Overview and setup

Dataiku AI in Agent Management can operate in two modes:

  • Operated by Dataiku. Processing is done through Dataiku’s cloud-based servers

  • Bring your own LLM. You can choose which LLMs to use for the AI Services, and the processing is done through LLM connections that you configure and control using LLM Mesh. Your input is not sent to Dataiku.

Using AI Services requires acceptance of the Dataiku AI Services Terms of Use.

Requirements

Using the Dataiku AI Services is subject to acceptance of our Dataiku AI Services Terms of Use, which are linked from the “AI Services” page in Admin.

Once you have accepted the Terms of Use, you can turn on AI Services

AI Services require that the Dataiku Agent Management server be connected to Internet, in both “Operated by Dataiku” and “Bring your own LLM” modes. The URL to whitelist is https://ai-gateway.api-services.dataiku.io.

Use of the AI Services involves metered usage. AI Services may include a limited amount of included free usage and/or additional credits subject to fees as set forth in applicable Orders. Please see Usage of AI Services for more details.

“AI Services Operated by Dataiku” mode

By default, AI Services use Dataiku’s own server, hosted and managed by Dataiku. In that mode, your input is processed by Dataiku and by Third Party AI Providers.

“Bring your own LLM” mode

Alternatively, you can choose which LLMs to use for the AI Services, and the processing is done through a LLM connection that you configure and control through the LLM Mesh.

When users make requests to AI Services in this mode, the requests are directly sent to your choice of third-party AI provider. Dataiku (as a company) does not process the requests and does not generate outputs.

Note: Stories AI is not available in “Bring your own LLM” mode.

Choice of LLM

In order to use AI Services in “Bring your own LLM” mode, you must select a LLM. AI Services, notably Cobuild, require high-performance models, able to reliably perform long reasoning and tool calling.

The following models have been tested to provide good results. Note that most of these models can be obtained through multiple providers.

  • OpenAI Astra

  • OpenAI GPT 5.6 Sol and Terra

  • OpenAI GPT 5.5, 5.4

  • Anthropic Fable 5.1 and 5

  • Anthropic Opus 5, 4.8, 4.7, 4.6 and 4.5

  • Google Gemini 3.8 Flash, 3.6 Flash, 3.7 Flash and 3.5 Flash

  • z.AI GLM-5.3 and GLM-5.2

  • xAI Grok 4.6 and 4.5

The best overall results are obtained with OpenAI GPT 5.6 and Grok 4.6.

The following models have been tested to provide “average” results. Using these models is not recommended as experience may vary:

  • OpenAI GPT 5.3-Codex and 5.2

  • DeepSeek V4 Pro

  • DeepSeek V4 Flash

The following models have been tested and do not provide good results. Using these models (as well as all inferior models) will lead to a disappointing experience, and is strongly discouraged:

  • OpenAI GPT 5.1

  • OpenAI GPT 4.1

  • OpenAI GPT 4o

  • Qwen3-235B-A22B

  • Nemotron 3 Super

  • DeepSeek V3.1

More models will be added to these lists. As AI Services evolve, some models may change category. We recommend checking this list regularly.