A reminder that italicized text is my own authorship; otherwise, it’s analysis from AI models.

Starting this week, the newsletter is going to take on a slightly different approach. There's a lot that happens week to week within the world of AI. New models are released, new things happen, but not all of it necessarily applies to the world of market research.

So, what I'm going to do moving forward, is after each story, I'm going to write about what lessons we can take as market researchers from each one of these stories. So you're not going to see a section dedicated to “what this means for market research,” but, instead, a breakdown within each section of what we can take away from what's happening in the world of AI.

I hope you enjoy and benefit from this new direction. Let's move on to this week's news!

AI is making execution cheap. Deciding well is becoming the expensive part.

For years, the promise of AI has been that it could help people produce more work, faster. This week, several different stories pointed to the consequence of succeeding at that: when producing options gets cheap, deciding which option is right—and knowing when AI should be involved at all—becomes more important.

The common thread is not that AI has failed. It is that faster execution increases the value of deciding what deserves to be executed in the first place.

Z’s Take

I think the researchers’ role is starting to move to the outer ends of a research project.

Upstream, a researcher is becoming even more important in defining the actual problem to be researched. They’ll also need to be able to identify existing data sources, whether those data sources are sufficient to answer the question, and what additional data needs to be collected.

Downstream, researchers will still need to do what they've always done: find the insights. But they will also need to know whether an AI analyzed the data correctly, used and called on tools correctly, and whether the conclusions are supported by data or just confidently wrong.

Researchers will need to know more than just how to use AI to write prompts to help them write reports or create discussion guides. Researchers will need to know AI well enough to judge where it fits into the research process well, where they need to check what the tools are doing and how thoroughly to check the output, what tools work well for what tasks, and perhaps even what models to use in order to keep costs reasonable.

Stories that got stronger as the week went on — or are new this week

Small teams are becoming systems of humans and AI agents

An AI agent is software that can carry out multiple steps toward a goal rather than simply answering a prompt. This week, the conversation moved beyond individual agents toward teams designed around networks of agents, tools, knowledge and human checkpoints.

Open models are becoming credible participants, not just cheaper alternatives

Open-weight models are AI models whose trained model files can be downloaded and run outside the company that created them. Their improving capability is making it increasingly practical to use several model providers—or even ask several models to tackle the same problem and compare their answers.

Human review is being redesigned rather than simply added

“Keep a human in the loop” sounds reassuring, but this week brought evidence that a human checkpoint does little if the person has no meaningful basis for evaluating what the AI did.

Z’s Take

If you're a freelance market researcher, I don't know that the news about open models or open weight models necessarily applies to you, unless you're comfortable using command line interfaces or if you have the computing capacity to load and run the open-weight models locally.

I think the news about Anthropic defaulting Claude to automatically approving tasks because it turned out that most people didn't know exactly what they were approving in the first place actually serves as a reminder that as researchers, no matter what tool we're using, we shouldn't just be blindly accepting whatever the output gives us. Additionally, we need to be sure that we're not just approving access requests from LLMs to everything on our computers. Be sure that you are aware of what your LLM has access to if you’re a freelance researcher so that you aren’t introducing security risks to yourself, your customers, or participants in your studies.

If you are on the brand side and you're working with AI tools that have been approved by the company you're working in, you’re in a safer position than a freelancer, but it's still worth checking your settings to make sure that your conversations are not being used to train the AI models.

Signals appearing in 4 or more of the last 8 weeks (Long-term Continuing)

AI work is becoming orchestration, not prompting — 10+ weeks running

The recurring shift is from asking one model to perform one task toward designing repeatable systems of agents, context, tools, permissions, tests and handoffs.

Human judgment remains the verification layer — 8+ weeks running

AI keeps getting better at producing outputs that pass automated checks, while questions about whether the work is appropriate, original, trustworthy or worth doing continue to require judgment.

Using multiple models is becoming an operating strategy — 15+ weeks running

Organizations and practitioners continue moving away from assuming one AI provider or model should handle every task, using routing to manage capability, cost, access and risk.

Z’s Take

Now that different models within AI ecosystems have very different price points (for example, Claude's ecosystem now has Fable, Sonnet, and Opus, and ChatGPT's ecosystem now has Sol, Terra, and Luna), it’s worth knowing when to use models that cost the most versus when to use models that cost the least.

The general rule is to use the most expensive models for planning, and the less expensive models for execution. Another strategy is to use the expensive models for when execution needs to be less error-prone.

For researchers, this might look like using a Fable or Opus to plan a multi-step research project from start to finish, then using Sonnet to execute analysis steps within that project. Or it might mean using ChatGPT Sol to plan how to organize a research repository, use Luna to pull all the files into the needed sub-folders, and then use Terra to do a quality check on the work.

There’s been a little less said about token costs in AI news recently, but knowing model capability differences is something to keep in mind.

Signals appearing in 2–3 of the last 4 weeks (Short-term Continuing or Emerging)

Research is moving closer to the decision itself — 3 weeks running

As AI compresses the time required to build and test ideas, traditional research timelines increasingly risk delivering evidence after teams have already made the decision.

Z’s Take

The rest of this article by Constantine Papas is worth reading. He describes two other models of research that reduce the time between research and decision-making: consultative, where one person is the reference point for decisions, and integrated, where knowledge of the customer is integrated into the organization so that it surfaces without the researcher having to always be in the room at the time decisions are being made.

I’d argue that while AI has made it easy to build 3 different prototypes to test, the question remains: are any of the prototypes addressing the actual customer pain? And that’s where having a way for the customer information to be integrated into the organization improves outcomes, but also where I think researchers are still needed.

*This newsletter covers August 7 – August 13, 2026. Sources: Slow AI, AI Risk Management Newsletter, Unpromptable, Neatprompts, The Signal, AI Governance, Ethics & Leadership, Human+AI, Every, AI Daily Brief, The Voice of User, Lenny's Newsletter, AI Maker, The AI Architect, Prompt-Led Product.