People ask us this most, so here is the straight answer.
The short version
Terazu is for companies where one number comes out of several systems, someone has to defend it, and the data stays on their own servers.
It pays off fastest just before an AI assistant starts answering business questions. Then a wrong number stops being an argument and becomes anyone’s answer.
The estate
A typical one: Oracle or SQL Server for operations, Informatica, Airflow or SSIS moving data, a warehouse, and Power BI, SSAS or Cognos on top.
What matters is one metric defined in a BI model, a procedure and a spreadsheet, with nobody sure which the board saw.
- More than one system. Terazu reads Postgres, SQL Server, Oracle, BigQuery, Informatica Cloud and Airflow live, and uploaded BI models, SSIS packages and Cognos models.
- History. Old and new systems side by side after a migration, or two sets after an acquisition.
- Disagreement. The same name, different numbers, each with an owner who’s sure.
- A reason to keep data at home. Regulation, contracts or policy. Terazu runs on your own servers, and can run air-gapped.
The certified route, from question to certified number, is measured on PostgreSQL so far.
The moments people go looking
Nobody looks for this on a quiet day. These moments start the conversation.
An AI assistant is about to go live on company data
An assistant that writes its own SQL sounds just as sure when it’s wrong.
On Spider 2.0, a benchmark of real enterprise databases, its authors’ agent on o1-preview solved 21.3% of 632 tasks, against 91.2% on Spider 1.0.1
dbt Labs put it plainly: “With text-to-SQL, failure looks like a plausible but incorrect answer.”2
Terazu answers from what your team certified, and your data team also gets working answers, marked not certified.
Someone who can only ask gets a certified number, an offer of one, or a refusal.
Two numbers reach the board
Finance and the business bring different figures under one name. In October, our simulated retailer’s warehouse gave three answers for last month’s sales:
3,807,294.42 SAR for calendar September, 4,564,960.60 for fiscal period 9, and 4,653,345.11 for period 9 before markdowns.
Your data team certifies one definition, every answer says which it used, and Terazu measures where else it’s computed. The owners make the call.
Old and new run side by side
Migrations leave two sets of systems running longer than planned. SQL Server 2016 left extended support in July 2026.3
From SQL Server 2025, Reporting Services gets no new versions; on-premises reporting moves to Power BI Report Server.4
A migration first needs a census: what exists, what feeds what, which reports still matter. It’s also the first thing Terazu builds.
Someone asks how a number was made
Regulators increasingly ask for definitions with owners, bound to where the data lives.
- Banks. The European Central Bank’s risk data guide expects banks to keep a dictionary of their main business definitions, with clear ownership.5
- Insurers. Solvency II requires processes to ensure “the appropriateness, completeness and accuracy of the data used in the calculation of their technical provisions”.6
- Listed companies. The SEC expects a key metric to come with “a clear definition of the metric and how it is calculated”, naming same-store sales.7
- Health plans. Under a CMS rule, many US health plans must publish prior-authorisation metrics yearly. The first were due by 31 March 2026.8
Terazu doesn’t make anyone compliant. It gives the accountable people a definition, an owner, the evidence and a trail for each number they certify.
Where it matters most
That points to banks and insurers, health plans, listed retailers and consumer brands, the public sector and utilities.
Our demo is a simulated retailer: Oracle tills, a Postgres warehouse, Airflow, Informatica Cloud and Cognos.
If a number has to hold up in front of someone who can say no, it needs an owner, a definition and a receipt.
The people inside
Everyone on the data team uses the same console. A role decides who may change and certify the contract, not who may read it.
- Analysts write down what numbers mean, and work out or check a number they were asked for, with its receipt.
- Engineers see what feeds what, from source to report, and which definitions rest on a table before they change it.
- Stewards certify definitions with the evidence attached, and get them back to review when something underneath changes.
- A newcomer reads the whole contract from day one as a viewer, and gets a certified number, an offer of one, or a refusal.
- The chief data officer usually buys it, and gets a map of the estate naming its blind spots, and the open questions to define next.
- Claude Code, Copilot and Cursor ask over MCP, seeing how a question was read and which definition answers it. The figure opens for the person.
Business users aren’t the audience. They keep their BI tools; publishing certified definitions to those tools is on the roadmap.
Three stories
Composites of the estates we build for, not customers. Terazu has no outside users yet, and its certified route is measured on PostgreSQL so far.
A grocer in the middle of a migration
Finance and merchandising bring the board different comp-sales figures. Terazu measures the gap; the next board pack cites the one both owners certify.
A bank after a supervisory finding
Exposure differs across an Oracle core, a SQL Server risk mart and an SSAS cube. The risk owner certifies each indicator, on the bank’s servers.
A health plan publishing its prior-authorisation numbers
Two teams disagree on the approval rate. Terazu drafts it, grain stated, for the director to certify. “Active provider” stays refused until someone defines it.
Who it isn’t for, yet
- One platform, one well-kept model. If everything lives in one warehouse with one semantic layer and nothing disagrees, you may not need Terazu yet.
- Only Snowflake or Databricks. Both connectors are on our roadmap, not in the product yet.
- Dashboards, ETL or document search. Terazu reads pipelines and BI models; it builds no dashboards and searches no documents. Writing pipelines is on the roadmap.
- AI that takes actions. The assistant in Terazu is read-only against your data.
- No owners. If nobody can own the definitions, nothing gets certified, and Terazu refuses, naming what’s missing. Honest, but not much use.
- The whole enterprise first. Modelling everything before anyone asks? We’re the wrong tool. We start with the numbers you already argue about.
How to find out
Run a census on the systems behind the numbers you argue about. Within a read budget you set, it shows what’s there, what disagrees and what it can’t see.
Sources
- Lei et al., Spider 2.0: Evaluating language models on real-world enterprise text-to-SQL workflows, ICLR 2025.
- dbt Labs, Semantic layer vs. text-to-SQL, April 2026.
- Microsoft, SQL Server 2016 lifecycle.
- Microsoft, Reporting Services consolidation FAQ.
- European Central Bank, Guide on effective risk data aggregation and risk reporting, May 2024.
- Directive 2009/138/EC (Solvency II), Article 82.
- US Securities and Exchange Commission, Release 33-10751, guidance on key performance indicators and metrics in MD&A, January 2020.
- Centers for Medicare & Medicaid Services, Interoperability and Prior Authorization Final Rule (CMS-0057-F).