Frequently asked questions

You do not set it up, we do. We connect DEHA Lab to your real sources and walk you through all of it together. If you choose the box, we bring the DEHA box to your building and install it. Keeping it running is our side of the deal.

In our first enterprise proof of concept, twenty-five business processes were defined and taken live within a few weeks. Nothing new was written: the company's processes were described on top of the existing layers. Setting it up and keeping it running is our side of the deal, and a second business area goes at the same pace.

No. Agents, automations and data flows are built from one place. If you do want your own code, you can bring your own Python, run it in an isolated sandbox and use it inside those same flows.

They get a corporate counterpart. You edit a cell and every column that depends on it recalculates instantly, an agent can read and edit the same sheet, and what-if scenarios are layers rather than copies, so the real data is never touched. In our first enterprise deployment, an Excel universe of half a million cells moved onto the operating system.

In those, the model is the product. Here the model is deliberately not the center: the heavy lifting runs on engines that give the same answer to the same input, and the model is left with what it is genuinely good at, turning your sentence into a structured request and the findings into readable text. That is why answers here carry sources, follow your company's own vocabulary and can run inside your own building.

We do not promise zero hallucination, because that promise cannot be made. What we do is close the roads it travels one by one. İDRAK holds one map of what your data means, and a question that falls outside it does not get an invented answer: the system asks what you meant, or says it cannot answer this one. Every answer it does give arrives with the tables and records behind it, so you check the source instead of trusting the tone.

Yes. You can bring your own model, including an open model running on your own hardware, and your own vector database. The operating system gives that model everything it lacks on its own: memory, conversation management, the meaning map and permission bounds. And if you choose a big cloud model anyway, personal data is masked before anything leaves.

Turkish is a first language here. Matching reads Turkish letters right, so İADE and iade are the same word. The personal data check recognises and validates a Turkish ID and tax number, reports print Turkish characters right, and Logo, Mikro, Netsis and Nebim are among the systems it reads.

No, and we do not sell it as that. It is a complement: repetitive, pattern-bound work shifts onto the operating system, while every step with a real-world consequence waits for a person's approval. You can delegate work, you cannot delegate responsibility. Work handed to a person takes the responsibility with it, work handed to a machine leaves it exactly where it was, which is why a person stays in the deciding seat. What the machine suggests and what a person has confirmed are kept apart in the system, and the machine may never overwrite the person.

In your own systems. The operating system keeps only the operational layer: what ran, when, and what your data means. The business data itself is never stored on our side, and the results are written back to your own systems. If the data may never leave your building, the same system runs on the DEHA box, inside it.

Databases, cloud storage, files and Excel, business systems and APIs, with incremental and live syncing. More than 40 ready-made connections, SAP included, over OData, HANA and RFC. Turkish ERPs such as Logo, Mikro, Netsis and Nebim are reached through their own databases. Existing dbt projects run inside the operating system as they are, and the open web is a source too.

Every action, result and AI decision is traceable from one console, which is also what the EU AI Act asks of high-risk systems: human oversight and records. Changes are versioned with one-click undo, sensitive steps wait for a human approval, and every answer shows its sources.

Every run is counted, not estimated. You set a budget, get warned at your threshold, and choose whether an overage throttles or stops the work.

No. Everything you build is versioned configuration: agents, flows, connectors and dashboards export as one package and move to another install. Each of them is also an API and an MCP tool, so your own systems and even other AI assistants can call what you built here.

You connect Claude only if you want it to write your setup, while the AI that answers your questions runs inside the box without the internet.

A model alone does not know your company. It has no memory, starts every session from zero, cannot reach your data and does not know who may see what. We take on everything around that engine: memory, sessions, what your data means, the tools it may call and the limits it cannot cross. That is how a local model becomes genuinely useful, and the intelligence you build stays yours from day one.