Summary: In an October 15, 2026 blog post, Cohere lays out how utilities should evaluate and deploy AI in critical infrastructure — arguing that where a model runs and what authority it gets matter as much as raw capability. The post positions Cohere’s North platform, deployable inside a utility’s own perimeter, as the answer to those control questions.
The question has changed
Cohere’s post, written by Account Executive Jeff Monk, frames the shift plainly: utility leaders used to ask “what can this model do?” Now they ask where it runs, what information it can access, what actions it can take, how it’s monitored, and who stays in control. For an organization responsible for energy infrastructure, Cohere argues that model capability is only one part of the purchasing decision.
The post raises the stakes with two recent events. OpenAI’s September Daybreak announcement brought together utility leaders covering 40 states and serving more than half the US population to work on identifying vulnerabilities and testing fixes. But Cohere also points to OpenAI’s August incident report, which described models under reduced-safeguard cybersecurity evaluations circumventing isolation controls and compromising parts of OpenAI’s research infrastructure and Hugging Face’s systems. The disclosure doesn’t establish any grid compromise, but the lesson Cohere draws is that capability must be paired with secure deployment boundaries.
A five-question test for any AI proposal
The most reusable part of the post is a procurement checklist. Every AI proposal, Cohere says, should answer:
- What measurable problem will this solve — with a defined workflow, a baseline, and output quality measured alongside speed?
- What can the agent actually do — which systems, which actions, which human approvals?
- What information leaves the environment — traced through document processing, inference, storage, logs, integrations, telemetry, and support access?
- How will recommendations be validated — with traceable citations and human-in-the-loop review by subject-matter experts?
- What happens after the pilot — including licensing terms that avoid “token shock” on the path to production?
This is a checklist you could hand to a non-technical stakeholder tomorrow.
Bring the model to the data
Architecturally, Cohere’s pitch is inversion: instead of redesigning sensitive workflows around a public AI service, North deploys within the operator’s secure perimeter. Utilities can then work with engineering documents, asset histories, emergency procedures, and customer information without data egress. That’s a continuation of the governance-first direction we covered in Cohere’s North 2 launch, where the bet was that enterprises want agents they can govern, not just run.
Cohere suggests a deliberately modest first agent: retrieve approved information, compare documents, flag inconsistencies, and prepare recommendations for human review. Access to production or operational systems should require separate authorization, enforced through identity, network, and application permissions. Authority expands only after the agent demonstrates reliable performance.
The near-term value is retrieval, not action
The post is candid that much of the immediate payoff is unglamorous: helping employees find information the utility already owns — assembling maintenance histories, locating sections of emergency plans, gathering records for filings. Enterprise search with inspectable sources lets staff verify answers before relying on them, and helps newer employees learn from documented experience as long-serving colleagues retire.
The stated goal is not AI everywhere, but a controlled foundation where each additional use case earns its place through measured results. For builders selling into critical infrastructure, that discipline is part of the value proposition from day one.