DataFocus · Journal
Starting from zero: is OT Flash a sensible AI data foundation for a small team?
For a small company with no data platform, OT Flash's one-engine design and one-ontology-per-question approach is a reasonable starting point, provided you budget for servers, model calls and operations.
Yes, with conditions. For a small or mid-size company that has no warehouse, no pipeline team and no semantic layer, OT Flash is a sensible first foundation because DataFocus designs it around one engine (Doris as both compute engine and data warehouse) and one ontology per business question, which keeps the number of moving parts low. The conditions are real: "free" self-hosting still means paying for a dedicated server, model calls and someone to run it. If your data fits in one analytical engine and you can name one business question with an owner, start there; if you already run a Hadoop-class platform or need heavy pipelines, look elsewhere.
What do small teams actually lack?
Most small companies do not lack AI tools. They lack three things underneath them: one place where the numbers live, agreed definitions of those numbers, and a person who owns each answer. Without those, every AI pilot ends up rebuilding context from spreadsheets.
The external evidence points the same way. MIT Project NANDA's July 2025 report The GenAI Divide: State of AI in Business 2025 found that for enterprise-grade GenAI tools, 60% of organizations evaluated them, 20% reached pilot and 5% reached production, and its authors attribute the stall mainly to tools that "don't learn and don't integrate well into workflows". The authors also say these figures are "directionally accurate" and based on interviews rather than audited reporting, so read them as a pattern, not a measurement. BCG's September 2025 study of 1,250 executives (press release, report) put 60% of companies in a laggard group getting hardly any material value from AI.
Neither study is about small companies specifically, and neither proves anything about OT Flash. What they show is that the gap is usually integration and definitions, not model quality. A small team cannot afford to close that gap with a five-component data stack.
Why does one engine reduce moving parts?
A typical "modern data stack" for analytics plus AI has separate pieces for ingestion, storage, transformation, a semantic layer, BI and an AI interface. Each is a contract to maintain, a login to manage and a place where definitions can drift.
DataFocus describes OT Flash differently on its edition page: Apache Doris serves as both the in-memory compute engine and the data warehouse, under the line "One engine. A fast start.", and the edition is aimed at teams that want one business domain running quickly with data that can sit in a single analytical engine. The Flash agent manual adds that Flash queries go through Doris and that Trino, Hive and Hadoop are not part of its agent entry points.
What that means in practice for a small team:
| Concern | What the Flash sources state |
|---|---|
| Where data lives | Imported business data is stored in Doris internal tables (GitHub README) |
| How data gets in | Local CSV, Excel or JSON files, external databases, API and HTTP sources; scheduled table import tasks (r57 manual) |
| Light shaping of tables | Join Worksheet, Merge Worksheet, SQL views, datasets and formulas, which the manual describes as table-level tools, not a full ETL engine |
| Where definitions live | A business definition is set once and used by search analysis, dashboards, control panels, the ontology agent and external agents (product brief) |
| How AI answers are grounded | Each query returns the SQL that ran, the tables used and a data fingerprint, and answers come from SQL that actually ran rather than figures estimated by a model |
| What runs on the box | The monitoring page lists containers such as df-webserver, df-search, df-biserver, df-doris-fe, df-doris-be, df-pgsql, df-redis and df-nginx |
That last row is a useful reality check: "one engine" does not mean one process. You still operate a set of services. But they ship together and are monitored from one System Manager page, which is a different operational burden from wiring and upgrading separate products yourself.
The trade-off is explicit too. DataFocus lists ETL and AutoML for OT Heavy; the product brief shows them as not offered for Flash. Flash still imports and refreshes data, but if your plan depends on multi-stage pipelines, Flash is not built around that.
Why one ontology per business question?
In OT Flash, an ontology is a knowledge space around one business topic. It gathers related tables, business documents, indicators and pinboards, and comes with an ontology agent that answers from that ontology's data and documents.
The product brief's suggested sequence is narrow on purpose: pick one business question that has data, policy documents and a clear owner (order delivery and purchase approvals are its examples), bring the relevant tables and rule documents, build the first ontology, get business users working on it, then authorize an AI agent, and only then move to the next business area.
For a small team this is the right shape. One question means one owner who can say whether an answer is right, one set of definitions to agree, and one small surface to put permissions around before any agent touches it.
What does a first 30-day scope look like?
The plan below is a hypothetical starting scope for a company with roughly one data-minded engineer and one business owner. It applies DataFocus's published getting-started guidance; the week boundaries are our suggestion, not a DataFocus procedure.
- Week 1: define the task. Write down the business question, its business owner, acceptance criteria and permitted scope, as the getting-started guide asks. Example: "Which open orders will miss their promised delivery date this week?" owned by the operations lead.
- Week 1: prepare the server and decisions. Work through DataFocus's preparation checklist: usage scope, OS and architecture, capacity, data authorization, model choice, data egress and backup ownership.
- Week 2: start with synthetic data. The guide recommends it, and says not to enter customer information into the interactive demos. Use this week to learn the ontology modules without risk.
- Week 2–3: connect real sources read-only. The r57 manual recommends a dedicated read-only database account for OT Flash with access only to the databases and tables you want analyzed. Import the two or three tables the question needs and decide how often they refresh, which DataFocus's selection notes ask you to plan.
- Week 3: agree the definitions. "Promised date", "open order" and "late" get defined once in the ontology, with the owner signing off.
- Week 4: business users first, then an agent. Let the owner use search, answers and pinboards. Only when the numbers are trusted, authorize an AI agent on that ontology.
Acceptance at day 30 is simple: the owner gets the same figure from search, a pinboard and the agent, and can see the SQL behind it.
What costs remain with free self-hosting?
DataFocus is direct about this: OT is self-hosted on your own servers, and "You pay for servers, model calls and third-party services." Budget for these lines:
- A dedicated server. System requirements and supported platforms are listed in the Flash release notes; plan resources around data volume, concurrency and model requirements.
- Model calls. Administrators configure the LLM service. If you use an external model, DataFocus says to confirm what is sent to it under your data policy.
- Authorization and setup. The README says new installations require instance-specific product authorization and model configuration.
- Backups. The README describes Doris online snapshots to S3-compatible storage, and notes that restoring interrupts service. Someone has to own and test that.
- People time. Upgrades, connector checks and user support.
Both editions are free to download from the download page, and the installers and release notes are published in the OT Flash and OT Heavy GitHub repositories.
When should you outgrow Flash?
Flash stops being the right fit when the constraints that made it attractive start to pinch:
- Your data no longer fits one analytical engine, or you already run a Hadoop/Hive/Trino platform. DataFocus positions OT Heavy for large data volumes, an existing big-data platform and a need for pipelines and modelling.
- You need built-in ETL or AutoML. DataFocus lists these for Heavy, not Flash.
- You need a platform or high-availability setup beyond what the Flash release notes list.
A decision rule that holds up: if you can name one business question, one owner and fewer than a handful of source tables, start with Flash. If your first question already needs a pipeline team, Flash is the wrong first step.
Next step
Read the edition comparison on the download page and the OT product brief, then write down your first business question and its owner before installing anything.