Mistral AI's Aaron Delp argued that enterprises get ROI from AI they own, trained on their own data and deployed where they choose. He walked through Mistral's approach to a first use case and pilot, then demoed admin controls, workflows and skills in Mistral AI Studio.
Key takeaways
- Own your AI, don't rent it. Delp argued that control of data, weights and keys leads to lasting ROI.
- Benchmarks aren't business outcomes. 'Leaderboard maxing' doesn't integrate the domain knowledge enterprises need, he said.
- Start with one measurable use case. Mistral aims to finish feasibility pilots in three to four weeks.
- Admins need real controls. Audit logs, org-wide connectors, feature toggles and token limits were shown in the console.
- The hard part is fit. The biggest agent orchestration problem is fitting model, harness and workflows together.
Enterprises should own their AI rather than rent it, Mistral AI's Aaron Delp told attendees of a Data Science Connect webinar sponsored by Mistral AI. "This isn't renting AI. This is owning your AI," said Delp, Global Head of AI Technical Marketing at Mistral AI, "full stack top to bottom."
The session, "The Future of Agentic AI: Scaling Efficiency and Control with Mistral's Open Source Platforms," set out to show how a team could get started with agentic AI in as little as 10 to 15 minutes, using Mistral's AI Studio.
Why don't off-the-shelf AI tools deliver enterprise ROI?
Delp said off-the-shelf AI often optimizes for the wrong outcomes. He pointed to "leaderboard maxing," the race to top benchmarks for a day or two, and asked what that means for business results. He cited the widely referenced MIT study that found about 95% of enterprises don't see the return they want from AI.
The reason, he said, is that generic AI doesn't integrate domain-specific intelligence or offer deployment flexibility. Mistral's pitch is an "AI flywheel," in which an organization's own data and expertise feed models that are continually pre- and post-trained, served, measured and improved, like an application moving from version 1.0 to 1.1.
He stressed deployment choice for organizations with data sovereignty and compliance concerns: on-premises, edge or any cloud, with the customer owning "the data, the weights and the keys."
What is in Mistral's enterprise AI platform?
Delp described a build, activate and realize sequence. Building happens in Mistral AI Studio and Forge, which he described as a subset of Studio for fine-tuning and training Mistral's generalist and specialist models on a customer's workflows. Activation happens in Vibe, Mistral's interface for business work and for code. Value is realized through custom solutions, such as customer support agents, contract drafting and compliance automation.
He named customers in regulated and demanding sectors. Mistral has worked with Cisco on agentic workflows for in-house customer service systems, and with HTX, part of Singapore's Home Team, on a co-developed model family fine-tuned for the national language and for answering and routing government service queries. He also mentioned work with the European Space Agency on images and documents.
How should enterprises start with agentic AI?
Start with one "iconic use case," Delp said, chosen with subject matter experts and tied to a measurable business outcome. "If there's not a measurable business outcome, none of this matters," he said. "This is all technology for technology's sake."
Mistral then runs what it calls tech feasibility, a pilot it aims to wrap up in three to four weeks, including data mapping and targeting the deployment environment that fits best. After sign-off, the work turns to production and scale.
Proofs of concept are notorious for working on controlled data with controlled tools, he said, then struggling when usage grows from hundreds to thousands or millions. Studio is meant for that stage, with workflows, an AI registry, a context engine, model customization, observability and evaluations. Once the first project is built, "the acceleration is in the repeatable model," he said, with traceability and reuse across a unified catalog.
What does an agentic AI admin console need?
In a live demo of the Mistral console, Delp started with administration. Administrators can create organizations, members and workspaces, and review audit logs that show whether a human, a skill or something else ran each action. Connectors such as GitHub, Slack and Gmail can be enabled once for the whole organization.
He turned off Vibe's work mode, which handles agents and multi-step tasks, and showed the tab disappear for a user, then restored it. Administrators can also cap token usage per second or per minute and disable file downloads, which he presented as a check on "token maxing."
How do AI workflows and skills work in Mistral Studio?
In Studio, Delp showed speech-to-text, text-to-speech and document AI for OCR, available in the interface, in batch and through the API. He ran workflows from Mistral's cookbook examples, including insurance claims, cargo release compliance, dangerous goods validation and a long-running code modernization task that edits files one by one. Workflows can be connected to GitHub or run on schedules. He said he set up the examples in 10 to 15 minutes.
He also created a simple "hello world" skill as an administrator, published it to Vibe, and ran it as an ordinary user.
During Q&A, asked how Mistral compares with Qwen models for local use, Delp said it depends on the size, features and optimizations a team needs. Mistral has 20 to 30 models across generalist sizes and specialist uses, he said. Asked about the biggest problems in orchestrating agents, he pointed to fitting the model, the agent harness and agent workflows together. "The easiest way to think about a harness is a wrapper around the models," he said.
See more agentic AI demos
Agentic AI Showcases and AI Demo Days are among four series in Data Science Connect's 2027 schedule of 24 free live webinars. See the webinar schedule for upcoming sessions.
Frequently asked questions
What is Mistral AI Studio?
Mistral's platform for building, deploying and managing AI agents, workflows and skills, with observability, evaluations and model customization, as Delp described it.
How do you start an enterprise agentic AI project?
Delp recommends identifying one use case with a measurable business outcome, then validating it in a short pilot, which Mistral aims to complete in three to four weeks.
What is an agent harness?
Delp described it as a wrapper around a model that orchestrates and manages agents.
Can Mistral models be deployed on-premises?
Delp said Mistral's stack can be deployed on-premises, at the edge or in any cloud, including self-hosted and hybrid environments.
Watch: The Future of Agentic AI with Mistral AI: Scaling Open-Source AI for the Enterprise | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
