The tool was never the hard part

Both made headlines. Neither is the story that should change your Monday.

The quieter one is this: the tools are working. Most companies just haven't done the work to turn them into results.

Buying AI is a lot like buying gym equipment.

The purchase was never the hard part. The hard part is whether anyone actually uses it, and whether you can tell if it's making a difference. Plenty of companies now have the equivalent of a room full of expensive machines and no idea if they're any fitter.

The value was never in the tool. It's in the boring work around it.

Your middle managers are carrying the rollout.

HBR put a name to where AI initiatives quietly stall: the middle.

Managers are told to drive adoption on top of running their teams.

No extra time. No playbook. No training for the new job they just got handed.

Executives set the ambition. Vendors ship the tool. The work of turning that into daily practice lands on the layer with the least slack to absorb it. So when a rollout underdelivers, it usually isn't the software. It's that nobody resourced the people asked to make it stick.

Most leaders can't tell you what AI returned.

MIT Sloan laid out three ways to measure AI's return, from single-function tracking up to portfolio-level governance.

Most organizations are on the bottom rung. Plenty aren't on the ladder at all.

A lot of leaders can tell you they're "using AI." Far fewer can tell you what it gave back. Without a number, every AI decision runs on faith, and faith doesn't survive the next budget review.

And nobody's checking the answers.

A separate study found employees tend to take the AI's answer and move on. They skip the extra digging that might complicate a clean recommendation.

Fluent answers feel finished. The more natural the tool sounds, the less anyone checks it. Adoption without scrutiny doesn't raise the quality of your decisions. It just lets you make the bad ones faster.

What to do this week: Pick one team already using an AI tool. Ask its managers a single question: what rollout work landed on you that nobody planned for? Then name the one number you'll use to decide, by the end of the quarter, whether that tool is earning its keep.

Ask why your teams still can't use the AI tool they keep requesting.

The honest answer usually isn't "it's too risky." It's that nobody has done the work to make it safe to approve. Governance gets treated as the function that says no. The companies winning with AI use it to say yes, on purpose, with the controls to back it up.

Samsung went from banning ChatGPT to deploying it everywhere.

In 2023, Samsung banned ChatGPT company-wide after employees pasted source code into it. This week it did the opposite. It rolled ChatGPT Enterprise and Codex out to its entire Korean workforce and its global device division, one of the largest deployments OpenAI has run.

Nothing about the underlying risk changed between the ban and the rollout. What changed is that Samsung built the controls to manage it. The takeaway isn't "be more permissive." It's that the road from no to yes runs through governance, not around it.

An agent only knows the rules you write down.

HBR's instruction for anyone deploying an agent: write down how your organization actually makes decisions. The unspoken rules. The judgment calls a good employee makes without thinking about them.

An agent can't infer the standards you never wrote down. Give it a vague mandate and it will optimize for the wrong thing, confidently. The work of getting value from an agent starts before the agent. It's making your own decision rules explicit enough to hand over.

The safety side is catching up.

A developer let roughly 2,000 people try to break into an AI assistant built on a current model. Six thousand attempts to trick it into leaking its secrets. It held.

The "is it safe to give AI access to our data" question has better answers than it did a year ago. Not perfect ones. But good enough that "we can't risk it" is getting harder to lean on as the reason to never start.

What to do this week: Take the one AI tool your teams keep asking for and you keep stalling. Write the three controls that would let you approve it: how it handles your data, who gets access, what it's allowed to touch. Then ask the vendor what red-team testing it's been through, and what leaked.

This week handed you the AI-and-jobs story from both ends.

One company blamed AI for 21,000 cuts. Another quietly rehired the people AI couldn't replace.

The useful read isn't "AI takes jobs" or "AI is overhyped." It's that the line between what to automate and what to keep human is a decision. Right now, most companies are letting a vendor pitch or a board narrative draw it for them.

Oracle put it in writing.

Oracle told the SEC it cut about 21,000 jobs last year, roughly 13% of its workforce. The filing names AI adoption as a direct cause, and warns there may be more. It spent $1.8 billion on severance while pouring money into AI data centers.

Plenty of companies have quietly blamed "restructuring" for AI-driven cuts. Oracle wrote it into a federal filing. That makes it the example your board reaches for when it asks the obvious question: is that us? Better to have an answer that's a decision than a flinch.

Ford spent three years rehiring humans.

Same week, the opposite direction. Ford has spent three years bringing back veteran inspectors and engineers, around 350 of them, after its AI inspection tools kept missing defects a human eye catches. It now uses AI to assist inspectors, not replace them. Ford just topped the latest J.D. Power quality ranking.

The work AI handled worst was the judgment-heavy, edge-case work. Which is exactly the work that's tempting to automate, because it's expensive. Ford's correction cost it years and a lot of money. The cheaper version is to ask the question before you cut the role, not after.

What to do this week: Borrow Ford's test. Before you automate any role this quarter, write down what a person catches in that work that the tool won't. If you can't answer it specifically, you're not ready to automate the role. You're ready to pilot alongside it.

The EU AI Act's high-risk deadline is August 2. As of that date, the Act's high-risk rules apply to AI used in hiring, lending, and healthcare that touches the EU. The ask is binary: does anything you run fall in scope, and who owns getting it compliant in time?

Copilot Cowork is now generally available. Microsoft's push to put agent-style help inside the Microsoft 365 apps your teams already live in went GA this month. Before you buy a separate agent tool, it's worth ten minutes to see what the license you already pay for now does.

It's tempting to read a week like this as a verdict on whether AI works. It isn't.

Oracle, Ford, Samsung, the manager drowning under a launch nobody planned for: each is a story about the work around the tool, not the tool. The software shows up ready now. Whether the organization does is the part nobody can buy.