Data Readiness And Governance

Why Data Decides The Outcome

In week 2 you scored feasibility, and data was one of the four questions. In practice it is often the deciding one. An AI tool is only as useful as the information it can read. A model with no access to your price list will invent prices. A forecast built on inconsistent sales records will be confidently wrong. Companies that succeed with AI usually spent as much effort on their information as on the tool itself.

Data readiness means that the information a use case needs can be found, trusted, and used legally and safely. Governance means the rules and the people that keep it that way. This week gives you plain methods for both, and none of them needs a data science degree.

You do not need to fix everything. Fix the data for the one use case you are about to try.

Where Company Data Lives

Most organizations hold information in more places than anyone lists. Start with systems of record, which are the official sources for a type of fact: the ERP (enterprise resource planning) system for finance and supply, the CRM (customer relationship management) system for customers, the HR system for staff. Next come the places where work really happens: email, chat tools, shared drives, wikis and document libraries. Then comes the shadow layer: spreadsheets on personal drives, exports saved for a quarterly report, and notes in private folders.

Add outside sources too, such as supplier portals, bank files and data from partners. Finally, note paper records and scanned images, which hold surprising amounts of history. For each source write down three facts: what it contains, who looks after it, and how someone reaches it. That page is your data inventory, and it is the most useful document of this week.

Structured And Unstructured Data

Structured data fits into rows and columns with fixed meaning, such as an order table with a date, a customer and an amount. Unstructured data has no fixed layout: emails, contracts, call recordings, slides, images. Industry estimates often say that most company information is unstructured, which is why language models matter to business: they are the first widely usable tool for reading it.

The two kinds need different preparation. Structured data needs clear definitions, so that 'customer' means the same thing in sales and in finance. Unstructured data needs to be findable, current and free of duplicates. A common pattern for question answering is retrieval-augmented generation (RAG): the system first searches your approved documents, then gives the best passages to the model so that it answers from them. RAG only works as well as the documents behind it. If your intranet holds three versions of the leave policy, the tool may quote any of them.

Data Quality In Six Dimensions

Quality is easier to improve when you split it into parts. Accuracy asks whether values are correct. Completeness asks whether required fields are filled. Consistency asks whether the same fact is the same everywhere. Timeliness asks whether the data is recent enough for the task. Uniqueness asks whether duplicates exist. Validity asks whether values follow the expected format and rules, such as a date that is a real date.

You can check these with small samples. Pull 50 records for your use case and review them by hand against the six dimensions. Count the problems and note the type. A sample like this takes an hour and often reveals that one dimension, such as missing fields or duplicate customers, causes most of the pain. Record the share of records affected, because that number tells you whether a quick cleanup is enough or a larger project is needed.

Access, Privacy And Permissions

Access answers who may see which data. The standard principle is least privilege: each person and system gets only the access needed for the job. This matters twice with AI. First, a tool that searches company documents must respect the permissions of the person asking. If an employee asks the assistant a question, it must not retrieve a salary file that the employee could not open directly. Second, when you connect a tool to many systems at once, a single broad account can quietly expose everything.

Classify data by sensitivity. A simple scheme has four levels: public, internal, confidential and restricted. Restricted usually covers personal data about customers and staff, financial results before release, and trade secrets. Each level gets rules on where it may be stored and which tools may process it. Week 5 links these levels to KVKK, GDPR and acceptable-use rules. For now, tag the sources in your inventory with a level.

Ownership And Governance Roles

Data without an owner decays. Governance starts by naming people. A data owner is a senior person accountable for a body of data, such as customer records, who decides who may use it and for what. A data steward is the practical caretaker who fixes errors, keeps definitions and answers questions. A custodian, usually in IT, runs the systems that store the data and keeps it secure. One person may hold two roles in a small team, but all three jobs must be done.

Governance also needs a few light routines. Keep definitions of key terms in one shared glossary. Record where each dataset came from and how it changes, which is called lineage. Review access every quarter. Define how errors are reported and fixed. A catalog, which is a searchable list of datasets with owners and descriptions, helps once you have more than a handful. Resist heavy bureaucracy: start with one use case, one owner and one page of rules, then grow.

A Practical Readiness Check And First Steps

Apply the week to your top use case from week 2. Write the data it needs. For each item record where it lives, its sensitivity level, the owner, the quality findings from your sample, and whether the tool may legally and technically reach it. Then mark each item green (ready), amber (usable with fixes) or red (blocked).

Treat red items as work for the plan, not as failures. Typical fixes are agreeing one definition, removing duplicates, moving a key spreadsheet into a shared system, asking an owner for access, or scanning and tagging a set of documents. Give each fix an owner and a date. If a use case has several red items on its main data source, move it from quick win to bet on your matrix. That honest re-scoring is a sign that the method works.

Some fixes are cheap and fast, and they pay off beyond AI. Merging duplicate supplier records improves reporting and payments as well as any model built on them. Building a clear document library helps new staff find answers. When you explain the work to leaders, show these wider benefits, because they make data funding easier to approve.

Key Terms

System of record

The official source for a type of fact, such as the CRM for customer details.

Structured data

Information in rows and columns with fixed meaning, such as an order table.

Unstructured data

Information without a fixed layout, such as emails, contracts and recordings.

Retrieval-augmented generation (RAG)

A method where the system searches approved documents first, then gives the best passages to the model.

Data quality

How well data is accurate, complete, consistent, timely, unique and valid.

Least privilege

Giving each person and system only the access needed for the job.

Data owner

The senior person accountable for a body of data and for decisions on its use.

Data steward

The practical caretaker who maintains definitions and fixes errors in a dataset.

Lineage

A record of where data came from and how it changed along the way.

Exercise: Inventory The Data For Your Top Use Case

1. Take your top quick win from week 2 and list every piece of information the task needs.
2. For each item write the system or file where it lives, who looks after it, and how someone reaches it.
3. Tag each item public, internal, confidential or restricted.
4. Pull a sample of 50 records or 10 documents and check them against the six quality dimensions. Count problems by type.
5. Mark each item green, amber or red and write one fix with an owner and a date for every amber and red item.
6. Re-score the feasibility of the use case and decide whether it stays a quick win or becomes a bet.
7. Share the one-page inventory with the named owners and ask each to confirm or correct it.

Further Reading

• DAMA-DMBOK: Data Management Body of Knowledge, DAMA International
• Data Quality: The Accuracy Dimension, Jack E. Olson, Morgan Kaufmann
• Data Mesh: Delivering Data-Driven Value at Scale, Zhamak Dehghani, O'Reilly Media
• Hidden Technical Debt in Machine Learning Systems, D. Sculley and colleagues, NeurIPS 2015
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, Patrick Lewis and colleagues, NeurIPS 2020