The Signal

A weekly read for operators thinking seriously about AI. Less noise, more signal.

The Lead

Faster at the wrong thing

Most AI products being built right now have the same shape. The user has a problem. The system gives them an answer. Better answers faster, at scale, in any domain — that is the dominant design pattern and it is what the entire industry talks about as if it were the only one. The product roadmaps, the demos, the funding rounds, the keynote slides; all of them are versions of "we make humans faster at getting answers."

It is the wrong design pattern for most of the high-value work humans actually do, and the operators who internalize that this year will spend the next decade ahead of the operators who do not.

A piece in the technology press this week profiled an AI coaching tool that does something almost no other AI product does. It refuses to give advice. The CEO behind it, Emma Weber, built the system specifically to facilitate reflection. Sessions are time-bounded. The agent asks questions designed to push the user to examine their own thinking. It does not solve. It does not recommend. It helps the user have, in her words, a better conversation with themselves. Across roughly twenty-four thousand sessions, the rate of escalation to a human was approximately zero-point-zero-zero-eight percent.

That is a contrarian design choice in 2026. It is also the right one for the category. Soft skills, leadership development, professional growth, the kinds of changes that compound over years — none of these are answer problems. They are noticing problems. The work is not "tell me the right thing to say." The work is "help me see what I keep missing." An AI that gives empathetic-sounding scripts can make a leader sound better in a single conversation. An AI that asks the leader why they always avoid the same kind of conflict can change the leader. Those are different products serving different timescales of value.

This distinction extends far beyond coaching. The same principle applies to any domain where the constraint is human attention rather than human computation. Strategic decisions. Sales judgment. Hiring calls. Founder priorities. In all of these, the bottleneck is not "what is the answer" but "am I asking the right question." Build an AI that answers faster and you get a leader who is faster at the wrong thing. Build an AI that surfaces what they are not noticing and you get a leader who is, slowly, becoming better at the work.

The market is going to figure this out, eventually. The first wave of AI products won the demo and lost the year. The category that wins the next wave will be the one that makes humans more thoughtful, not just faster. The signals are already there for operators paying attention. The reflection-based coaching tools. The decision support systems that prompt rather than prescribe. The agents that ask follow-up questions before they act. None of these win the speed comparison. All of them produce the durable behavior change the speed comparison cannot.

If you are buying AI to make your team faster, you are buying the easy version of the product. The harder version, the one most vendors are not building yet, makes your team better. That one is worth waiting for. It is also worth building yourself, if you cannot find it on a vendor's roadmap.

Three Things Worth Knowing

The contrast exercise is underrated

Dr. Art Markman, a cognitive scientist at the University of Texas, described a law school exercise this week worth borrowing. Students wrote letters to clients, then asked AI to rewrite the same letter with more empathy and less technical language. The rewrites were not the lesson. The contrast between the original and the rewrite was — students could suddenly see what their original drafts had omitted. This is a cheap, repeatable technique applicable to almost any operator output. Take something you wrote, ask an AI to rewrite it with a specific quality you suspect you are missing, and read the diff. The diff is the teacher. The AI is just the comparison surface.

Imitation is not understanding

The same Markman piece flagged a quieter risk. AI can teach a person to use empathetic language. It cannot, by itself, teach them empathetic attention. A leader who uses an AI to draft warmer client emails has not necessarily become a warmer leader; they have become a leader whose emails sound warmer. The two are not the same and the gap eventually shows up — in meetings, in conflict, in the conversations the AI was not in the room for. This is the operator's warning label for almost every AI product currently being sold as a soft-skill enhancer. If the tool changes the output without changing the attention, the change does not generalize.

The knowing-doing gap is where AI training programs fail

Weber's framing of the "knowing-doing gap" is the most useful diagnostic any operator can carry into a training procurement conversation. A manager rehearses constructive feedback in a Tuesday training session and reverts to avoidance with a real direct report on Thursday. The training did not fail to teach. It failed to transfer. Practice alone is not adoption, and AI tools that demo well on rehearsal scenarios produce no measurable behavior change in production unless they explicitly engineer for the gap between session and situation. The question to ask any vendor selling AI training: what specifically in your product addresses the knowing-doing gap? If the answer is more practice, more reps, more simulations, the answer is no.

From the Practice

Build advisors, not answer machines

There is a design choice that comes up in almost every client engagement and almost nobody surfaces it explicitly. Should this agent give the user the answer, or should it help them find the answer themselves? The first agent is faster to build, easier to demo, and feels productive in the first month. The second is slower to build, harder to demo, and produces real behavior change by month six.

The rule we apply: if the user has to make the same kind of decision repeatedly and the decision quality compounds across cases, build an advisor that prompts. If the user needs a one-off output and never has to make a similar judgment again, build an answer machine. The mistake is using answer machines for compounding decisions. The user gets the answer; the user does not get better. Six months in, the user is more dependent and no more capable, which is the opposite of what most operators thought they were buying.

Reflection-by-design is harder to build because it requires constraining the system, not extending it. The instinct of every engineer is to give the user more. The instinct of every good coach is to give the user less and ask them more. The agents that change behavior are built by people who can tell the difference.

Looking at an AI deployment in your business and unsure whether you have built an advisor or an answer machine? That is one of the most useful questions an operator can ask, and exactly the kind of conversation a Discovery call is for. Bring the use case; we will work through which mode the system should be in.

Book a Discovery Call

Or read this issue on LinkedIn.

Applied AI strategy, training, and implementation.

Ryan Rezeika · Founder · Vancouver, BC
ryan@resonai.ai · resonai.ai · LinkedIn

You're receiving this because you subscribed at resonai.ai/signal. · Unsubscribe