³ The first version of the specification was published in January 2026 under an Apache 2 licence, and Databricks, Qlik and AtScale joined.⁴ By April 2026 the work was in incubation at the Apache Software Foundation under the name Ossie.⁵

Platforms have moved in the same direction. Databricks launched Unity Catalog Metrics in public preview in June 2025, explicitly to move metric definitions out of individual BI tools and onto the data platform.⁶

So the question for a vendor has changed. Ask whether the product can export its semantic definitions in the open format today, and ask to see it done. A demo of an export is worth more than a roadmap slide.