# Cohere and Waterloo: the missing training is not technical

- Published: 2026-07-03
- Authors: CORTEXA
- Category: Opinion
- HTML: https://researchhub-vert.vercel.app/blog/cohere-waterloo-ai-change-management

Most failed AI deployments do not fail on model quality. They fail on process, ownership, and nobody deciding what the system is allowed to do.

Cohere announced a partnership with the University of Waterloo on a certificate in AI transformation and change management.

The instinct is to file this under corporate outreach. The category it addresses, though, is where most enterprise AI actually fails.

## The failures are rarely technical

Post-mortems on abandoned AI deployments tend to converge on the same set of causes, and model quality is seldom among them:

- Nobody owned the system after the pilot team moved on.
- No one specified what the system was allowed to decide unsupervised.
- The workflow it was meant to improve was never mapped, so the integration point was wrong.
- Staff were not told whether it was assisting them or evaluating them, and assumed the worst.
- There was no escalation path when it was wrong, so people stopped using it after the first bad answer.

Every one of those is a change management problem. None is solved by a better model.

## Why this is genuinely hard to teach

The reason a certificate in this area is unusual is that it sits between two established disciplines and belongs to neither.

Technical curricula teach how the systems work. Business curricula teach organisational change in general. The specific combination — what changes about change management when the thing you are deploying is **probabilistic, opaque, and improves under you** — is not standard content anywhere.

That combination has real content. Deploying a system that is right 92% of the time requires different process design from deploying one that is right or broken. Deploying a system whose behaviour shifts when the vendor updates it requires different governance from deploying software that changes only when you change it.

## The obvious conflict, stated plainly

A model vendor co-designing the curriculum that teaches organisations how to adopt models has an interest in the conclusions. That is not disqualifying — vendors also have the most deployment experience, which is exactly what the curriculum needs — but it is worth naming.

The signal to watch is whether the material is **vendor-neutral in substance**: whether it teaches evaluating alternatives, whether it covers when *not* to deploy, whether the case studies include failures.

## Why the trend matters regardless

The shortage of people who can run an AI deployment competently is more binding than the shortage of people who can train a model. There are far more organisations needing the first skill than the second, and almost no structured path to acquiring it.

Whether this particular certificate is good, more institutions treating deployment competence as a teachable discipline is the right direction. The current default — learning it by failing at it, expensively, once per organisation — is a bad way to run an industry.