Eight things you may be living with, and what each becomes.
Each opens with the situation as we hear it, says what changes, and ends with a way to prove it on your own data rather than take our word. Skip to the one that is yours.
Concern 1
Two systems, one truth, never in step
What you live with
The database is the truth and the search engine is a copy, and everything between them is your code. The copy lags, the queue backs up, a deploy drops a message, and the first anybody knows is a customer who cannot find a record that exists.
What changes
Every index is built by the engine that holds the records, inside the same transaction. There is nothing to keep in step, because there is only one thing.
Prove it yourself: Write a record through the API and search for it by keyword, by meaning and through a relationship in the next statement. There is no delay to measure because there is no second system to be behind.
Concern 2
The AI cannot see the data, and what it sees is stale
What you live with
Every assistant, copilot or agent project begins by copying the data somewhere the model can reach it. The copy is a second security surface, a second cost, and a second thing to keep current - and it is out of date by the time the demo runs.
What changes
The assistant sits on the same surface as the rows. A question is answered from the database as it is now, with its evidence cited, against your own model and your own key - and you can read exactly what would be sent before it is sent.
Prove it yourself: Ask a question through the console, read the citations, then call /query/prepare and read the exact submission the model would have received.
Concern 3
The cost is unknowable until the invoices arrive
What you live with
Six vendors bill in six units. A search vendor charges per search and per record; a database per compute hour and per gigabyte; a model provider per token. Nobody can say what a feature will cost before it ships, and finance reconciles a month late.
What changes
One published unit - 2 cents per thousand - and every other rate on one card you can read on the console. What an ingest actually costs is checked against your budget at every stage, and work stops above 95% of it. Usage accrues live, itemised by meter, and the bill is those figures.
Prove it yourself: Open the pricing page in the UI, then run an ingest and watch its charges accrue, meter by meter, as it runs.
Concern 4
Every new kind of data is a new vendor
What you live with
Audio needs a transcription API, video a frame service, PDFs a document-AI product, relationships a graph database. Each is a contract, a credential, a data-processing agreement, and one more copy of your data outside your control.
What changes
Audio, video, images, thirty-one document formats and tabular files all arrive through the same pipeline into the same database - transcribed, read, understood and indexed. Each is a line on the same bill, not a new contract.
Prove it yourself: Drop an hour of audio, a slide deck and a CSV into one submission - or a zip of all three - and query all three in one statement afterwards.
Concern 5
Eighteen languages of users, one language of search
What you live with
Translation runs after the fact, if at all; the search index only ever saw the original; a customer searching in their own language finds nothing and concludes the product is not for them.
What changes
Text is translated on write into up to eighteen languages, stored beside the original and indexed in each, so a search in any of them finds a document written in any other. The translator is priced once per gigabyte of source, however many languages you choose.
Prove it yourself: Write a record in one language and search for it in another, in the same session.
Concern 6
The security review is where a small vendor loses
What you live with
Who may see which column, who saw it, what left the building and for whom - answered in slides, promised in a roadmap, and never demonstrated. The review stalls, and the project with it.
What changes
Rights down to a single column, for a person or for a group your own directory asserts at sign-in, enforced identically on every door; residency per database that refuses a write outside its region; and an audit line for every act that nothing can remove.
Prove it yourself: Seven demonstrations, run live in week one of the proof of concept, in the order your reviewers will ask for them. Read them on the proof-of-concept page.
Concern 7
The surface is HTTP and the team wants a package
What you live with
Every example is a curl, every integration begins with someone writing the client again, and the AI project begins with an export because nothing in the framework knows the database.
What changes
Packages for .NET, Python and JavaScript with the streaming, the errors and the retries already handled; LangChain and LlamaIndex retrievers that ask the database for passages instead of exporting it; and an MCP server that gives an agent four tools on its own key. An Entity Framework Core provider shipped in September 2026.
Prove it yourself: pip install dataoil, then stream a query; add @dataoil/mcp to Claude Desktop with one configuration line and ask it about your data.
Concern 8
Moving in is easy; moving out is the risk
What you live with
Every platform is easy to join. The question a careful buyer asks is what it costs to leave: proprietary query languages, formats nobody else reads, exports that take a week.
What changes
The query language is SQL. A PostgreSQL port serves the same databases to the tools you already own. An export is one call, the whole state of a database is files in object storage you can read, and a backup restores anywhere - which is also how you would take everything with you.
Prove it yourself: Connect psql to the PostgreSQL port and run the same query the API answered. Then export the database and count the rows.
When two of them are yours
That is enough to write the goals for a proof of concept.
Four weeks, your data, the goals agreed in writing in week one, and a written answer either way.