Common questions
Questions we get asked
Short, honest answers. If yours is not here, ask.
Does the AI see all our data?
No. Access is enforced beneath the AI, at the row and the column, before the model is involved. It writes the question; it does not decide what you may see.
Do we have to move our data to you?
No. One company, one source database, one deployment, running where your data already lives. It is not a shared cloud service.
Does it work without an internet connection?
An evaluation copy does, once installed -- Docker Desktop and nothing else. It even works before you give it an AI key, using questions that have verified answers stored with them. Running an open model on that same machine keeps every live question inside the building too.
What does hosting cost?
You pay for the computer it runs on, not for each person who asks. A small cloud machine is on the order of a hundred dollars a month. A box you own, with an open model, is a once-off purchase and then electricity. The comparison is on Hosting costs.
Is there a charge per user?
No. One company, one deployment. Cloud or a machine in the office -- neither adds a cost when another person starts asking.
What does a question cost?
You choose the model and see its per-question cost and privacy implication before you commit. Every question is then recorded with what it actually cost. Models that fail an accuracy floor are not offered.
What if the question is ambiguous?
FactSmith asks which date or measure you meant, rather than guessing. You answer, then it runs.
Is FactSmith just a chatbot on a database?
No. The trust comes from a governed layer of business meaning and confirmed facts, not from the model. That layer is the product.
What does FactSmith itself cost?
It is sold per deployment, to one company at a time. We will give you a real number on a call rather than post a figure that changes next month.
Who is it for?
Distributors, multi-store retailers, manufacturers and franchise networks with real operational data -- and no dedicated analytics team. Large enterprises that already run specialist teams are not the first customer.
Short sector pages: Distributors, Retail, Manufacturing, Franchise.
How long to stand up an evaluation?
An evaluation copy is a folder and an installer. It needs Docker Desktop and nothing else. It works before you give it an AI key, using questions that have verified answers stored with them. The facts library and example dashboards are already seeded, so the first session is not a blank product.
Which warehouses does it read?
Dimensional, Kimball-style models. The evaluation copy runs against PostgreSQL. SQL Server is on the adapter path for a customer warehouse; that path is not what the package proves end to end today. If the schema is not dimensional, it says so rather than guessing.
Do we need Easy Data first?
No. Ask is what is available now. Easy Data is in early development: it is how messy sources will become that dimensional foundation. The two are halves of one job, and they are not the same stage.