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.
―――――――――――
A companion for every employee - [Fund]
Meta — The Future is for Everyone
Meta argues that the biggest opportunity from superintelligence is not automation, but individual empowerment. Its vision is for everyone to have a highly capable personal agent that understands their goals, works on their behalf, helps them learn and create, and makes it possible for individuals and very small teams to do things that currently require far more time, expertise and people. Meta’s broader bet is that widely distributing this capability could increase invention and individual productivity faster than AI replaces existing work.
The Signal: If that direction is right, AI starts to look less like software your company buys and more like the laptop: a basic layer of capability every employee eventually gets.
The interesting shift is from generic AI access to a companion built around the person doing the work. One that understands their workflows, recurring decisions, context and the things they are trying to accomplish.
That changes the question from “Where can we automate?” to “What could this person do if they had much more capability?” Pick one role you have five or more of. Sit with three people doing that job and find where more time, knowledge or capacity would materially change what they can accomplish. Build the companion around those constraints first. If it works, you have a pattern you can repeat across the company.
―――――――――――
Write the eval before you build the agent- [Pilot]
Hiring Agents Is the Easy Part
The argument here is that agent adoption will not be limited primarily by AI capability. The harder problem is verification: defining what good performance looks like, evaluating it continuously, assigning responsibility for feedback, and deciding what an agent is allowed to do when something goes wrong.
The Signal: This is a management problem wearing a technology costume. Most agent conversations start with models, frameworks and tools. Far fewer start with the question that will ultimately decide whether the agent works: How will we know it is doing a good job?
Humans learn many company standards by watching other people. Agents cannot reliably absorb everything that way. Tacit expectations have to become explicit—and many organizations have never written them down because they never needed to.
Before building an agent, write the evaluation first. Take 20 real examples from the last month, mark each as good or not good, and explain why. If you cannot create that document, you probably are not ready to automate the work yet. And even if you never build the agent, the exercise is worth doing.
―――――――――――
You're probably overpaying for routine AI work - [Pilot]
Ramp — Cracks in the AI Thesis
Ramp’s spending data shows adoption of model-serving platforms continuing to grow while growth at the largest frontier-model providers slows. At the same time, businesses are directing more AI spending toward open-source and alternative models rather than relying exclusively on the most capable and expensive options.
The Signal: Someone probably connected your AI stack to a frontier model when the project started because it was the safest choice. There is a good chance nobody has looked at that decision since. Now classification, extraction, tagging, summarization and routine first drafts may all be running through the same expensive model—even when better reasoning adds almost no value. The emerging model is simple: use the best models for the hard work and cheaper models for routine work. Pull last month’s usage, find your three highest-volume, lowest-risk tasks, and test them against a cheaper model for two weeks while sampling outputs manually. It is not glamorous AI strategy, but it may be one of the easiest permanent savings available.
―――――――――――
Robots are waiting on us, not themselves- [Watch]
Bloomberg — Workers Are Teaching AI-Powered Robots to Take Over Their Jobs
Tens of thousands of workers in India are being paid to record themselves performing ordinary physical tasks from a first-person perspective. The footage is being used to train robots to operate in the physical world—creating a dataset that, unlike text and images on the internet, largely did not already exist.
The Signal: That makes physical data an important bottleneck for robotics. Before a robot can learn thousands of ordinary human tasks, somebody has to record, label and structure those tasks for training. That suggests physical AI may progress less like a sudden software breakthrough and more like a long process of building the missing dataset. Watch what companies are paying people to record. It may be one of the clearest signals of which physical tasks they expect to automate next. There is also a governance question here. As companies document workflows to make them AI-ready, those recordings, demonstrations and procedures become valuable training data. Businesses should start deciding who owns that data—and what employees have actually agreed it can be used for.
―――――――――――
Don't build an agent to see the future- [Skip]
The Frictions That Make AI Forecasting Hard
The argument is that AI forecasting often mistakes improvements in a single task for changes in an entire system. Real-world outcomes are constrained by regulation, institutions, physical infrastructure, organizational behaviour, trust and other bottlenecks that may not improve simply because an AI capability does.
The Signal: I work in strategic foresight, and I would not hand the conclusion to an agent. Here is the failure mode: AI discovers that a task can now be completed 10× faster and concludes that the industry will move 10× faster. But the actual constraint may be regulatory approval, a union agreement, customer trust, procurement, physical infrastructure or something that still has to be manufactured and shipped. Those frictions rarely appear in AI benchmarks, but they often determine what happens next. Use AI extensively inside foresight work: scanning for signals, synthesizing research, challenging assumptions and drafting scenarios. That is real leverage. But don't outsource the final judgment. The most important part of forecasting is often understanding the constraints the model cannot see.
―――――――――――
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