AI brokers fail on enterprise knowledge for a predictable cause: they’re requested to guess what the enterprise means. Level an agent at uncooked tables and it has to deduce what “income” consists of, which joins are secure, and who’s allowed to see the consequence. It guesses otherwise every time, so the identical query returns completely different numbers.
A semantic layer for AI brokers sits between your knowledge and the brokers that question it, offering a ruled mannequin of enterprise that means. It defines metrics, entities, relationships, and permissions as soon as, so brokers choose trusted definitions as an alternative of writing their very own logic towards uncooked tables. The result’s constant solutions, enforced entry guidelines, and numbers anybody can hint again to a definition.
The urgency is actual. In March 2026, Gartner predicted that by 2030, common semantic layers might be handled as vital infrastructure, alongside knowledge platforms and cybersecurity. This information explains what brokers want from a semantic layer, how the items work collectively, and how one can get an present semantic layer prepared for brokers.
Key Takeaways
- A semantic layer for AI brokers defines enterprise that means as soon as, so brokers choose ruled metrics as an alternative of producing SQL towards uncooked tables.
- Textual content-to-SQL offers an agent entry to knowledge; a semantic layer offers it an agreed definition of what the information means.
- Brokers are a special sort of knowledge client: they question repeatedly, at excessive quantity, and more and more take actions, not simply learn outcomes.
- An AI-ready semantic layer wants seven properties: shared definitions, readable metadata, per-request permissions for reads and actions, open interfaces, predictable efficiency, traceable solutions, and secure change administration.
- Definitions in legacy BI instruments are invaluable enter, not an agent-ready layer; most groups migrate that logic right into a semantic layer designed for brokers relatively than modeling from scratch.
- A semantic layer covers that means; accepted sources, enterprise priorities, and deeper observability belong to a broader context layer.
What’s a semantic layer for AI brokers?
A semantic layer for AI brokers is a ruled mannequin of enterprise that means that an agent can uncover, question, and belief. It interprets technical buildings, similar to tables, columns, and be a part of keys, into enterprise ideas like clients, orders, and internet income. The agent works with these ideas; the semantic layer handles the interpretation to SQL.
For an agent, the semantic layer provides 5 issues it can not reliably infer by itself:
- Metrics: actual definitions of measures similar to internet income, churn, or lively customers, together with filters and time logic.
- Entities and dimensions: the enterprise objects an agent causes about, similar to buyer, area, or product line.
- Relationships: accepted be a part of paths and grain, so a query that spans two tables doesn’t double-count rows.
- Permissions: guidelines that resolve which rows, columns, and actions a given person or agent could entry.
- Descriptions and synonyms: human-readable context that maps on a regular basis language (“gross sales”, “bookings”) to the precise object.
The layer itself just isn’t new. Enterprise intelligence (BI) instruments have used semantic layers for many years to maintain dashboards constant. What modifications with AI brokers is the patron: the identical definitions now should serve software program that asks questions by itself, at machine pace, on behalf of many various individuals.
Why AI brokers want a semantic layer (and why text-to-SQL is not sufficient)
AI brokers want a semantic layer as a result of the that means of enterprise knowledge just isn’t saved within the schema. Ask “What was This autumn income?” and three solutions can all be legitimate: gross sales counts booked offers, finance counts acknowledged income, and treasury counts money collected. A desk named orders doesn’t inform a language mannequin which one you meant.
Textual content-to-SQL (translating a natural-language query instantly right into a SQL question) solves entry, not that means. Trendy LLMs write syntactically appropriate SQL with ease. The failures occur one degree up, in 4 recurring patterns.
What’s metric drift?
Metric drift is when the identical query produces completely different numbers as a result of the agent re-derives the metric on each immediate. One run excludes refunds, the subsequent consists of tax, the third makes use of ship date as an alternative of order date. Every question is legitimate SQL. None of them matches the quantity finance experiences, and customers cease trusting the agent after the primary mismatch.
Why do fallacious joins and grain matter?
