The AI Efficiency Trap
Melina Palmer: Hello. Hello everyone. My name is Melina Palmer and I want to welcome you to the Brainy Business Podcast.
By now, I’m guessing you have at the very least started to play around with AI in your life and business. From the work perspective, what has your gut instinct led you to use it for? Or how have you seen those around you start to use and test it out?
For a lot of people and companies, the default idea is to do more faster. It becomes about efficiency over everything else. You can have even more reports before going into that meeting, and instead of writing one article, you could have 47 social media posts, and you can do it all with the same amount of people or less. But is any of it actually better? Are you getting better results? Better team dynamics? A better feeling at work amongst the people who are still there, and happier customers? More engaged interaction with the content you’re churning out one prompt at a time?
For most people, the answer is no. And a big reason is because our brains have led us to this efficiency trap. When it comes to applying AI, that isn’t the only option. And in a lot of cases, it definitely isn’t the best option.
My guest today to discuss this is Will Leach. Will is the founder and CEO of Mindstate Group, a behavioral research and brand consultancy built around the Mindstate model that he introduced in his bestselling book, Marketing to Mindstates. He’s spent over 25 years applying behavioral science to marketing for Fortune 500 companies like PepsiCo, Mars, and New York Life. He’s also the creator of Bevy, an AI platform designed to give brands ongoing access to the psychology behind their customers’ decisions.
Will and I teach together via the Texas A&M Applied Behavioral Economics Certificate program through the Human Behavior Lab, and I am delighted to have him back joining me again today, as he’s always full of great insights.
As you listen today, I encourage you to think about how you’ve started using AI, how it’s working, and what an optimal scenario could look like. You don’t have to have all the answers, but knowing what’s good and what needs improvement for you specifically is useful to have in mind as you hear the suggestions from Will and myself during the discussion.
And really quickly, before we get into the conversation, I want to be sure you know there are links in the show notes for my top related past episodes and books, ways to get in touch, and more. It’s all waiting for you in the app you’re listening to and at thebrainybusiness.com/597. Now, let’s jump right in.
Introduction to Will Leach and Behavioral AI
Melina Palmer: Will Leach, welcome back to the Brainy Business Podcast.
Will Leach: It is so great to see you again, Melina. It’s been a while.
Melina Palmer: Yes! I mean, it’s been a while in the land of podcasts, but thankfully we’ve seen each other more than just once in seven years or whatever it’s been. When your first episode came out, Will, you were actually one of the first interviews I did. I think if I remember correctly, your episode was around 88 or something, and the first interviews I did were around 78. So you were one of the first.
Will Leach: I was a guinea pig for you! Okay, hopefully it was a good experience.
Melina Palmer: Yeah! Well, I mean, in reverse, hopefully it was a good experience for you, too!
Will Leach: Really, I didn’t know what I got into, but I did my best.
Melina Palmer: For everyone who doesn’t yet know you, can you share a little bit about yourself and the work that you do?
Will Leach: So, everybody, I’m Will Leach and I am the founder of Mindstate Group. I run a behavioral research and brand consultancy. We specialize in understanding those non-conscious factors of human behavior that you learn about on this podcast all the time.
I’ve been in the industry for quite a while, and I wrote a book in 2018 called Marketing to Mindstates, which was my take on understanding how people make decisions very fast and very fluidly, using a framework called the Mindstate Model. It’s a practical guide to helping marketers really use behavioral science in their profession.
Melina Palmer: Yeah, I remember reading your book way back in the day around when it first came out, and I remember reading it thinking, “Man, this guy’s got it together! He’s got this amazing framework, so much interesting stuff.” I really loved how everything was presented, so I was delighted to have you on then and back again today.
So Will had reached out to me because we’ve been friends over the years, and we both teach together through the Texas A&M Human Behavior Lab Certificate in Applied Behavioral Economics—the longest name ever!
Will Leach: It sounds so official, though!
