McKinsey Found a Third of Agent Rollouts Made Things Worse

So, What Actually Happened?

Monday, and I cannot get past one number from the weekend. Thirty percent. Not thirty percent faster. Thirty percent of companies running agentic AI in production where productivity actually went down. That is McKinsey counting, not a sceptic posting. We scanned 190,000 articles this week so you don't have to, and the same weekend handed me two more numbers pointing the same direction: internal Bing data showing Copilot cut publisher click-throughs by up to 94%, and Kakao quietly shutting down its Kanana chatbot after a run of AI projects that never landed. For two years the question was what these systems can do. Somebody finally added up what they did.

The Bottom Line: The first honest scoreboard for agentic AI arrived this weekend, and a third of the teams on it are worse off than before they started.

The Tracks That Matter

1. McKinsey Counted, And A Third Of Rollouts Went Backwards

A McKinsey report landed showing productivity fell in 30% of companies that put agentic AI into real operations. Not plateaued. Fell. The striking part is not the figure, it is that somebody measured at all, because most agent rollouts I see carry no before-picture, which makes a decline invisible and a gain unprovable. The explanation arriving alongside it is that the generic assistant was simply the wrong shape, and enterprise AI has to move to specialised vertical agents that know one workflow deeply instead of touching forty shallowly. A general-purpose agent dropped into a specialised process behaves like a stand-in DJ handed somebody else's crowd. Technically playing. Reading nothing.

Here's what works: Before your next agent goes live, write down the one number it must move and today's value of that number. No baseline, no result.

2. Bing's Own Data Shows Copilot Taking 94% Of The Clicks

Internal Bing figures show Copilot cut publisher click-through rates by up to 94%, which is the first time that number has come from inside the house rather than from a furious publisher. Read it next to Conductor's benchmark across nearly 14,000 domains, where AI referral traffic averages about 1% of sessions, and the shape of the problem changes completely. The traffic did not move to a new channel. It stopped being traffic. Anyone whose funnel opens with a search result now runs a business whose top of funnel is counted by someone else's model, in someone else's dashboard, with no line anywhere for the visit that never happened.

Here's what works: Pull six months of organic sessions and split them by whether an answer engine could have answered the query outright. That split is your exposure.

3. A Breach Went Unreported Because The Company Defines The Word

The RubyGems supply-chain compromise was never reported to Brussels under the EU AI Act's disclosure rules, and the cause is not a missing form. The company classifies its own incidents, so something that looks like a reportable breach from outside sits comfortably inside the definition of something else. Utah is attacking the same gap from the opposite end, testing AI rules inside a regulatory sandbox where the state watches real systems run instead of reading attestations about them. One approach trusts the definition and waits. The other goes and looks. Every team signing an AI vendor contract this quarter is picking one of those without realising a choice was on the table.

Here's what works: Ask your AI vendor who decides what counts as a reportable incident. If the answer is the vendor, your disclosure clause is decoration.

Quick hits:

• The slowdown talk went to court. A lawsuit alleges the largest labs made an illegal agreement to coordinate an AI slowdown, which converts a safety conversation into an antitrust exhibit and gives discovery a reason to read everyone's email.
• Two governments bet on goal-free AI. Canada and Germany put CAD 300 million behind Yoshua Bengio's safe-by-design approach, the first serious state money aimed at systems deliberately built without objectives of their own.
• A screenshot tool leaked 23.6 million users. The Gyazo breach hit exactly the kind of small supplier that never appears on a third-party risk register and sits on half your team's desktops anyway.

Signal vs. Noise

🟢 Signal: data security. The weekend's movement was in the plumbing, not the models: a package registry compromised, 23.6 million accounts exposed, and the people who broke into OpenAI going to the press to say the industry has a security problem. Most coverage still files all of this under model safety, which is not where a single one of these incidents actually happened.

🔴 Noise: ”compliance” as a category. The word pulled one of its heaviest weeks of coverage this month and lost more of its grip on real events than anything else we track. That is what happens when one term covers both the thing that stops an incident and the paperwork that describes one afterwards. No compliance programme caught any of this weekend's failures.

From the 190K

We scanned 190,000 articles this week. Here's what no one's talking about:

NVIDIA shipped a tool that inspects what an AI agent is permitted to invoke, a startup launched a model that refuses to write a single word of prose on purpose, and Alibaba open-sourced a reviewer that reads AI-written code before a human sees it.

Three desks, three filings, no overlap. The chip press covered SkillSpector as agent security. The model watchers wrote up Jev, which emits typed structures instead of sentences, a design its makers describe as a system that cannot produce prose at all. The developer-tools wires ran Alibaba open-sourcing OpenCodeReview. Read them on the same morning and they are one build: not making the model better, but narrowing what it is allowed to emit and checking the output before it moves. That is a quiet admission from three unrelated corners that the expensive failures are not stupid models, they are capable models producing plausible work nobody inspected. Which is exactly where McKinsey's 30% comes from.