Mistaken joins silently inflate outcomes. Actual warehouses include fan-out joins, slowly altering dimensions, and a number of other tables that each one appear like “the client desk.” If an agent joins orders to order traces earlier than summing order totals, income doubles and nothing flags the error. A semantic layer fixes the be a part of path and grain upfront, so the agent can not select a fallacious one.
Why would not a SELECT assertion implement permissions?
A SQL question has no built-in notion of who’s asking. An accurate question and a question that exposes one other buyer’s knowledge look an identical to the mannequin. If entry guidelines reside solely in a dashboard or in a immediate, an agent can bypass them with one surprising query. Permissions have to use when the question is generated, not after the consequence comes again.
Why are unexplainable solutions an issue?
A solution that no person can hint can’t be utilized in a choice. When an agent returns a quantity backed solely by 40 traces of generated SQL, an analyst has to reverse-engineer the logic earlier than signing off. When the reply names the metric definition, filters, and time vary it used, evaluate takes seconds.
Because of this GoodData.AI takes a metadata-first place: the LLM works with ruled metadata and by no means guesses SQL towards uncooked tables. The reasoning behind that design is defined in why AI in analytics wants metadata.
Go deeper: Most present semantic layers have been constructed for dashboards, not brokers. The whitepaper Why Brokers Will Hallucinate on Your Present Semantic Layer explains how scattered BI logic causes agent errors and how one can refactor it.
Brokers are a special sort of knowledge client
AI brokers change who consumes the semantic layer, not simply how usually. A BI semantic layer was designed for individuals studying dashboards at human pace. In the present day the identical definitions serve a spectrum of customers, from analysts exploring a dashboard to autonomous brokers that run a enterprise course of finish to finish. Every step alongside that spectrum raises the bar.
| Client | Who asks | How usually | What it wants from the semantic layer | Value of a fallacious definition |
|---|---|---|---|---|
| Dashboards and self-service | An individual exploring knowledge | Often, at human pace | Constant metrics throughout experiences | A deceptive chart somebody could discover |
| Knowledge purposes | Many finish customers inside a product | Repeatedly, many customers without delay | Constant metrics plus per-customer isolation | Mistaken numbers proven to clients |
| Assistants and copilots | An individual delegating evaluation | Many follow-up questions per session | Readable descriptions, synonyms, and traceable solutions | A assured fallacious reply utilized in a choice |
| Autonomous brokers | Software program performing by itself | Continuously, many queries per job | The entire above, plus permissions for actions and approvals | A fallacious motion, similar to an up to date report or a triggered workflow |
The sensible conclusion is to run one semantic layer for all 4, not a separate layer per client. Two layers drift aside the second somebody edits a metric in solely one in every of them. One ruled layer means a dashboard, a customer-facing app, and an agent return the identical This autumn income, calculated the identical method, below the identical entry guidelines.
The shift additionally explains why agentic analytics raises new necessities for the information basis. Brokers don’t simply learn extra; they ask unanticipated follow-up questions, chain a number of queries into one job, and more and more act on the outcomes. Serving all 4 customers nicely takes one ruled serving basis for individuals, purposes, and brokers, not a stack of separate integrations.
What an AI-ready semantic layer wants
An AI-ready semantic layer wants seven properties {that a} dashboard-era semantic layer usually lacks. The primary two make that means express; the opposite 5 make that that means secure to serve to brokers at scale and secure to vary over time.
Enterprise that means outlined as soon as
Each metric, entity, and relationship ought to have precisely one ruled definition that each one customers reuse. If “internet income” exists in a dashboard components, a SQL view, and a spreadsheet, an agent has three candidates and no method to decide on. Defining it as soon as, ideally as version-controlled code, turns a guess right into a lookup.
Metadata brokers can really learn
Brokers depend on names and descriptions way over human analysts do. A metric referred to as ASP_v2_final with no description leaves the agent guessing. “Common promoting value (ASP): internet income divided by models offered, excluding returns” tells it what the metric means and the way it’s calculated. AI-ready metadata consists of clear titles, descriptions that lead with enterprise that means, synonyms, expanded abbreviations, and an proprietor.