Melina Palmer: It is! We are “totes official,” as I would say—no one else would say it that way, but I would.
Will had reached out because, as with everybody, he’s been making some leaps and forays into this wonderful land of AI. Actually, Will, you were the first person I knew who was actively using Claude. I remember we were at an IIEX event—shoutout to GreenBook—in Austin, and we were chatting and you mentioned how you were having a lot of success with Claude and it was doing a good job. That would have been 2023.
Will Leach: 2023 sounds right, yeah.
Melina Palmer: So how did you start using AI so early? Can you tell us a bit about your journey on the behavioral side of AI stuff?
Will Leach: I was lucky enough to have an agency where a gentleman was asking me about how I do copywriting and things like that. He talked about a new tool he heard about through a company in Austin called Jarvis—I think they’ve renamed themselves a bunch of times since. He said, “You should look at this new platform; it’s really good at copywriting.”
Now, I used to do all my own copywriting, so back then it was kind of weird. But I thought, “Maybe this will help me make my own business faster.” When I jumped on, it was a platform that was very hard to use. It was forcing me into set questions and answers. It asked me a question, but I accidentally answered the wrong question. I remember it came back with a response that shouldn’t have made sense based on the prompt, but it did! I thought to myself, “Wait a minute, the question has nothing to do with what I put into this box. It’s trying to force me into a workflow, but I think I can just answer anything or ask any question I want and a response will come back.”
So I did that, and all of a sudden I started realizing almost immediately: wait, there is this world called Large Language Models, and I can access it now! That was very early on.
Then I hired a high school senior from Silicon Valley who had just been accepted to UC Berkeley. I found him on a Reddit thread. He knew about behavioral science—he was just a savant—and he was also really early with OpenAI. Somehow OpenAI had given him and a couple of other people early access to these models. Literally, one time he said, “If you just pay for my burritos, I’m happy to talk to you,” because he just loved behavioral science! He was so much farther ahead because he lived in Silicon Valley. That’s how I got my first introduction to all of this.
Melina Palmer: That’s fascinating! When you were saying you answered the wrong question and it gave you a good response, I was wondering if it was like satisficing, or the problem that most people come across now. We’ve all probably experienced forgetting an attachment and saying, “Please review this and give me feedback,” and even without the attachment, it writes a whole write-up anyway! You’re like, “Hey bro, there’s no document there! What did you base this off of?” I love all the memes these days about AI making things up!
Will Leach: It used to be a lot worse, by the way! But that’s kind of what happened: I answered the wrong question, it took that little bit of context I gave it, and it wrote things back to me.
Context is a big deal in what we do and how to properly work with AI. In the marketing and brand management world with large corporate clients, employees are not being trained fast enough on how to use it. Most of them are just using AI as a co-pilot, so they only know how to use one model, and they’re in meetings all day, so they really don’t know how to use it to its fullest.
The Importance of Context and Iteration in AI
Melina Palmer: Let’s dig in on that a little bit. Of course, you’ve got Copilot, you’ve got Claude, and you’ve got ChatGPT, which I think most general users jump into because of the headlines. Many large businesses require Copilot because of enterprise security. A lot of people in the writing, creative, and thought-leadership space are using Claude more.
As you’ve seen, what is the problem with the way people are jumping in without training? What I see is people saying, “It doesn’t give me the output I want, so it’s easier to just do it myself.” There’s a self-fulfilling prophecy there, or an ostrich effect where people avoid it. What else are you seeing as far as consequences of not getting these things right?
Will Leach: It comes down to two things from what I see. One is that, at least in big corporations, they’re not being told to build context for AI.
What I mean by context is that you must treat your AI agent or assistant—whether it’s Claude, Copilot, or Gemini—as if it knows nothing about your business at least once. You have to build context. You can literally go in and type or say: “Claude, I want to work with you to build context for us to have better engagements. Interview me about my job, my role, and my life so you better understand what I deal with every single day and how to give me better output.”
