Mandating AI Isn't an AI Strategy. What the Microsoft Staff Letter Gets Right.

This week one of the most searched business stories in India wasn't about a product launch or a share price. It was about how people inside one of the world's biggest AI companies feel about AI.

According to The New York Times, a group of current and former Microsoft employees calling themselves Concerned Microsoft Workers has circulated an open letter about the company's AI push. They describe a "speed-at-all-costs" approach. They say AI tools are being pushed into almost every part of their jobs, often with little training, and that the tools frequently create extra work because people have to fix poor output. They ask Microsoft to stop tying AI use to performance rewards and reprimands, and to give workers a real say in how AI is rolled out. For many of them, the letter says, work has become "joyless and demoralizing".

The letter surfaced the same week Microsoft reportedly offered voluntary buyouts to about 60,000 staff.

We don't know the inside of Microsoft, and we're not going to pretend we do. But the pattern in that letter is one we see in companies of every size. It's worth taking seriously, because it's the clearest description we've seen of how AI adoption goes wrong.

Usage is the wrong number to chase

Here's the trap. Leadership invests in AI. Leadership needs to show the investment is working. The easiest thing to measure is usage: how many people opened the tool, how many prompts, how many agents built.

So usage becomes the target. Then it shows up in reviews. Then people use the tool because they're told to, not because it helps.

That's how you end up with what the Microsoft letter describes: AI everywhere, and people spending their day cleaning up after it.

The number that matters isn't how much AI your people use. It's how much work actually left their desk, and whether what's left is the work only they can do.

Agents should take work off people, not add review work

A lot of enterprise AI today is designed like this: the AI drafts, a person checks everything, the person fixes what's wrong, the person sends it.

On paper that's "human in the loop". In practice the person now has two jobs: their old one, and proofreading a machine. If the AI is right 80 percent of the time, the person still has to read 100 percent of the output to find the 20 percent that's wrong.

That isn't a productivity gain. It's a new kind of tiredness.

The fix is to design the work before you deploy the tool. For every step in a process, decide which of three things it is:

1. The agent does it alone. Reversible, low-cost, well-defined: reading, matching, sorting, internal updates.
2. The agent does it, a person approves. Anything that leaves the building or moves money.
3. A person does it. Judgement, exceptions, relationships, anything you'd be embarrassed to explain was done by a machine.

When that's written down, people know where they fit. The agent isn't watching them or replacing them. It's a teammate with a clearly smaller job.

What it looks like when it's done right

When we built a yield-monitoring operation for DataBeat, the analysts' mornings were the problem: four to six hours every day going through dashboards to find what moved overnight and why.

We didn't hand the analysts a chatbot and ask them to use it more. We took the routine itself, the pulling, comparing, checking and writing up, and gave it to a team of more than a dozen specialised agents with a manager agent coordinating them. That work now takes minutes. Analyst time across those workflows fell by 93 to 99 percent.

The analysts didn't become proofreaders. They moved to the part of the job that needed them in the first place: deciding what to do about what the agents found.

That's the difference. In one model, people serve the AI. In the other, the AI serves people's decisions.

Give people a say in how agents join the team

One of the demands in the Microsoft letter is simple: workers want a voice in how AI is deployed in their own work.

That shouldn't be controversial. The person who has run distributor claims for years knows exactly which claims are routine and which ones smell wrong. The finance lead knows which supplier always sends invoices in a strange format. That knowledge is what makes an agent safe and useful. Leave those people out and you get an agent that's confidently wrong in all the places they could have warned you about.

In the operations we've built, the people closest to the work own the rules the agents follow. At Wyra, ex-AWS sales leaders own the playbooks their outbound agent executes. At Mivi, the product team shapes each AI avatar's personality. The agents run the work. People own the judgement.

Agents need managers too

There's a flip side. If you're going to give agents real work, you need to manage them like you'd manage any team member:

• Clear limits on what each agent can do without asking.
• A budget that actually stops spending, rather than a report at the end of the month.
• Regular checks on quality, because agents drift as data and processes change.
• A record of what each agent saw, decided and who approved it.

This is what we mean by governance. Not committees and policy PDFs. Just the basics of running a team, applied to a team that now includes software.

What we'd tell any leadership team this week

If you're under pressure to show AI progress, and most leadership teams are, here's what we'd suggest instead of a usage mandate:

• Pick one process with real volume and real pain. Not "use AI more". One specific operation.
• Ask the people who run it where the time goes. They'll tell you in an hour.
• Write down who decides, step by step. Agent alone, agent then person, person only.
• Measure hours returned and decisions improved, not prompts sent.
• Keep people's names on the judgement calls. It's better for quality, and it's what keeps the team on your side.

AI that people are forced to use will always be resented. AI that quietly takes the worst part of someone's job off their desk gets defended by the very people it works for.

That's the version worth building.