Automated checks assist right here. In GoodData.AI, the Semantic High quality Agent flags duplicate or too-similar titles and descriptions and unknown abbreviations throughout metrics, attributes, details, and date datasets, so you possibly can repair them earlier than brokers encounter them. AI Reminiscence and AI Data then floor the assistant in every workspace’s enterprise phrases, so the synonyms and conventions your groups use level to the precise objects.
Permissions for what brokers can see and do
Agent permissions should cowl each knowledge entry and actions, they usually have to be evaluated on each request, not as soon as per connection. Being related to an accepted software or holding a broad function just isn’t sufficient. The safer sample is that an agent works below the identification of the individual it acts for: each request carries that person’s permissions, so the agent can see, compute, and alter solely what the person may. Row-level and column-level guidelines apply when the semantic layer compiles a question, so an agent performing for one person can not return one other person’s knowledge.
Actions want the identical self-discipline. Making a metric, a visualization, or an alert ought to depend upon express rights, and modifications an agent proposes must be reviewable earlier than they’re saved.
In GoodData.AI, one permission mannequin covers BI and AI. Object-level permissions management who can view, share, and edit particular person columns, metrics, and visualizations, and the UI, APIs, SDKs, AI Assistant, Good Search, and the MCP Server run every request with the authenticated person’s permissions. Requests from exterior brokers over A2A additionally respect which AI brokers the person could entry. Creating metrics and visualizations is a separate permission, and when the AI Assistant builds a dashboard that provides new metrics or visualizations, it reveals what might be created earlier than something is saved.
Platforms that implement permissions in each AI interplay deal with this as governance by design, not as an instruction in a immediate. In customer-facing merchandise, every tenant wants strict isolation; a multi-tenant structure retains one buyer’s agent from ever reaching one other buyer’s knowledge.
Open interfaces for brokers
Brokers want an ordinary technique to uncover and request ruled metrics. The Mannequin Context Protocol (MCP) has turn out to be the frequent interface: an agent lists obtainable metrics and dimensions, then asks for them by identify as an alternative of writing SQL. Agent-to-agent (A2A) protocols, SQL, and SDKs cowl different purchasers. The requirement is that each interface serves the identical definitions and the identical guidelines. For a sensible instance, see how GoodData.AI constructed an MCP server for ruled analytics. Solutions may come again as visuals: with MCP Apps, GoodData.AI visualizations render instantly inside MCP purchasers similar to Claude, with ruled metrics and the person’s permissions.
Predictable efficiency and price below agent load
Agent site visitors is steady and bursty, not like the regular load of individuals opening dashboards. A single agent job can generate tens or lots of of short-lived queries, and lots of of brokers can run without delay. Dealing with that load takes greater than uncooked question pace. Workload isolation prevents one agent from slowing down everybody else. Scheduling, prioritization, and caching assist handle each interactive and automatic requests.
Reusing cached outcomes for repeated questions retains response instances and prices predictable. In GoodData.AI, caching with validity durations and pre-aggregation preserve agent site visitors off manufacturing warehouse compute, and every knowledge supply can expire cached outcomes after a set interval or on a schedule that follows your knowledge refreshes. The GoodData.AI structure overview describes the strategy to caching and high-concurrency workloads.
Solutions you possibly can hint again to a definition
Each reply an agent returns ought to embody the metric definition, filters, dimensions, and time vary used to supply it. If somebody questions a quantity, they will shortly examine the way it was calculated. When the underlying knowledge is incomplete, stale, or ambiguous, or when a query falls outdoors the outlined semantic, coverage, or knowledge boundary, the system ought to say so as an alternative of manufacturing a believable quantity. Tracing is strongest when it lives the place the agent runs: an observability layer contained in the platform that runs the agent already is aware of the workspace, the person’s function, and the semantic mannequin behind every name.
Modifications which can be versioned, examined, and reversible
Enterprise definitions, insurance policies, and knowledge sources change continuously, and brokers multiply the affect of each change. A small edit to a metric definition or an entry rule can alter 1000’s of automated solutions and actions in a single day. An AI-ready semantic layer treats mannequin modifications like software program releases. Modifications are versioned and examined towards recognized solutions earlier than deployment, with rollback obtainable if wanted. Storing the mannequin as code makes this sensible; step 4 within the readiness information under covers testing.
