Industrial AI used to answer questions. Now it does the work: it opens work orders, reschedules crews, chases suppliers and commits inventory. When AI starts acting, the question changes from, "Is the answer right?" to "Who authorized the action?"
IFS launched IFS Intelligence at IFS Unleashed in Orlando this week. It is a useful case study because it answers that question with a design principle, not just a feature list. IFS calls the principle “staged autonomy.”
What IFS Intelligence is
IFS describes IFS Intelligence as a product that collects data from across an industrial operation (assets, systems and processes) and turns it into action. It works from live data in IFS and third-party systems and acts only within the rules the business sets.
People in the field, on the plant floor, on a rig, or in the office ask questions in plain language through a single conversational interface. The system reasons across live operational data, explains what is happening, shows its sources, and plans the next step. It can then carry out that plan.
The product offers three ways to work:
- Cowork (ask and act): Investigate a problem in plain language and prepare actions that run only after a person approves them. An answer can be saved as a "Lens," a live view that keeps updating. Work that succeeds once can be saved as a "Skill" that anyone can repeat. Cowork also reaches connected systems such as Microsoft 365 through MCP (Model Context Protocol, an open standard for connecting AI to other software).
- Loops (delegate): Hand defined workflows to "digital workers" (AI agents that carry out multi-step tasks). They run in the background and escalate when judgment or approval is needed. Teams build their own digital workers in a tool called Studio.
- Runway (extend): Use plain-language requirements to speed up migrating, configuring and extending IFS software.
The control layer underneath
All three run on what IFS calls the Industrial AI Harness. IFS says the AI works through the customer's existing identity, permissions, business rules, approvals and audit controls. The AI "cannot and will not be able to see or do something the signed-in user is not authorized to see or do," and no raw data is sent to external AI providers, according to IFS.
The Harness has six parts, each meant to prevent a specific kind of failure:
- Planning and orchestration: breaks a goal into steps and coordinates the agents and tools.
- Skills and connectors: defines and checks which systems an agent may use.
- Memory and context: carries progress forward and pulls in only the information the task needs.
- Guardrails and permissions: enforces identity, approvals and limits on risky actions.
- Verification and escalation: checks the work for correctness and policy compliance.
- Observability and auditability: records every decision and action so it can be traced and reviewed.
Staged autonomy: the idea worth borrowing
Under IFS's description, staged autonomy moves a digital worker from copilot assistance, to recommendations, and then to autonomous action, with escalation whenever the AI is uncertain.
That maps onto how we already treat people. Nobody hands a new hire the keys to the plant on day one. They shadow first, then recommend, then act. Their supervisor can pull them back at any time. Staged autonomy applies the same logic to software that acts on physical operations.
You can see the stages in IFS's own examples. In Cowork, a production manager who learns a large order is at risk can ask one question. The system finds the risk, suggests likely causes, and recommends a next action, which the manager approves and sets in motion. In IFS's examples for maintenance and back-office work, no record change is committed until someone approves it. Loops is the next stage: background work with escalation to a person.
Why the stages matter now
IFS's own research shows why this framing is timely. Its Industrial Index 2026 found that 43% of frontline teams have trained and introduced agentic digital workers in the past year. Only 11% have validated AI outputs for safety-, quality- or mission-critical processes, and only 4% require human oversight or approval for AI-supported decisions. Yet, more than 70% say they are confident in their safety and data policies.
Research from The Futurum Group cited by IFS points the same way: 66% of enterprises plan to invest in digital workers within 12 months, but only 5.7% trust AI to act fully autonomously.
A vendor sponsored that research, and the same vendor sells the remedy, so read it with that in mind. Still, the pattern holds, whoever you buy from. The front line is moving faster than the controls, and confidence in a policy is not evidence of control.
What makes staging real
Staged autonomy is only a control if the stages are real. Whether you use IFS or anyone else, three disciplines decide that:
- Discovery: List every agent, the systems it can reach and the actions it can take. Classify each action by its consequences and whether it can be undone. You can't stage what you haven't listed.
- Agility: Test demotion. Can you move an agent from acting back to recommend-only in minutes, without stopping operations? A promotion you can't reverse is not a control.
- Governance: Put a named owner on every promotion decision, and back it with evidence: error rates, escalation logs and audit trails. "It's been fine so far" is not evidence.
The layer under the layer
Every control described here rests on identity, permissions and tamper-evident audit logs, and all three rest on cryptography. The trust layer for AI agents is the same layer that quantum computing threatens. Organizations building agent governance now should build in crypto-agility (the ability to swap encryption methods without rebuilding systems) from the start.
The technology is new. The discipline is not: know what you have, keep the ability to change it, and make someone accountable. IFS Intelligence puts staging into the product. Making the stages count is still the customer's job.
Edited by
Erik Linask