Claude will then have a conversation with you. You should not try to type this out—get on a microphone and talk! When we type, we constantly edit our thinking. When you talk, you just let everything out. What AI is great at doing is taking all that junk, organizing it, putting meaning behind it, and summarizing it really well.
I did a 30-question interview when I initially started working with Claude. I wanted Claude to know my background, my work in behavioral science, and my book. But I also said, “Here’s the company I run, here’s my role, here are who my clients are, and here’s what they’re asking me.” Claude will then create context files that reside on your desktop or project space so the AI looks at that before it ever goes searching for answers. That matters because you want answers customized to you so you don’t have to write massive, complex prompts every single time telling it to “act as the world’s greatest research analyst.” You don’t have to do that anymore if your AI has context.
The second thing I see is that people forget that the very first response you get from AI is oftentimes not very good. You look at that response and think, “I could have done better than that.” When you have that feeling, the first thing you should do is type: “I don’t think that’s a very good response. Get rid of the jargon,” or “This sounds like any generic AI agent could tell me this; I want you to go deeper.”
When you tell your AI that, it’s like telling an employee, “I think this is really mediocre work; you’re smarter than that.” An employee would go back and say, “Point taken, let me go deeper.” Your AI works the exact same way. It will go in, look for blind spots, and correct its own errors.
Oftentimes, people copy and edit the very first response, getting an okay output. My rule is to iterate three times, but if you ask at least twice, it improves dramatically. On the first answer, say, “This looks pretty generic; I want it to be more specific to my audience or my Vice President.” It will come back better. Then say, “Any AI agent could have done this; I want something even smarter.” It will come back with an even smarter answer, and then you can tweak it from there. I find that on the second or third iteration, especially for deep thinking, you get much better results. People don’t push back enough on AI, and then they get mediocre results.
Melina Palmer: When you push back, do you read what’s there and give specific feedback, like “I don’t think that’s the right bias to lean into, let’s focus on this one,” or “Can you confirm this is accurate?” Or do you literally just say “That’s generic, give me a better one” without wasting time reading the first pass?
Will Leach: Honestly, for the first pass, I usually don’t even look at it! I honestly go back and say, “Any agent could come up with this; give me something smarter.” That’s almost my rule of thumb now. On the second pass, I’ll review it. By the third pass, that’s where I’m really digging in and saying, “No, I don’t agree with that authority point,” or tweaking specifics.
In fact, I’ll use Perplexity a lot to verify things. What will happen sometimes is that ChatGPT or Claude will give you a stat, and the stat itself is correct, but the inference or claim drawn from that stat is exaggerated. I’ll take the stat to Perplexity and ask it to verify the claim. Perplexity will often say, “This stat is real, but the liberties taken to get to that conclusion are not backed up by research at all.” So I use Perplexity to audit stats.
Other than that, I read through it and ensure my point of view comes across. It might feel like that’s not saving time, but it is saving a ton of time compared to doing the first two or three drafts from scratch. Once it has your tone of voice, even if editing takes an extra hour, your content is so much better.
Melina Palmer: Yeah, and hopefully the world will get past the phase of “Great, I can publish 500 AI-generated posts instantly.” If you can’t bear to finish reading the output yourself because it’s so obviously generic AI, don’t post it!
Will Leach: I love that! That’s so great. If you don’t read it yourself, why should you expect anyone else to?
Melina Palmer: AI tends to over-explain some points and completely gloss over others in a way that’s opposite to how I share content, so it requires finagling.
Will Leach: To that point, Melina, I wanted context for my AI agents to understand my writing style, so I uploaded my book into it. Uploading your books and podcast transcripts helps a lot! Over time, it learns where you like to place emphasis and how you think through things. People complain that AI doesn’t have their tone of voice, but that’s simply because they haven’t given it context.
Melina Palmer: Exactly! It’s just building off of what everyone else on the internet says. I saw something recently where someone pointed out, “Romance novels weren’t built on AI; AI was built on romance novels!” Because there are so many of them in the training data!