What changes today: find the one place an AI writes something in your business that goes straight out without a second pair of eyes. That is the cheapest risk you will retire this quarter.

By The Numbers

• Productivity fell in 30% of companies using agentic AI — McKinsey counted the deployments rather than the demos, which makes this the first decline number a board will actually accept.
• Copilot cut publisher click-through rates by up to 94% — the figure comes from internal Bing data, so the platform doing the cutting is the one reporting the size of it.
• AI referral traffic averages about 1% of sessions across nearly 14,000 domains — Conductor's benchmark, and the distance between that 1% and the 94% above is the entire measurement problem in two numbers.
• 23.6 million Gyazo users exposed in one breach — a free screenshot utility, which is the category of supplier that never reaches a risk review and always reaches the laptop.
• Canada and Germany committed CAD 300 million to safe-by-design AI — public money aimed at systems built without goals of their own, arriving beside private rounds still funding systems with more of them.
• See what's rising across AI and data this week →

Deep Dive: The Thirty Percent Who Got Slower

I emptied dancefloors, often. But I always knew the risk I was taking, because I knew what worked and could fill the floor again. Sometimes one record fixed it. Sometimes it took half an hour to rebuild. The risk was never the empty floor. The risk would have been not knowing the way back.

Nobody wrote down the before

A third of agentic rollouts got slower, and most teams in that third cannot prove it either way. They measured the demo, then called the demo a deployment. When the cycle time drifts three weeks later, there is no baseline to compare against, so the conversation becomes opinion versus opinion and the loudest person wins.

The agent replaced the step, not the judgment

Automation pays when it removes a step nobody should be doing by hand. It costs when it removes the step where someone was quietly catching things. A generic agent cannot tell those two apart, because from the outside they look identical: a person, a screen, some typing.

Getting back is dearer than going

Rolling forward takes a sprint. Rolling back takes a quarter, because the manual process has already been dismantled, the people who ran it have been reassigned, and nobody wrote down how it worked. That asymmetry is why a reversible pilot beats an impressive one.

What Actually Works

1. Baseline before you build: one metric, measured for four weeks, written down where the team can see it. Without it you cannot prove a win or spot a loss.
2. Keep the manual path warm: do not dismantle the old process until the new one has survived a bad week. Redundancy is cheap right up until you need it.
3. Automate the typing, not the catching: map where judgment happens in a workflow before an agent goes anywhere near it.
4. Set the kill criteria first: write down the number that ends the pilot, before anyone is emotionally invested in its survival.

An empty floor is survivable when you know which record brings them back. It is a career event when you don't.

What's Coming

The Slowdown Argument Gets A Docket

The question of whether the industry is really ready to slow down stops being a panel topic the moment an antitrust claim attaches to it. Expect the safety language in public statements to get carefully lawyered within weeks, because anything said about coordinating restraint now reads as evidence of coordination.

Open Weights Become The Answer To The Productivity Bill

StepFun's Step 5 Preview arrives at 600 billion parameters with a 1M context window and open weights due October 15. When a third of deployments are showing negative returns, the first lever finance pulls is not better outcomes, it is cheaper inputs. Open weights are about to get a hearing from people who were not listening a quarter ago.

Memory Becomes The Line Item Nobody Modelled

China's CXMT says a new memory-chip platform has entered mass production. Everyone budgeted for GPUs and treated memory as an accessory. Memory pricing is the part of the AI bill that moved most while nobody watched, and a second serious supplier changes that negotiation in both directions.

For Your Team

Tuesday's meeting prompt: ”Name the AI system we are proudest of. Now tell me the number it was supposed to move, what that number was before we started, and what it is today. If we cannot answer all three in this meeting, what exactly have we been reporting upward?”

Share-worthy stat: McKinsey found productivity went down in 30% of companies running agentic AI. Not flat, down. Sit that beside the same weekend's finding that Copilot cut publisher click-throughs by as much as 94%, and this becomes the first week the AI conversation had real subtraction in it.

Go deeper: Track where enterprise AI spending is actually landing →

The Track of the Day

”The honest middle path is to report what you can measure, label what you're estimating, and name what you can't see at all.”
Israel Acheampong, on measuring what AI search is really doing

Three imperfect numbers you can defend will outlast one clean number you cannot. That is true of AI search, and it is true of every agent dashboard that only shows green.

We scanned 190,000 articles this week so you don't have to. Data Pains → Business Gains.

Published: September 21, 2026 | Curated by Yves Mulkers @ Ins7ghts