AI-ready semantic layer guidelines
| Property | Why brokers want it | How you can confirm it |
|---|---|---|
| Enterprise that means outlined as soon as | Prevents metric drift throughout prompts | Every core metric has one definition, saved as code |
| Readable metadata | Brokers select objects by identify and outline | No undefined abbreviations; each metric has an outline and proprietor |
| Per-request permissions for reads and actions | Stops knowledge leaks and unreviewed modifications | Each request runs below the requesting person’s permissions; creating content material wants express rights, and agent modifications might be reviewed earlier than they’re saved |
| Open interfaces | Brokers uncover metrics as an alternative of guessing SQL | Brokers can checklist and request metrics via MCP or an API |
| Predictable efficiency | Agent site visitors is bursty and excessive quantity | Latency and price keep steady because the variety of brokers grows |
| Traceable solutions | Customers should confirm numbers earlier than performing | Every reply returns the metric, filters, and time vary used, and flags unsure knowledge |
| Secure change administration | One edited definition can change 1000’s of agent solutions | Mannequin modifications are versioned, examined earlier than launch, and might be rolled again |
What a semantic layer would not cowl by itself
A semantic layer defines what the information means, nevertheless it doesn’t resolve which paperwork rely as accepted sources, which enterprise priorities ought to form a solution, or how an agent’s full reasoning is reviewed afterward. These wants sit in a broader context layer, coated within the part on semantic layer vs. context layer under.
The way it works: from a query to a ruled reply
A semantic-layer-grounded agent by no means writes SQL towards uncooked tables; it selects ruled objects and lets the semantic layer generate the question. Take a regional gross sales supervisor who asks an agent: “What was internet income by area final quarter?” Here’s what occurs step-by-step.
- Uncover. The agent asks the semantic layer, usually over MCP, which metrics and dimensions can be found to this person. It receives a catalog with names and descriptions, similar to net_revenue and area, not desk definitions.
- Choose. The agent builds a structured request: metric net_revenue, grouped by area, filtered to the earlier fiscal quarter.
- Compile below the person’s permissions. The semantic layer turns the request into SQL, applies the accepted be a part of path, and provides the person’s entry guidelines. A supervisor who sees solely the EMEA area will get EMEA rows, even when the query requested for all areas.
- Execute. The question runs towards the information supply, or returns from cache if the identical consequence was computed lately.
- Return with proof. The agent receives the numbers along with the metric definition, filters, and time vary it used.
- Clarify. The agent summarizes the end in plain language and cites the definition, so the supervisor can examine precisely which “income” the reply makes use of.
Actual agent duties not often cease at one query. An agent requested why income is declining would possibly examine areas, then product traces, then particular person accounts, with every consequence deciding the subsequent question. The identical steps repeat for each request in that chain, so “income” retains one that means and the identical entry guidelines apply from the primary question to the final.
The definition the agent depends on is written as soon as by the information group. In GoodData.AI’s analytics-as-code format, a metric is a YAML file with a MAQL expression:
kind: metric
id: net_revenue
title: Web income
description: >
Accomplished order income after refunds and reductions, excluding tax.
Additionally referred to as "internet gross sales". Owned by Finance.
maql: >
SELECT SUM({truth/order_amount}) - SUM({truth/refund_amount}) - SUM({truth/discount_amount})
WHERE {label/order_status} = "accomplished"
format: "#,##0.00"
The metric carries no mounted date: the agent picks the time dimension and vary per query, so the identical definition solutions “final quarter” and “yr so far.” The title and outline matter as a lot because the components: they’re what the agent reads when it decides which metric solutions “internet gross sales final quarter.” For the implementation particulars of 1 such interface, see the GoodData.AI MCP Server documentation.