Building Specialized AI Agents for Your Team
Melina Palmer: When thinking about AI agents, employees, and teams, it’s easy to get myopic in our daily work. Whether it’s a CEO asking AI to write an article or review an ad, people rarely step back to look at the bigger picture and ask, “What are the gaps on our team that an AI agent could fill?” For example, if you don’t have an SEO expert, an editor, or an email specialist, you can create agents for those specific roles.
At The Brainy Business, we created custom GPTs as specialized agents to review content and offer feedback. It’s never going to write thought leadership content for me—I want to have those original thoughts—but I can ask an agent for the best strategy to distribute it on LinkedIn versus Instagram. Can you share what you’ve done with building AI employees for your team?
Will Leach: Yes! I went through an exercise similar to what you did. I have a market research branch of my business, and a newer platform called Bevy Behavioral Intelligence, which gives clients 24/7 access to customer psychological insights.
The very first agent I ever created was a Chief of Staff—someone who could help me organize things, summarize incoming emails, and help me operate my business better. I went to my AI Chief of Staff and uploaded my capabilities deck and my product spec sheet. I said, “Based upon everything you know about my business, help me understand what AI employees or specialized agents I should be investing in to make my business run better.”
It came back with a list, and I am now up to 32 specialized agents! I have Nancy, my newsletter specialist; Parker, my podcast specialist—who prepared me for this podcast episode with you today, Melina! I have Brooke, who handles Bevy onboarding, and Sam, my SEO search expert.
Claude helped me identify those functional gaps. I then asked Claude to generate the exact system instructions to build custom GPTs for all of those roles. It took hours, but I generated custom instructions for 18 agents in one weekend.
Then I made sure these agents “knew” about each other so they could hand off work. For instance, my newsletter editor agent can take an article I wrote and pass it to my LinkedIn specialist agent.
When reviewing the generated instructions, I spent time refining them because some things didn’t fit my style. For example, in my business, “Mindstate” is spelled as one word, but AI naturally defaults to two words (“Mind State”). If I didn’t correct that in the custom instructions, it would misspell my brand name every time.
Out of those 32 agents, I regularly use about five of them. But it’s great knowing that if I need a specific skill set, I have an agent built for it.
For anyone listening, the key lesson is: don’t just build a general “marketing expert” agent. A marketing expert prompt yields generic results. You want to build agents with deep, narrow expertise in their specific lane. If you hire a human who claims to know everything about marketing, sales, SEO, and operations, they’ll give you mediocre results. You need to break those skills out.
Melina Palmer: Absolutely! Think about it like onboarding a human team member. You need to invest time up front to help them understand your business and ask, “What questions do you have for me? What context is missing?” Having that dialogue is essential. Expecting any AI to be flawless on day one without guidance is unrealistic.
You don’t have to build 30 agents over a weekend. For corporate teams with limited resources, I advise leaders to step back and audit their workload: What tasks do your human team members actually enjoy doing? Where is their passion? And what are the mundane, repetitive tasks that everyone hates doing?
Data processing, initial synthesis, and formatting take up tons of time and energy. AI agents excel at those tasks. Delegating mundane work to AI frees up human energy for high-level creative and strategic thinking.
Moving Beyond Efficiency to Effective AI
Will Leach: I’ll take that a step further. I work with market research departments in big corporations. They are overwhelmed with data coming from everywhere, sitting in meetings all day, representing the voice of the customer, and trying to make sense of massive data sets. That workload is exhausting.
A few weeks ago at IIEX, I gave a presentation on “Effective AI” versus “Efficiency AI.” Right now, corporate America is focused almost entirely on Efficiency AI—doing work faster and cheaper. In theory, doing work faster frees up time to be strategic. But in reality, as technology makes research faster, companies don’t give employees time to think; they just demand more volume! It used to take two weeks to run a concept test, then overnight, and now managers ask, “Can you run three concept tests in four hours?” The efficiency gains are consumed by more grind work.
