DataFocus · Journal
How to pick and scope the first business question for an OT Flash pilot
A checklist for choosing one business question for a first OT Flash pilot: data, policy documents and a named owner, written acceptance criteria, synthetic data first, and a go or no-go rule.
Pick one business question that already has three things: data you can name, policy documents that define the terms, and one person who owns the answer. Write down the acceptance criteria and the permitted scope before you load anything, and start with synthetic data. DataFocus's own getting-started guidance says to define the task, its business owner, acceptance criteria and permitted scope first, and to start with synthetic data. If the question fails any of those tests, choose a smaller one.
Why scope the first question this tightly?
Unclear value is a known risk for AI projects. In June 2025 Reuters reported, via The Star, a Gartner prediction that more than 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs and unclear business value. That is a prediction about agentic AI projects in general, not a measured result for any product, and it says nothing about OT Flash. It does support one practical habit: decide up front what "worked" means.
A first pilot on OT Flash is not a rollout. It is a test of one question against one set of data and documents, with a result that someone can accept or reject.
What makes a good first question?
DataFocus's product brief suggests starting with a business question that has data, policy documents and a clear owner, and gives order delivery and purchase approvals as examples. Use this checklist.
- It has tables. The data the question needs sits in a few tables you can export or connect.
- It has documents. The meaning of the key terms lives in written policies, rules or metric definitions, not only in people's heads.
- It has one owner. A named person decides whether an answer is right and settles disputes over definitions.
- It recurs. Someone asks it every week or month, so a repeatable answer has value.
- It is narrow. One business area, not "all of finance".
- A wrong answer is recoverable. For the pilot, the output informs a person. It does not trigger a payment or an order change.
The reason for items 1 to 3 is how OT Flash is organized. The r57 manual describes an ontology as a knowledge space around one business topic that brings together related tables, business documents, indicators and pinboards, with an ontology agent that answers from the ontology's data and documents (OT Flash r57 user manual). A question with no documents, or no owner for them, leaves the agent with little to cite.
A hypothetical example of a well-scoped question
Consider a hypothetical mid-size distributor. This is an illustration, not a DataFocus case and not a customer.
The finance lead wants to know: "Which purchase requests above an approval threshold skipped a required review step last month?" The data is two tables (requests and approval events). The rule lives in a one-page approval policy. The finance lead owns the rule and the answer.
Compare that with "improve our procurement." There is no owner of the answer, no document that defines success, and no way to say when it is finished.
How do I write acceptance criteria and scope?
Write these before the pilot starts, on one page. The structure below is an illustrative template, not a DataFocus form.
| Item | What to write | Hypothetical example |
|---|---|---|
| Question | One sentence, with the time window | Which requests over the threshold skipped review last month? |
| Owner | One named person | Finance lead |
| Data in scope | Table names, nothing else | Requests, approval events |
| Documents in scope | File names | Approval policy v3 |
| Terms to define | The words that could be read two ways | "Threshold", "skipped", "last month" |
| Permitted scope | Who may see which data during the pilot | Finance team only |
| Accept if | Checks the owner can do by hand | Owner's own list for ten sampled requests matches |
| Reject if | A clear failure condition | Any listed request that the owner cannot trace to a source row |
The "Terms to define" row matters most. DataFocus's guidance warns against equating production completion, dispatch and customer acceptance without confirmation, which is the same issue: if two people mean different things by one word, no tool can settle it for them. Get the owner to settle each term in writing, and put that text in a document the ontology can hold.
Why start with synthetic data?
DataFocus's getting-started guidance says to start with synthetic data, and not to enter customer information into the interactive demos. The demos themselves are different from a pilot: the four homepage scenarios use simulated data in the browser, change data only in the current page and reset on refresh, according to DataFocus's demo notes. They are for looking, not for running your question.
For your own environment, build a small synthetic version of your tables first: same columns, invented values, a few rows you have planted on purpose (one request that skipped review, one that did not, one with an ambiguous date). Then you know the right answer before the agent gives one. The r57 manual itself uses sample data whose names start with Manual_Sample_, which is a useful naming convention to copy so nobody confuses test tables with real ones.
Move to real data only after the synthetic run passes your criteria, and only for the permitted scope you wrote down.
What does the run-through look like?
DataFocus's building guidance suggests a sequence: inspect the version and capabilities, inspect sources, confirm meaning, draft, review the diff, test, obtain authorization, publish and read back, then verify a business task. For a business lead, the parts that matter are these.
- Create the ontology. In r57 this takes a name and an icon. Then you import the related tables in Data Tables and upload the business documents in Document Center.
- Wait for documents to finish processing. A document's status must be Completed before the agent can search and cite it.
- Ask your question in the Agent module. An LLM service must have been configured by an administrator first. The manual says an answer usually takes tens of seconds, longer for complex questions.
- Read the evidence, not just the answer. The answer states which tables the data came from, and "View processing details" shows which data the agent queried and which tools it called. DataFocus's brief says each query returns the SQL that ran, the tables it used and a data fingerprint. Check the SQL against your definition of the terms.
- Compare against your hand-made list. Use the planted rows from the synthetic run, then the sample you set in the acceptance table.
Two cautions from the manual. Executions are Python scripts that really read or change data; it advises saving one only after a test run confirms correct results, and granting it only to people who need it. Control-panel action buttons perform real business actions on the current object, so confirm the object first. For a first pilot, leave both out. Questions and answers are enough.
How do I decide go or no-go?
Decide on the criteria you wrote down, not on how the demo felt. A simple rule:
- Go to the next step (more data, more users, a second question) if the owner's checks all pass and every wrong answer can be traced to a missing definition or a data problem you can fix.
- Stay in the pilot if answers are right but the owner could not trace some of them to a source.
- Stop if the owner rejects results and the cause is not a definition or data fix, or if the question turns out not to have an owner after all.
Also note what the pilot did not test. DataFocus's brief states that OT is self-hosted and that administrators configure the model service; with an external model, confirm what is sent to it under your data policy. DataFocus's deployment notes say free self-hosting still needs infrastructure, model APIs and operational support. You can get OT Flash from the download page or the GitHub repository, where the release notes list system requirements and supported platforms. Read them before you plan anything beyond a trial, and treat an IT-led install and a production decision as separate steps.
When this approach is not the right fit
If the question has no owner, no written definitions, or its data cannot be exported at all, a pilot will not fix that. Settling the definitions is work to do first. If your priority is large pipelines or modeling across a big-data cluster, the brief lists ETL and AutoML for OT Heavy rather than Flash, so read the edition comparison before choosing.
Next step
Read the getting-started guide and fill in the one-page table above for your candidate question. If you cannot fill every row, pick a different question.