--- title: "AI on Your Data: Governed, Private AI | Smart Data" url: https://www.smartdata.net/services/ai-on-your-data description: "Governed AI on your own data: schema-grounded SQL with answers you can check, access rules enforced at query time, private and local models, AI cost control." source: "Smart Data (smartdata.net). Markdown copy of the page; the HTML page is canonical." --- AI # Governed AI on your own data A language model pointed at a lakehouse with no structure around it runs up a large bill and gives answers nobody can check. We design and build AI that runs on your own data, inside your access rules, on models sized to the job, with the query behind every answer in plain view. Sound familiar? ## Where AI on enterprise data stalls Most companies already use AI somewhere. The trouble starts when it meets the data the business runs on. The pilot impressed in a demo and never reached production It ran on sample data with nobody’s access rules applied. Production has both. We start from the governed data the pilot skipped. Start with: the data the use case depends on. The assistant answers confidently, and sometimes wrongly A model guessing at table names and business definitions will guess wrong. Give it the schema and the definitions, show the query behind every answer, and a wrong number gets caught before anyone acts on it. Start with: one governed domain and its definitions. Usage keeps climbing, and so does the bill Sending every question to the largest hosted model is the most expensive way to answer it. Routing work to right-sized and private models turns AI spend back into a number you can plan. Start with: what is used today, by whom and at what cost. How we approach it ## What we design and build Context and semantics come before agents. These are the pieces of a governed AI system on your own data, usually built in this order. Schema-grounded answers ### Schema-grounded SQL with checkable answers The model gets your schema, your business definitions and the question, and writes the SQL that answers it. The system around the model is built to keep it from making answers up: queries are validated against the schema, run with the asker’s permissions and returned with the answer, so anyone can see how the number was reached. - The generated query shown with every answer - A router sends each question to a translator that knows that part of the business, working from the semantic model and the schema - Queries validated against the schema before they run - Business definitions kept in a semantic layer, not buried in prompts Tools we work in - Microsoft Fabric - Databricks - Snowflake Governed access ### Governed access and audit Access rules, sensitivity labels and audit trails come from the data platform, so the model sees only what the person asking is allowed to see, and every question and answer can be traced afterwards. - Row- and column-level security enforced at query time - Every question, query and answer logged for audit - Sensitive fields classified before any model can reach them - Microsoft Purview - Unity Catalog - Microsoft Entra ID Private and local models For high-volume or regulated workloads, a small private model can cost less to run and be easier to govern than a public API. We design for hosted, private-cloud and on-premises models, chosen by what the data allows and what the task needs, and keep regulated data inside your environment. - Open-weight models run in your cloud tenant or on your own hardware - Hosted frontier models used where the task needs them and the data permits - One interface for your people, whichever model answers - Ollama - Azure - AWS Model routing and AI cost ### Model routing and AI cost control Not every question needs the largest model. We route each task to the smallest model that answers it well, cache what repeats and measure cost and accuracy per answer before a workload scales, so you know what the AI bill will be before it arrives. - Routing rules that match each task to a right-sized model - Cost per answer tracked alongside accuracy - Spend visible by team and use case before the invoice arrives Orchestration ### AI orchestration Once answers are grounded and governed, they can drive work: multi-step workflows that look something up, check it against policy and hand it to the next system or person. A person approves anything that changes a system of record. - Workflows built from grounded, permission-aware steps - Human approval before any write to a system of record - Every step logged, so a result can be traced back to its sources Our point of view ## Don’t just point a language model at your lakehouse Treat AI spend the way you treat cloud spend: track the cost of each answer, where its data came from and whether it was right. That’s how AI on your own data stays governed, and its cost stays planned. ### Context before agents Schema, definitions and a semantic layer come first. ### Cost is a design input Model choice, routing and caching are decided at design time, with a cost per answer attached. ### Your data stays yours Private and local models where the data requires them, and hosted models only where it allows. ### Every answer can be checked The query, the sources and the model behind each answer are visible to the person who asked. How this connects to AI ## Not sure your data is ready for AI? Not sure your data is ready? The AI Enablement Guide covers what the AI programs that reach production do differently, and how to assess your own readiness. How an engagement runs ## One governed use case first, then the next Aim for one Gold use case in 90 days, not a full data warehouse rebuild. Each phase ends with something you can use and a decision about the next. ### 01Workshop One to two weeks on your data, systems and use cases. Where AI is already in use, what it costs, what data it touches and which use case has the data to support it. Deliverable A readiness scorecard, a 90-day roadmap of use cases ranked by impact and data readiness, a recommended architecture and a scope for the first build. ### 02Prove it on one domain A context layer on one governed data domain, then schema-grounded answers on top of it, on your real data and in your