Data Analytics Services in the AI Agent Era: Why Semantic Layers Matter

By now, a lot of companies have plugged an AI assistant into their data warehouse, the central database behind their reports. A sales director types “What was revenue last quarter?” and has a number on screen seconds later, with no request to the analytics team.
In practice, the numbers often come back wrong. An assistant asked for last quarter’s revenue may return a figure that does not match the finance report. This is seldom a fault in the model. The assistant looked for “revenue”, saw a column with exactly that label and totaled it. It could not know that finance leaves out refunds, test orders and one subsidiary when it reports revenue.
Data platform vendors now treat this as a problem they need to solve. At its AI World conference in late October, Oracle announced a Business Semantic Layer for its Autonomous AI Lakehouse. InfoWorld reported that the goal is to give AI agents a consistent understanding of enterprise data. Snowflake is working on something similar under the name Horizon Context. Databricks and Microsoft already offer something comparable: a metric is defined in one place and other tools pick it up from there.
Why AI agents give inconsistent answers
Until recently, most business questions reached the data through analysts. Analysts learn details that never make it into documentation. They know which tables contain test records, which region reports in a different currency and which system has its own definition of “customer”. They use this knowledge in every report, usually without writing it down.
None of that reaches an AI agent. An agent sees far less: the names of the tables and columns and, where someone took the trouble, a line of description. It turns the question into SQL and reports what the database returns. When the question could mean two things, the agent goes with one and keeps quiet about the other. A wrong answer therefore looks just as confident as a correct one.
This explains a common pattern in early AI analytics projects. Usage is high in the first weeks. Then somebody sets the assistant’s figure against an existing report, the two do not match, and people stop opening the tool.
Example: one question, three answers
Consider one ordinary request: “How many active customers did we have in Q3?” At a mid-sized company the CRM, billing and product analytics can all answer it, and each understands “active” in its own way.
| Source system | What “active customer” means there | Illustrative Q3 count |
| CRM | Account with an open deal or a renewal date in the quarter | 4,180 |
| Billing | Customer with at least one paid invoice in the quarter | 3,620 |
| Product analytics | Account with at least one login in the last 30 days of the quarter | 5,940 |
Each number is correct within its own system. The trouble starts when an agent takes the first table it finds and reports that number as the company’s count of active customers. Different people then get different answers to the same question, and someone has to spend time working out which number is right.
In the same InfoWorld article, Amit Kumar Jena of Kanerika described a pharma project on Microsoft Fabric. The team standardized its KPIs before the agents went live. According to Jena, “investigation and reconciliation time fell by about 70%.” Reconciliation here means finding out why two reports show different numbers. This is one consultant’s account of one project. It shows where inconsistent definitions usually cost the most: in the hours people spend comparing numbers.
What a semantic layer does
In simple terms, a semantic layer is the company’s dictionary for its own numbers. “Revenue” is written down there once, with its formula. A dashboard reads that entry, and so do an analyst’s notebook and an AI agent, so all three return one number.
The layer normally stores four things:
- Metric definitions. Each has a formula, filters and a named person who approves changes.
- Dimensions and hierarchies. A dimension is anything a metric can be split by: region, then country, then city, or product family and then single product.
- Relationships between tables, so a query joins orders to customers on the right keys and counts nothing twice.
- Access rules, so that each user sees only the data they are allowed to see, through a dashboard or through an AI agent.
According to InfoWorld, Oracle builds these models from catalog annotations, SQL annotations imported from its database and analytic views suggested by AI. Oracle says the feature will be available within the next 12 months, so companies cannot use it yet.
The software is only part of the work. As Jena pointed out, a semantic layer requires companies to define and maintain their business metrics, relationships and rules. Someone in the business has to decide, for example, that “active customer” in company-wide reports means a paying customer. That person also has to document the decision and update it when the business changes. These decisions have to come from the company. The vendor’s software stores them and applies them.
How AI agents change the scope of analytics work
Once AI agents use company data every day, several tasks that used to be optional become a standard part of an analytics project.
- Metric catalog. Definitions, their owners and the change history are handed to the client as a deliverable, the way a dashboard is.
- Data quality checks that never stop. A data pipeline is a scheduled job that takes data out of a source system and gets it into shape for reporting. Quality tests belong inside that job: they look for out-of-date records, rows loaded twice, empty fields and joins that have stopped matching. The agent answers from whatever the warehouse holds at that minute, and after a failed overnight load that can be half a table.
- Access control for agents. An agent that acts for a user must have the same permissions as that user, down to individual rows and columns.
- Lineage. This is the record of where a figure came from and how it was calculated. Every figure needs one, so that people can explain and check an answer.
- Answer testing. Somebody on the project keeps a file of real business questions with verified answers. When a data model, a definition or the AI model changes, the whole file gets run again.
That list of five is, more or less, the job description for data analytics services in 2026. Sombra, to take one provider, runs an analytics project in six stages, from strategic alignment and a current-state data assessment through to ongoing support. Conflicting definitions are usually found at the assessment stage. Fixing them there costs much less than fixing them after an assistant has been rolled out across the company.
Why AI answers depend on the data layer
Employees notice only a small part of all this: the box where they type a question, the summary at the top of a report, the short note on why a metric went up or down. All three features are exactly as good as the data layer they read from. Snowflake makes this argument for Horizon Context, a feature it presents as a governed semantic foundation for AI and BI.
