Cohere’s utility AI guide starts with permissions, not model size

Utilities considering AI have a more useful starting checklist than “which model is biggest?” Cohere’s new guidance says to pin down what a system can access, what it is allowed to do, how its answers will be checked and who remains accountable.

Cohere Watch analysis

What happened

The guide sets out five questions for utility leaders: what measurable problem the AI should solve; which systems and actions an agent can reach; what information leaves the organisation; how recommendations will be validated; and whether costs remain predictable after a pilot.

It suggests starting with bounded tasks such as retrieving approved information, comparing documents and preparing recommendations for an employee to review. Access to operational or production systems, it says, should require separate authorisation and approval controls. Cohere also argues that deployment inside a utility’s controlled environment can help keep sensitive data within its perimeter. Read Cohere’s guide to AI for utilities.

Why it matters

Utilities hold maintenance records, emergency procedures and other operational knowledge that could make useful AI search and document tools. But an agent connected to critical infrastructure needs more than a capable model: teams need to know what it can reach, what it can change and how to stop it when something goes wrong.

The guide gives readers practical evaluation points: establish a baseline, check output quality as well as speed, trace recommendations to their source material, assign an accountable owner and expand a workflow only after its value and safeguards have been demonstrated.

Our read

This is sensible advice from a company selling enterprise AI, so its pitch for private deployment is part of the package, not a neutral verdict on every architecture. The stronger takeaway is the operational one: start with a narrow job, limit access and keep a person responsible for reviewing recommendations. “Let the agent figure it out” is not a control framework, however confident the demo looks.

What to watch

• Whether utility pilots publish measured results against a clear baseline.
• How teams restrict and audit agents’ access to sensitive information and systems.
• Whether successful document and search tasks lead to requests for broader authority.

Discussion spark: Should utilities begin with AI that searches and recommends inside tightly controlled systems, or is that too cautious to deliver meaningful benefits?

not affiliated with or endorsed by Cohere