How Cognite moves AI agents from prototype to production through a graduated trust ladder and built-in governance
Executive summary
Industrial companies operate in high-stakes environments, and Industrial AI must be designed for them from the start. Asset performance, human safety, and regulatory compliance are not negotiable, and a mistake costs far more than a bad answer. As agents spread across industrial operations, enterprises need to move quickly without giving up control or trust.
Cognite Atlas AI is built for that balance. Agents start with narrow, low-risk capabilities and gain authority step by step as domain experts see them prove their reliability. We call this the trust ladder. Governance is part of the platform from the start, so agents can reach production without needing to be rebuilt later to satisfy security and operations teams.
The industrial challenge: from prototype to production
Plenty of AI projects look impressive on a laptop and stall there. Industrial plants hold AI to the same standards as process control systems, and three demands explain most of the gap.
95% of AI prototypes never reach production as successful pilots are notoriously difficult to deploy in a governed way. Industrial organizations hold AI to the same standards as process control systems, and three demands explain most of the gap between a working prototype and a production system.
- Safety and asset exposure. Unbounded autonomous execution creates real operational and human safety risk. A general-purpose agent allowed to write and run any code it likes does not belong in a chemical plant.
- Data isolation and permissions. An agent must see only what its user may see, and it must use that information appropriately to answer questions or carry out the task at hand.
- Governance beyond the model. Generic agent harnesses sit apart from enterprise identity, permissions and operational context. That gap is what keeps a successful pilot from scaling.
Walking up the trust ladder
Rather than granting broad autonomy up front, Atlas AI expands what agents can do one level at a time. Each level adds capability only after the one below it has proven safe and useful, and customers decide how far up they want to go.
- Level 1: Analyze (read-only). Agents search across sources, understand documents and images, run analysis code and answer questions, using the credentials of the person running them. They see only what that person can see, and nothing they do changes anything in the real world.
- Level 2: Use tools and create artifacts. Agents use tools their user is already allowed to run, such as physics simulators, and produce reports, charts or a root cause analysis in Industrial Canvas. They have no path to act on what they build. A person reviews it and decides what happens next. Most industrial organizations are here today.
- Level 3: Trigger safe actions with a human in the loop. Agents can trigger actions in real systems, but a person reviews each one before it executes. Creating a work order notification is the typical case. We're opening this level carefully, one well-understood action at a time.
- Level 4: Closed-loop automation. Agents act toward a goal without approval at each step. This is where we want to go, and we aren't there yet. Doing it safely will take capabilities that don't exist today: fine-grained rules for which actions an agent may take and at what risk level, evaluation tied to promotion decisions, and guaranteed ways to halt an agent and fall back to manual operation. Customers will decide how far to go, and the platform will enforce those limits.
Four Principles for Governance
The ladder only works if the platform underneath it holds at every level. Four principles govern this ladder, and they apply whether an agent is analyzing data or taking an action.
- Secure by design. Running code makes an agent a useful analyst, so model-written code needs a safe home. Each session runs in its own hardware-isolated, read-only sandbox inside the same cloud, region and network boundary as the Cognite Data Fusion project, with outbound traffic restricted by default. Credentials and orchestration logic stay outside it. When agents begin to act, the sandbox gains no new powers. Actions go through separate, governed tools with human approval.
- Human-led control. Domain experts own agent workloads, and agents are granted through standard Atlas AI permission groups. An agent can't approve its own actions or expand its own permissions, and its builder can't change its authority alone. Tools are never assumed. Customers decide which are connected.
- A controlled lifecycle. We strongly recommend development, test and production projects with well-defined promotion processes to production. Builders develop against a snapshot of real production data in development, experts review in test, and only then does an agent roll out to the wider team in production. Builders can unpublish an agent, and administrators can revoke access to agents or models.
- Trusted operations. Customers choose their models and clouds, and their data isn't used to train models. The platform gets regular security reviews and penetration testing, and Cognite is SOC 2 Type 2 certified. Human-in-the-loop controls align with Article 14 of the EU AI Act.
What comes next
The work ahead for Atlas is to unlock more of what agents can do without weakening these industrial-grade controls: finer-grained action rules, evaluation built into promotion, tighter oversight of approvals, and faster ways to halt an agent. Each new capability has to earn its place through evaluation, expert review, and clear boundaries.
These ongoing advancements translate directly into concrete benefits for industrial operations:
- Faster path from pilot to production. Out-of-the-box governance means AI projects avoid getting stalled by IT, security, or compliance reviews, helping teams deliver real-world value sooner.
- Controlled risk and progressive autonomy. Operations teams can safely test and validate AI capabilities at lower trust levels before giving agents higher-stakes tasks or operational authority.
- Enterprise-grade security and governance. Deep integration with existing permissions, isolated sandboxes, and human-in-the-loop controls protect critical industrial assets and ensure regulatory compliance.
Success in this market takes more than a language model in a chat window. It takes an industrial agent harness that respects operational security, honors domain authority and scales safely. That's what Atlas AI is built to be.