AI Signal helps business leaders decide what to fund, test, or avoid. Each five-minute issue features five articles I choose, why they matter, and what I’d actually do about each one. I’m Yashar, founder of PivotPath, an AI product studio in Toronto.
Every item carries a tag — Fund, Pilot, Watch, or Skip. That's my call on what you should do about it in the next 90 days.
―――――――――――
AI is making employees faster. Companies still aren't seeing the money — [Fund]
McKinsey— The state of AI in 2026
McKinsey’s latest survey shows a familiar gap: 80% of respondents say AI improves personal productivity, but only 37% say it has contributed to company EBIT. Just 6% qualify as AI high performers. The difference is what those companies do with the technology. Nearly three-quarters of high performers say they are fundamentally redesigning workflows around AI, compared with about one-quarter of everyone else. They are also more likely to use AI for growth and innovation, not just efficiency.
The Signal: People are faster at their desks, but the company’s income statement can’t see it. That happens when you speed up individual tasks inside a process that hasn’t changed. One step gets faster, then the work waits in the same queue, approval process, or bottleneck. The high performers are the real finding: they redesign the workflow, not just the task.
Pick the process where your team already uses AI most heavily. Map it end to end and find the step that didn’t get faster. That’s where your ROI is probably stuck.
―――――――――――
Meta tried to replace people with AI. The numbers told a different story. - [Skip]
Reuters — Meta’s AI-native transformation
Meta explored shrinking some teams by as much as 60%, betting that AI agents and smaller teams could do much of the work previously handled by employees. But internal data showed a large gap between activity and results: code changes rose 220%, while changes that actually reached users as new or improved features increased only 36%. Technical and security incidents also rose 40%, while time spent dealing with them increased 70%. Meta ultimately cancelled plans for a second wave of restructuring.
The Signal: Meta had the money, talent, and computing power to run this experiment at enormous scale. The problem wasn't a lack of AI activity. It was that more activity didn't translate into the same increase in useful output. The important measurement was not how much code AI helped produce. It was how much better software reached customers — and what it cost to keep that software working.
Skip the headcount-first AI strategy. Choose the business outcome first, measure whether AI actually improves it, then change staffing. Cutting people before you have that evidence risks removing the people who would have spotted what wasn't working.
―――――――――――
Small businesses say AI is driving revenue. Most aren't measuring it. - [Pilot]
Clutch — What AI Maturity Looks Like in Small Businesses
Clutch surveyed 600 small businesses already using AI. The results look impressive: 83% report positive business outcomes and 81% believe AI has contributed to revenue growth. But only 54% actually measure AI's impact using defined metrics. That gap matters. One expert in the report points out that many small businesses don't know how long a process takes, how much it costs, or how many errors it produces before introducing AI. Without that baseline, there is no reliable way to prove whether the new system worked — and unsuccessful projects often end up quietly abandoned.
The Signal: If you can't measure it, revenue growth is still a feeling. Saying AI helped the business and proving it are different things. If you don't know what a workflow cost before AI, you can't calculate what it saved afterwards. Pick your most-used AI workflow and measure it for one week. Track how long it takes, how many people touch it, and how many errors or reworks occur. Then keep measuring after AI changes it. Do that before spending another dollar on tools. A simple baseline turns “AI seems to be working” into a number you can defend in a budget meeting.
―――――――――――
Companies can predict AI adoption. They’re much worse at predicting the benefits.- [Pilot]
U.S. Bureau of Economic Analysis — AI Expectations and Outcomes
A new BEA study compared what businesses expected from AI with what actually happened later. Companies were surprisingly good at predicting whether they would adopt AI: on average, forecasts of AI use six months ahead were within two percentage points of actual adoption. But predicting the benefits was much harder. When researchers compared companies’ original reasons for adopting AI — such as automating work or expanding products and services — with later economic results, they often found no clear link.
The Signal: We’re better at predicting adoption than benefit. That distinction explains a lot. Companies can usually tell whether they’ll use AI. They’re much less reliable at predicting what it will actually deliver. And a prediction you never write down can never be wrong. If your AI programme is expected to “improve productivity” or “drive growth,” you can always convince yourself it worked.
Write a one-page AI forecast for your business. What should change over the next 12 months? Give each prediction a number, a date, and an owner. Then schedule a review. A year from now, you’ll know something far more useful than whether you adopted AI: whether your assumptions about its value were actually right.
―――――――――――
AI is starting to control physical machines — and this time the results are measurable. - [Watch]
Anthropic — Model Hardware Standard research preview
Anthropic has introduced the Model Hardware Standard (MHS), a way for AI agents to communicate with and control different pieces of laboratory and manufacturing equipment through a common interface. Early tests produced unusually concrete results. At quantum-computing company QuEra, an existing laser-recovery system worked 58% of the time and took about 150 seconds; the AI-developed version later succeeded in 99.3% of 700 blind tests, with many recoveries taking just seconds. At Carnegie Mellon, connecting several lab instruments took about eight hours instead of several weeks.
The Signal: Notice what’s different about these numbers. There’s a before, an after, and a result you can actually count. 58% to 99.3%. 150 seconds to six. Weeks to eight hours. That’s what happens when the work has a measurable output and somebody was already recording it.
Back in AI Signal #0 — Cheap Now, Expensive Later, the signal around drone delivery was to start asking a make-or-rent question: if AI becomes strategically important to your operation, which capabilities should you own and which are you comfortable renting from someone else? The suggestion was to put that decision on the leadership agenda before the technology becomes critical.
This Anthropic work makes that question more relevant. AI is getting closer to the physical equipment that runs labs and production lines, and a common integration layer could make that capability accessible to smaller companies too.
It’s still a Watch. Claude’s physical reasoning remains limited, and expert oversight is required. But if you’re in advanced manufacturing, life sciences, or run programmable equipment that doesn’t talk to each other, pay attention. And carry the broader lesson back to any AI project: the clearest ROI this week came from work where somebody was already measuring the process before AI arrived.
―――――――――――
That's this week's signal.
One ask, and I mean it: hit reply and tell me what you're actually seeing — the pilot that stalled, the tool nobody uses, the thing that worked. I read every one.
Talk soon,
Yashar