On stage, I showed a photo of an old, dark vending machine covered in cobwebs. But instead of candy, it dispensed market research reports. I told the audience, “As an industry, we have built the world’s most sophisticated vending machine.” A vending machine is fast, cheap, and accessible, but nobody respects a vending machine! Things that are fast, cheap, and easy do not command respect or strategic influence.
When corporate budgets get cut or layoffs happen, “vending machine” functions get cut first because they aren’t viewed as strategic thought leaders.
Instead of just using AI for efficiency, we need to use AI for effectiveness. How can we use AI to make researchers and marketers more impactful, strategic, and respected?
That’s why we built Bevy. We built skills into the AI so a researcher can take unstructured data—like 40,000 customer reviews—and run a deep psychological analysis on it automatically. We built an “Ad Advisor” skill that evaluates creative ads using behavioral science to check if they align with the customer’s non-conscious drivers. We built strategic copywriting skills. We expanded the capabilities of the team beyond basic data gathering into strategic application. That is how you use AI to do your future job, before budget cuts force the issue.
Applying Behavioral Science to AI Data Sets
Melina Palmer: How can teams ensure they don’t gloss over behavioral science when using AI? Most AI models are trained on standard survey data, but as behavioral economists know, what people say in a survey rarely predicts what they actually do.
Will Leach: That is the single biggest flaw I see! Most companies build AI personas or models trained on legacy survey data and stated intent. When people take surveys, they are operating in a rational, System 2 mindset. They overthink and post-rationalize.
If you train an AI persona on stated survey data, you are building a “very confident lie.” When you ask that AI persona to make decisions, it will sound extremely confident, but its recommendations will fail in the real world because the underlying data is flawed.
To bring true behavioral science into AI, you must recognize that stated data is unreliable. You need conversational, unstructured data—how people naturally talk when they aren’t being forced into survey boxes.
While human speech still contains post-rationalization, AI can be trained to detect subtle non-conscious “tells” or behavioral signals in text—like emotional motivations, cognitive biases, and psychological framing—that humans miss.
We trained our models on the Mindstate Model to scan unstructured text for these behavioral cues. When clients ask to feed basic 1-to-10 NPS scores into AI, I cringe. A numerical rating isn’t predictive of future behavior or true emotional sentiment. But if you feed the open-ended text comments from those surveys into a behaviorally trained AI, now you can extract real psychological insights.
Don’t just scale bad survey data faster. Use unstructured text and train your AI models to look for behavioral signals.
Melina Palmer: Can you share a concrete example of how behavioral signals are extracted from unstructured data using AI?
Will Leach: We conducted a project for Lululemon analyzing roughly 40,000 customer reviews. Customer reviews are naturally polarized—people write them when they either passionately love a product or absolutely hate it. That polarization gives you closer access to authentic feelings.
A standard AI tool will summarize reviews into basic themes: “it looks stylish,” “it’s comfortable,” or “it’s expensive.” Any basic LLM can do that.
However, when you train AI on behavioral science frameworks, it looks deeper. When a high volume of Lululemon customers talk about how a product “fits my unique personality and makes me feel like myself,” the AI identifies underlying non-conscious motivators:
- Esteem Motivation: The desire to feel respected, admired, and stand out.
- Autonomy/Individuality Motivation: The desire for personal freedom and self-expression.
Knowing those core motivators allows you to craft precise messaging. Instead of just saying “our leggings are stylish,” your marketing copy can say: “Designed to help you stand out as an individual and express your unique style.”
That single sentence connects two deep psychological drivers. A standard ChatGPT prompt won’t give you that strategic link unless you train it on behavioral science frameworks. From there, you can prompt your AI copywriter agent to generate 30 email variations leveraging those specific motivators.
Melina Palmer: That depth of understanding is key! People don’t explicitly say in a focus group, “I bought these leggings because I wanted to show Susan I’m better than her!” They won’t articulate that consciously. But analyzing word choices across thousands of reviews allows AI to deduce drivers like esteem or belonging.