Suppose a language model gets one approved definition of the metric, the joins it should use and only the rows it needs. Most of its answers will then be right. The same model, let loose on a few hundred tables with unclear names, answers just as fast. The figure looks fine, and it may be wrong.
How to prepare data for AI questions
It is enough to model one business area before anyone asks a question; the rest of the warehouse can wait. Followed in order, the six steps below normally get that area working in a few weeks.
- Collect the questions employees actually ask. Go through the analytics request queue, internal chats and management meeting notes; 20 to 40 questions from there will cover a first scope.
- For every question, follow the number back to its source systems and write down where two systems define the same term differently.
- Sit down with the business owners and settle each definition. Record it, and put one named person in charge of every metric.
- Repair data quality along these paths before touching anything else. Add tests and monitoring for the tables and pipelines involved.
- Load the agreed definitions, with access rules, into a semantic layer. A new product is rarely needed: Snowflake, Databricks, Microsoft Fabric, a dbt semantic model or the modeling layer of the existing BI tool can all do the job.
- Turn each question from step 1 into an automated test with its expected answer, and run the tests ahead of every release.
Step 6 is the one teams tend to leave out. A developer renames a column, or the AI vendor ships a new model version, and “How many active customers did we have?” starts returning a different number with no error anywhere. Automated tests show the change before the wrong number reaches a report.
Data analytics examples from delivery projects
The three projects below show the data foundation work that reliable AI analytics depends on. They were reporting and data engineering projects, and none was presented as an AI deployment.
For a marketing solution provider, Sombra built a custom reporting solution that cut reporting time in half. Most of the gain came from two changes. Data sources were consolidated, and the preparation steps that analysts used to repeat by hand in every reporting cycle were automated.
For a leading logistics provider in the USA, Sombra’s work is credited with a 30% SLA increase. Service levels in logistics depend on operational data that arrives on time and in a consistent format, so a large part of the project was integration and data quality work.
The third client is an eClinical company. Its data pipeline was rebuilt by the data engineering team so that reports update in real time, and the new pipeline costs less to run. Clinical data is tightly restricted, so access controls were written into the pipeline design at the very beginning.
All three clients ended up with definitions that match, pipelines that somebody watches and access that is under control. An AI agent needs exactly these things before its answers can be relied on.
How to choose an analytics partner
A search for a partner often starts with typing “data analytics consulting near me” into Google. The “near me” part matters little now that most of this work is done remotely. What still counts is overlap in working hours, knowledge of the industry and knowledge of the regulations the company’s data falls under.
The six questions below work equally well for a local firm and for a provider of data and analytics services based abroad. Each one checks how ready the provider is for analytics work with AI agents in it:
- Suppose sales and finance each have their own definition of one metric. How do you settle it, and who signs off?
- Which semantic layer or metrics tools are live at your clients right now, and what platform does each run on?
- An agent answers on behalf of one of our employees. What prevents it from pulling data this employee has no access to?
- What does your testing of natural-language answers look like before a release and after it?
- Do you have data analytics examples from our industry, with results measured in numbers?
- A year after go-live, who is still looking after the definitions, the pipelines and the tests: your team or ours?
A provider that has done this before can show a metric catalog, or the set of test questions from an earlier project. Such a provider is more likely to deliver data analytics solutions that stay reliable as new tools are connected. A data analytics company that works mainly on dashboards and visualizations covers only part of what is needed now.
The last thing to check is how the provider’s data analytics services and solutions connect with the other systems the company runs. An analytics project today reaches into AI work, application integrations and cloud spending. A provider that has worked on all of these has a better idea of where to put a definition or a quality check.
How to measure the results
Compared with a new dashboard, data foundation work gives a budget holder very little to look at, which makes it hard to defend. So agree on five measures before the project starts, and note down today’s value for each one:
- Answer accuracy. Count the test questions that come back correct and express them as a percentage. What counts as good enough differs by use: a few misses are tolerable in a rough internal estimate and a serious problem in a financial report.
- Reconciliation time. Each month, analysts and finance people lose hours to explaining why two reports disagree. Add those hours up.
- Analyst request volume. Before the project, count the routine data requests the analytics team receives in a month. Count again once employees can ask the data themselves.
- Definition coverage. List the metrics people use most. The measure is the share of that list with an agreed definition and a named owner.
- Data incidents. Every broken pipeline or late load that reached a report or an answer goes into a log, with two facts: when it happened and when the team found out.
The same numbers point to the next job. If accuracy stays low for one department’s questions, a definition is usually missing there, or the data behind those questions has a quality problem.
What all this costs, and how long it takes, depends mostly on the state of the data at the start. Sombra’s own guidance is that simple analytics projects can start from a few thousand dollars and that most take from a few weeks to several months. An assessment of questions, sources and definitions is usually the first and smallest stage. Its findings set the budget for the rest of the work.
Summary
This year’s semantic layer announcements respond to a practical problem. More and more employees now get to company data through an AI assistant or agent. With no analyst in between, a conflicting definition or a bad table shows up far sooner than it ever did under traditional reporting.
A company that is planning analytics work for 2027 should do four things:
- define the main business metrics, with named owners, at the start of the project
- run data quality checks continuously on the pipelines that feed real questions
- apply the same access rules to AI agents as to the people they act for
- test the AI’s answers ahead of each release, exactly as developers test code
Experienced teams open a project with an assessment of the questions employees already ask and the data behind them. A few weeks is usually enough to learn which numbers an AI system can already be trusted with, and which need more work.