The place semantics can reside: three approaches
Semantic definitions for brokers can reside in three locations, and the precise selection depends upon how far your brokers attain. The important thing query is easy: will your brokers keep inside one knowledge platform, or serve a number of instruments, groups, and clients?
- Inside the information platform. Warehouse-native semantics preserve definitions subsequent to the information. That is handy when all knowledge and all AI reside in a single platform. Brokers, apps, and tenants outdoors that platform want their very own integration and their very own guidelines.
- In a standalone semantic layer. A separate modeling layer above the warehouse stays impartial throughout BI instruments and apps. You assemble the remaining your self: agent orchestration, permissions for actions, and observability.
- In an analytics platform with an built-in semantic layer. The identical definitions run dashboards, embedded apps, assistants, and brokers. GoodData.AI takes this strategy, serving one ruled layer via MCP, A2A, SQL, and SDKs, with governance, multitenancy, and observability in a single place. The trade-off is adopting a broader platform.
No strategy is universally greatest. Whichever you select, preserve definitions transportable. Apache Ossie (Incubating), previously the Open Semantic Interchange (OSI), is constructing a vendor-neutral specification for exchanging semantic fashions between BI, AI, and knowledge platforms, so metric definitions wouldn’t have to be rewritten for every software. Portability additionally means preserving knowledge in place: definitions ought to journey between engines with out copying the information beneath them.
Some groups lengthen the semantic layer with an ontology or data graph to seize richer relationships between ideas. That sample, defined in from RAG to GraphRAG, enhances a semantic layer relatively than changing it.
How you can get your semantic layer prepared for brokers
Getting a semantic layer prepared for brokers is an incremental challenge, not a rebuild. Most organizations have constructed up years of enterprise logic of their BI instruments. The duty is emigrate that logic, clear it up, and serve it via one ruled layer. 5 steps cowl the trail from a primary use case to manufacturing.
- Migrate the logic you have already got. Your metric definitions exist already inside Energy BI, Tableau, Qlik, or Looker recordsdata, usually in a number of conflicting variations. These definitions have been constructed for dashboards, so connecting brokers to them as they’re doesn’t make them agent-ready. Deal with them as enter: establish duplicates, resolve conflicts, and migrate the logic right into a semantic layer designed for brokers as an alternative of modeling from a clean web page. AI-assisted migration now makes it sensible to migrate logic out of legacy BI instruments with out rebuilding each definition by hand. Aklamio, for instance, moved its analytics from Tableau and Superset to GoodData after which prolonged the identical ruled setup to partner-facing analytics.
- Decide one use case that prices you time. Select a single course of with a transparent proprietor and measurable ache, similar to weekly income reporting or churn investigation. A slim first use case reveals which definitions brokers really want.
- Mannequin metrics, entities, and permissions as soon as. Outline the metrics, dimensions, relationships, and entry guidelines the use case wants, with descriptions brokers can learn. Retailer them as code so each change is reviewed and versioned.
- Check towards recognized solutions earlier than manufacturing. Construct a set of enterprise questions with solutions the information group has verified, together with ambiguous wording and denied-access circumstances. Run the agent towards them earlier than launch and after each mannequin change; the three constructing blocks listed here are traces, evaluations, and metrics.
- Broaden what runs. Every new agent or use case reuses the identical ruled definitions, interfaces, and guidelines. The second use case is quicker than the primary as a result of the muse already exists.
Semantic layer vs. context layer
A semantic layer tells an agent what the information means; a context layer tells it every part else it must act responsibly. The semantic layer covers that means and the entry guidelines utilized to queries. A context layer provides accepted sources, together with unstructured data similar to insurance policies and paperwork, enterprise priorities that form a solution, reminiscence throughout interactions, and observability of how every output was produced.
The 2 are usually not opponents. The semantic layer is the muse, and the context layer builds on it. In GoodData.AI, for instance, Agent Directions let directors outline enterprise guidelines an agent follows, similar to all the time displaying each Complete Gross sales and Complete Web Gross sales when somebody asks about gross sales. Shared definitions stay the precise place to start out, however because the article on why shared definitions resolve solely a part of the issue explains, manufacturing AI additionally wants governance, grounding, steerage, and evaluate round these definitions.