You can then use AI to draft headlines, test them against simulated customer personas, and narrow down your top options before spending budget on live A/B testing or field experiments.
Using AI Personas for Strategic Decision-Making
Will Leach: You hit on a massive topic: AI Personas. Creating digital representations of target customers was a huge topic at IIEX.
Let’s say your target consumer profile is named “Samantha.” You can build an AI persona representing Samantha based on your research data.
Here is the big warning when using AI Personas: Do not use an AI persona as a final testing vehicle.
Brands will show an AI persona three ad concepts and ask, “Samantha, which ad should we launch?” That is a mistake! That is the equivalent of asking a customer who knows nothing about marketing or strategy to design your campaign.
Instead, use the AI persona as a sounding board. Ask her: “Samantha, what thoughts or feelings come to mind when you read this headline?” She can give you qualitative reactions based on her psychological profile.
We work with Unleash Brands here in Dallas—they own Sylvan Learning and several family entertainment brands. Their brand managers use customer AI personas on their phones right in the middle of meetings. If a team is debating a new menu or program idea, they can query the persona: “We’re thinking about this concept—how does this land with you?”
The AI queries their past research data and behavioral drivers to give immediate feedback. It doesn’t replace final market testing for major decisions, but out of the 25 small decisions a brand manager makes every day, running 15 of them past an AI persona grounds daily choices in the voice of the customer.
Melina Palmer: That is such a practical way to streamline decision-making! We could do a whole second episode on AI Personas alone!
For listeners who want to connect with you, learn more about Mindstate Group, or check out Bevy, where is the best place for them to go?
Will Leach: Head over to mindstategroup.com to learn about Bevy and our consultancy. To follow my daily thought leadership, connect with me on LinkedIn—that’s where I post most frequently!
Melina Palmer: Perfect! We will have all those links in the show notes. Thank you again, Will, for joining me today. It’s always a pleasure chatting with you!
Will Leach: You too, Melina!
Wrap-Up and Final Thoughts
Melina Palmer: Thank you again to Will Leach for joining me on the show today. What got your brain buzzing in today’s conversation?
For me, it comes back to the conversation around efficiency versus effectiveness. It’s easy to treat AI like a human employee and try to force it into a traditional job description box. In business, small companies often look for “unicorn” hires—someone who can handle high-level strategy, copywriting, graphic design, SEO, social media, and project management all at once. That’s unrealistic for a single human.
The beauty of AI is that you can build hyper-specialized agents that do one specific job exceptionally well. You can have an SEO agent, an email optimization agent, an ad reviewer, and a research assistant. You don’t need 30 agents, but building three to five specialized agents to support your team’s actual needs can transform your workflow.
Personally, I’ve found that using AI to draft content from scratch is frustrating—it never sounds like me, and I end up rewriting it anyway. I realized I actually enjoy the writing process! What I don’t enjoy is spending hours summarizing research or building reading lists. Using AI for strategy planning, literature reviews, and research summaries frees me up to do the writing that is uniquely mine.
Audit your team’s workload: What tasks do you love doing, and what tasks drain your energy? Build AI agents to handle the draining tasks so you can focus on effectiveness and high-value work.
What are you going to do to enhance your use of AI starting right now? Share your thoughts with me on social media! You can find me as @TheBrainyBiz across platforms and as Melina Palmer on LinkedIn.
All relevant links, past episodes, and Will’s book, Marketing to Mindstates, are waiting for you in your app and at thebrainybusiness.com/597.
Thank you again to Will Leach for a fantastic conversation, and thank you all for listening! Join me next time for another episode of The Brainy Business Podcast. Until then, thanks again for learning with me, and remember to be thoughtful!
Announcer: Thank you for listening to the Brainy Business Podcast. Melina offers virtual strategy sessions, workshops, and other services to help businesses be more brain friendly. For more free resources, visit thebrainybusiness.com.
