Context engineering for enterprise data

What is context engineering?

Prompt engineering was about the wording of one request. Context engineering covers everything around it: which documents to retrieve, which data to query, which tools to call and which instructions apply. Models have become capable enough that wrong answers usually trace back to missing or wrong context.

Why does AI get enterprise data wrong?

When AI gets your data wrong, the model is rarely the problem. It can't find the right data, misreads what it finds, and doesn't know your rules.

• It can't find the data. AI finds data through table names, column names, descriptions and sample values. If the same customer is KUNNR in SAP, CustomerID in your CRM and client_no in a spreadsheet, it misses data that's right there.

• It misreads what it finds. A status column holding 1 to 5, an amount with no currency or VAT basis, or dates stored in two formats tell it nothing about meaning, so it guesses.

• It can't see your business logic. None of this is in the data: which of three revenue definitions finance uses, that intercompany sales are excluded, or that a supplier needs a valid VAT number.

• It inherits the errors. Where the data breaks your rules, through duplicates, missing fields or values that conflict across systems, the AI repeats the mistake.

Why is enterprise context so hard to get right?

Nobody designed this. Data estates grow one decision at a time, and each decision was reasonable when it was made.

• Systems get added for urgent needs, and each comes with its own definitions.
• Acquisitions bring their own ERPs, codes and definitions. The integration budget goes to keeping the business running, not to harmonising data.
• Migrations move the data but not its meaning. Old and new systems run side by side for years.
• The rules live in documents and people's heads. When people leave, the knowledge leaves with them.
• Data teams are small and stuck firefighting, so root-cause work never makes it to the top of the list.

Why don't prompts, RAG and catalogs close the gap?

Data teams know about this gap. They haven't closed it because capturing and agreeing on meaning by hand costs more than any team can afford, and the decisions still need people.

• Manual work (workshops, spreadsheets, consultants) is accurate for a moment. It gives you a snapshot that goes out of date and rarely scales beyond one domain.

• Catalogs and governance platforms store, govern and serve what people put into them. They assume someone fills them, and filling them is the hard part.

• AI assistants and home-built RAG retrieve what's written down. They can't decide which version is right, they inherit the conflicts, and their answers can't be traced to an agreed definition.

What does context engineering take for enterprise data?

1. 01

   Map what you have. Find the metric definitions, business definitions, business rules and data-quality rules in your data and documents, and link each one to its source and the fields behind it. If you can't see it, you can't fix it.

2. 02

   Harmonize the definitions. Where the same concept is defined differently, agree on one version. The inputs are existing definitions, documents, regulations and patterns in the data. The owners decide, starting with the critical 20%.

3. 03

   Resolve what matters first. Check the data against the agreed rules, rank what breaks by business impact and effort, and fix it at the source step by step. Automation makes this realistic: in one Metagem POC, it surfaced 445 discrepancies in under six weeks and replaced more than 120 person-days of manual discovery.

4. 04

   Sync it to your AI. Give agents the agreed definitions, rules and metric calculations, with their source and approver, through MCP, and let them query your data on those definitions. Push the metric definitions to your BI tools as well, so dashboards and agents calculate the same way. Clean data still needs this, because it doesn't say which definition applies or what a code means.

5. 05

   Keep it current. New systems, acquisitions, documents and regulations change your context. Re-check on a schedule and whenever a source changes.

Where do you start?

Start with one domain and one question the business keeps getting wrong answers to. Collect the definitions behind that question, find where they disagree, and get one owner to decide.