environment. Working software on your data, with accuracy and cost per answer measured. ### 03Scale and run More domains, routing and private models where they pay off, and the governance to keep it all auditable. We run it with you or hand it to your team. A routing and hosting design, a governance policy and a support plan. Industries ## Industries we work in ### Healthcare Patient, pharmacy and clinical data systems held to HIPAA, including where AI is allowed to touch them. - Medicaid programs - Pharmacy services - Clinical documentation [Healthcare](https://www.smartdata.net/industries/healthcare) ### Manufacturing and industrial ERP, shop-floor and supplier data on one platform, so plant reporting and finance reporting reconcile. - SAP and ERP reporting - Shop-floor integration - Supplier portals [Manufacturing and industrial](https://www.smartdata.net/industries/manufacturing) ### Retail and distribution Order, inventory, credit and customer data connected across portals, branches and the platforms behind them. - Vendor and customer portals - Order and credit systems - Data platforms [Retail and distribution](https://www.smartdata.net/industries/retail) ### Financial services Bank data platforms, reporting that holds up to an audit and managed support for the financial systems the business runs on. - Bank data platforms - Managed ERP support - Legacy migration [Financial services](https://www.smartdata.net/industries/banking-financial-services) ### Insurance Claims, member and policy systems for health plans and insurers. - Health plans - Claims data - Member apps [Insurance](https://www.smartdata.net/industries/insurance) Case studies ## The data and AI work underneath 78% ### of legacy transformation types handled by the AI-assisted migration tool An AI-assisted tool that converts legacy ETL into tested Databricks pipelines. In a three-month first phase across 177 legacy jobs and scripts, it handled 78% of the legacy transformation types, with tests generated alongside every conversion. - ETL & Data Pipelines - Claude [Legacy ETL to Databricks for a major national retailer: of legacy transformation types handled by the AI-assisted migrat](https://www.smartdata.net/case-studies/databricks-migration-national-retailer) $16M+ ### in lost Medicaid revenue returned to the health plan Hired to add a state-mandated code to a health plan’s nursing facility file, our team noticed the process was mishandling records. The plan confirmed it, filed corrections and recovered $16M+. - SQL Server [A routine Medicaid code update that surfaced $16M+ in lost revenue: in lost Medicaid revenue returned to the health plan](https://www.smartdata.net/case-studies/nfloc-platform-rewrite) ### Reducing data quality issues across four migrated programs Turnover had taken the onboarding knowledge with it. Recurring data quality issues across SSIS and MuleSoft feeds were fixed, onboarding was documented and four programs were migrated. - SSIS - MuleSoft [Reducing data quality issues across four migrated programs](https://www.smartdata.net/case-studies/enterprise-grade-data-quality-for-a-pharma-services-provider) A few practical questions ## Questions about AI on your data What data and IT leaders usually ask before AI touches production data. The model is given your database schema, your business definitions and the question, and writes the SQL that answers it. The query runs with the asker’s permissions and comes back with the result, so the answer can be checked. Riel St. Amand, our AI Solutions Advocate, presented the method at Dayton AI Day in August 2026 in a talk titled “Beyond RAG and Fine-Tuning: Schema-Grounded SQL Generation for Verifiable Enterprise Analytics.” Because it will answer whether or not it knows. Without the schema, the definitions and the access rules, a model guesses at what a column means and who may see it, and the cost grows with every question. That’s the work we do before anyone in the business uses it. Not always. Hosted models are the right call for many tasks. Private or local models make sense where the data is regulated, where volume makes per-call pricing expensive or where the model should stay in your environment. We design for the mix your data and budget support. No. The assessment shows which use cases your data supports today and what has to change before the rest become viable. The first build usually starts on the domain that is already in the best shape. We only use AI tools with training on your data turned off. If your contract restricts AI or requires stricter handling, such as zero data retention, your contract wins. We use US-based providers. Microsoft Fabric, Databricks and Snowflake, and the governance tools that come with them. The semantic layer and access rules live in the platform you already run, so AI inherits them. Related services ## Where AI on your data connects ### AI-assisted modernization Security backlogs, framework upgrades and legacy ETL moved forward by engineers working with Claude and other current models. [AI-assisted modernization](https://www.smartdata.net/services/ai-assisted-modernization) ### Data governance and quality Quality monitoring, access controls and lineage built into the platform your AI will read from. [Data governance and quality](https://www.smartdata.net/services/data-governance) One governed foundation for engineering, warehousing, Power BI and AI on the Microsoft estate. [Microsoft Fabric](https://www.smartdata.net/services/microsoft-fabric-consulting) ### AI readiness workshop What your data can support today, and which use case to build first. [AI readiness workshop](https://www.smartdata.net/schedule/ai-readiness-workshop) Next step ## Which question should your data answer first? Tell us the question, the data behind it and how AI is used today. Here is what happens after you reach out. ### A short call about the use case and its data What you want answered, where the data lives and who is allowed to see it. ### An engineer looks at what’s running Scope comes from your systems as they are, before anything is priced. ### You get a written proposal Covering scope, team, approach and price. [info@smartdata.net](mailto:info@smartdata.net)[(937) 886-9166](tel:+19378869166)