Cohere argues that companies are treating AI too much like ordinary software, when the bigger challenge is redesigning work around systems that can collaborate, act and change how decisions are made. Its practical message is clear: define what AI may do, what people must still decide, and who remains accountable when the output is wrong.
Cohere Watch analysis
What happened
In a new company article, Cohere sets out an approach to AI change management built around human-AI workflows rather than a one-off installation. It says organisations need to rethink roles, train people in new ways of working, communicate expectations and adapt oversight as adoption develops.
Cohere divides work into three useful categories. Verifiable tasks, such as code that must pass predefined tests, can be delegated more readily. Judgment-based work, such as choosing between business strategies, still needs a person to make the call. Hybrid work, including market research and business proposals, mixes both and needs continuing review.
The company also recommends giving AI relevant organisational context, using human-in-the-loop escalation for missing information or approval, and applying least-privilege access. It says organisations should monitor performance, behavioural drift and bias after deployment, rather than assuming governance ends when the tool goes live.
Why it matters
This is more than a piece of management vocabulary. If AI changes who prepares evidence, who checks it and who makes the final decision, then a conventional software rollout plan will miss the part where responsibility quietly moves around the room.
Cohere says employees may become managers of AI-assisted work, judging quality and deciding when human input is needed. People building internal AI tools may also become miniature product managers, responsible for users, business outcomes, maintenance and governance. That is a sizeable change in the job description, even if the organisation still calls it an efficiency project.
The company’s own framing is advocacy, not independent proof that these methods work everywhere. But it gives readers concrete questions to take into an AI deployment: can the task be checked objectively, does it require human judgement, and what happens when the system asks for more access?
Our read
Cohere is right to push the conversation beyond adoption figures and prompt-count theatre. The useful measure is not how many tokens a team consumes, but whether the resulting work is better, safer and properly owned.
The strongest idea here is to govern AI at the use-case level. A human-operated drafting tool should not face the same controls as an always-on autonomous system. That sounds obvious, which is usually a sign that somebody has had to write it down before a committee forgets.
What to watch
- Whether businesses publish real evidence of quality and outcomes, rather than usage totals alone.
- How least-privilege access and approval controls work in live agentic systems.
- Which roles gain responsibility for reviewing AI output, and whether they receive the training to do it.
- Whether regulators and internal auditors accept use-case-specific governance as a credible alternative to blanket rules.
Discussion spark: Should organisations measure AI adoption by usage, or should employees be judged mainly on the quality and consequences of the work produced with it?
Sources and evidence
- Source update (22 September 2026, 17:45 UTC)
not affiliated with or endorsed by Cohere