How GoodData.AI approaches the semantic layer for AI brokers
GoodData.AI’s place is that dashboards, knowledge purposes, assistants, and autonomous brokers ought to run on one ruled semantic basis, not on separate layers per client. In GoodData.AI, that basis is the Agentic Serving Airplane, the platform’s structure for serving knowledge, semantics, context, and coverage to each client. It combines a number of capabilities:
- Semantic mannequin and Analytics Catalog: metrics, attributes, and relationships outlined as soon as, with a catalog and a Semantic High quality Agent that flags unclear or duplicate titles and descriptions.
- Context Administration: accepted data, AI Reminiscence, agent directions, and entry guidelines carried throughout each AI expertise.
- One permission mannequin for BI and AI: object-, column-, and row-level permissions utilized to each request, whether or not it comes from an individual, the AI Assistant, or an exterior agent.
- MCP Server and open interfaces: ruled entry for exterior brokers and developer instruments over MCP, A2A, SQL, and SDKs, with MCP Apps that render visualizations inside MCP purchasers.
- Multitenancy: strict isolation for brokers embedded in customer-facing merchandise.
- Migration brokers: AI-assisted migration of dashboards, metrics, and enterprise logic from Energy BI, Tableau, and different legacy BI instruments into the GoodData.AI platform.
- AI Observability: utilization analytics that present adoption, failures, and timeouts throughout AI experiences, and explainable solutions: open any reply and see every step, what it used, and the place it failed.
To see how these items match collectively by yourself knowledge, request a GoodData.AI demo.
Often requested questions
A semantic layer in AI is a ruled mannequin of enterprise that means that AI techniques question as an alternative of uncooked tables. It defines metrics, entities, relationships, and permissions as soon as, so LLMs and brokers use accepted definitions relatively than inferring them. This makes AI solutions constant, safe, and traceable.
You want a semantic layer for generative AI as quickly as its solutions are utilized by somebody aside from the one that wrote the question. With out one, the mannequin re-derives enterprise logic on each immediate, so the identical query can return completely different numbers. A semantic layer fixes the definitions and applies entry guidelines earlier than any SQL runs.
Textual content-to-SQL just isn’t sufficient for AI brokers as a result of it offers entry to knowledge with out an agreed that means. The mannequin nonetheless has to guess metric definitions, be a part of paths, and permissions. A semantic layer lets the agent request ruled metrics by identify, so the exhausting logic is solved as soon as by the information group.
An AI-powered semantic layer serves enterprise definitions to AI brokers in a kind they will uncover and use. It exposes metrics and descriptions via interfaces similar to MCP, compiles agent requests into SQL below the person’s permissions, and returns outcomes with the definitions used. Some platforms additionally use AI to examine metadata high quality and emigrate logic from legacy BI instruments.
AI brokers usually entry a semantic layer via the Mannequin Context Protocol (MCP), an open customary for connecting AI fashions to instruments and knowledge. The agent lists obtainable metrics and dimensions, then requests them by identify. APIs, SQL endpoints, SDKs, and agent-to-agent (A2A) protocols serve different purchasers.
An present BI semantic layer is a helpful supply of enterprise logic, however brokers not often work reliably on prime of it as it’s. Legacy layers have been designed for dashboards, in order that they usually lack descriptions and synonyms, include duplicate metric definitions, cowl reads however not actions, and supply no machine-friendly interface. The sensible path is emigrate that logic, together with definitions nonetheless buried in particular person dashboards, right into a semantic layer designed for brokers.
A semantic mannequin for AI is the structured set of definitions inside a semantic layer: datasets, attributes, metrics, and the relationships between them. The semantic layer is the service that shops, governs, and serves that mannequin to customers. Be taught extra on this information to the semantic knowledge mannequin.
A semantic layer defines what enterprise knowledge means, whereas a context layer provides every part else an AI system must act responsibly: accepted sources, unstructured data, enterprise priorities, reminiscence, and observability. The semantic layer is the muse of a context layer, not a alternative for it.
