Clean data is what a redesigned operation produces, not what it needs first
OpenAI and Databricks now bring ChatGPT to governed company data. Most companies will read that as "fix the data first". The order runs the other way.
Waiting for clean data before starting with AI is the most expensive way to start. Data is messy because the operation that produces it is messy, so a data project that leaves the operation untouched gets dirty again by next Monday. Clean, trusted data is what a redesigned operation produces as it runs, not a gate you clear before the redesign begins.
What the OpenAI and Databricks announcement shows
On 13 November 2026, Databricks published how companies can use OpenAI's models and ChatGPT on their governed data. Employees can ask questions of company data from ChatGPT through the Databricks Genie plugin. In the post's words, "Genie applies the company's own definitions, so revenue in ChatGPT matches revenue in the finance reports," and "Users only see data they already have permission to see, and each answer links to its sources in Databricks."
OpenAI's president, Greg Brockman, is quoted describing AI without company data as "like hiring someone with 20 PhDs and never inviting them to a meeting."
It is a useful setup, and it works because the hard part is already done: one definition of revenue, permissions that mean something, sources an answer can point to. That is the part worth paying attention to.
The easy reading is "fix the data first"
The easy reading for every company without that foundation is a sequence: first a data programme, then a warehouse, then definitions, then, one day, AI. It sounds responsible. It is how many AI plans stall before they reach a single process.
We disagree with the order. Bad data is a symptom. The cause is an operation where nobody records the step, the stage or the owner at the moment the work happens. A case starts on WhatsApp, moves to an email, gets an update in a spreadsheet and closes in someone's memory. You can clean last year's records perfectly, and this week's cases will be recorded the same way as before.
A cleanup fixes the history. Only a change in how the work runs fixes the next record.
Where trusted data actually comes from
When an operation is redesigned, each step writes its own record as a side effect of doing the work. Nobody has to remember to update a sheet, because moving the case forward is the update.
- 01Start from the questionThe decision someone has to make
- 02Map the evidenceWhere each part of the answer lives today
- 03Record at the stepEach step writes its stage and owner
- 04Fix one definitionOne owner per term the question uses
- 05Answer and checkEvery answer points to its source
Each new question extends the same records instead of starting a new data project.
Three things change when you work in this order:
- Scope. You clean and connect the data one question needs, not the whole company's history.
- Ownership. Each record has a step that produces it and a person who answers for it, so it stays right after launch.
- Proof. The first useful answer comes early, and it tells you which data matters next.
The question decides which data has to be trusted. Without a question, "clean data" has no finish line.
Start from the question: the OCESA case
Sentynel's OCESA case, built with Smart Sponsorship for live entertainment in Mexico, started from a question rather than a data inventory. After an event, the evidence for its results sat in too many places: ticket sales, sponsorship results, ratings and media evidence. The case record describes the problem as "One event. Too many places to find the answer." and the system as "Connect the evidence to the question."
Five evidence sources were connected into one reporting operation: ticketing, sponsorship evidence, ratings, presentations and media files.
Source: Source: Sentynel project case record. Results are reported project outcomes; the cost comparison is normalized.
The case record reports a −70% reported operating cost. Results are reported project outcomes; the cost comparison is normalized, indexed to a starting cost of 100, not absolute costs.
The point for this argument is the order. The work did not begin with a company-wide data programme. It began with the question the business kept asking after every event, and the data that answer needed was connected and kept current by the reporting operation itself.
What to do on Monday
Do not start by asking whether your data is ready. Ask a narrower question:
- Write down one decision your team makes every week with numbers it does not fully trust.
- List where each piece of evidence for that decision lives today, including the spreadsheets and the inboxes.
- For each piece, name the step in the operation where it should be recorded, and the person who owns it.
If you can name the step, the data can be fixed inside the operation, and an agent can work with it from the first day. If you cannot name the step, no warehouse will help: the record has nowhere to come from, and that is where the redesign starts.