How to Build an AI Strategy That Doesn’t Suck (+ Actually Delivers ROI)

Reply to this email + I’ll send you your new nickname.

the problem: We know damn well that using AI in your biz feels like eating an elephant right now. Sometimes you don’t know where to bite first (any tusk people in the house?).

Most teams obsess over tools as execs deal with FOMO (McKinsey breakdown below).

Forget that. The companies winning with AI don’t start with the tools. They begin by securing buy-in, problem mining, and using AI as a tool to build a robust roadmap.

the strategy: Alex and Arman, the peeps behind Tenex and the company that started ultrathink, have run hundreds of these AI colonoscopies for $50B giants + scrappy upstarts.

Put your gloves on, Becky. It's time to avoid 2026 irrelevancy:

The first three steps determine whether AI becomes a profit engine or a pilot scrapyard. The last two, you’ll 100% want to see our complete playbook to understand how problem mining evolves into a tactical company report and roadmap.

get executive buy-in or bust: AI adoption dies in the hands of bureaucracy. You need a bigwig with social currency to become a bulldozer, champion the change, and defend the work.

New hires and mid-levels rarely have the juice. But in rare cases, they can spark change from the ground up (but they need to show a big win w/ scalable returns).

Before anyone starts pitching moonshot AI ideas, you need to be aligned on how AI adoption actually works. Show them these phases:

pro tip: Stay away from step 4 before crushing phases 1–3. It’s complicated without the foundational layer in place.

make the company confess: People won’t always tell managers what’s broken, but they will tell a survey.

Steal ours. We worked hard on it. Send it out to tenured ICs, mid-level managers, and other operators. The goal is to get at least 30 back.

Each question has a purpose:

interrogate power users: Skip the one-way mirrors and handcuffs—but record everything. We do this problem mining stage virtually, transcribing everything (crucial for the next stage: Claude Code synthesis).

We use Notion’s AI meeting recorder.

Here’s the spread:

Arman asks the same questions every time (+ personalizes them based on the surveys).

If they flagged “report building sucks” in the survey, go deeper in the call.

“You mentioned reporting takes forever—walk me through the exact steps. How many hours a week? Who else touches it? What breaks?”

➿ Enjoying the outline? Grab the full playbook of the exact $50K process (plus the steps that turn all surveys and interviews into an undeniable roadmap).

McKinsey just dropped its 2025 AI Report, and everyone’s talking about it.

Here’s the big three takeaways, but in English:

How does Google not win it all once it’s all said and done?

ELI5: They’re owning the entire race.

Not just saying that bc we have Logan Kilpatrick on air with us soon.

Google’s the only company that makes the apps, the brains, the cloud the brains live on, the math to power it, and the chips underneath it all.

Take a look at the current standings on Polymarket’s best AI model of 2025 bet (we see you, Moonshot).

“I built this to keep up with the technical wizards in the office,” Alex Lieberman (Tenex’s cofounder) wrote in our team Slack channel.

Here’s the full prompt:

Objective: Create a comprehensive research report on [INSERT TOPIC HERE]. The goal is to build a deep conceptual understanding of the topic — from its theoretical foundations to its real-world applications — so that I can use this as a launchpad for further exploration.

Audience: A non-technical but intellectually fluent reader. I’m comfortable following complex discussions, but I’m not formally trained in this technical domain.

Tone & Style: - Write in a clear, structured, and explanatory style. - Include technical depth, but translate every piece of jargon into plain English. - After each complex term, formula, or mechanism, provide: a) A plain-language translation (explain it like you’re teaching an intelligent layperson). b) A real-world, tangible example or analogy that makes the idea concrete.

Content Requirements:

1) Foundations Section - Define the core principles, vocabulary, and historical context behind [TOPIC]. - Explain why this field exists, what problems it solves, and who pioneered it. - Use simple examples to show the basic mechanics at play.

2) Core Concepts & Mechanics Section - Dive into the key theories, processes, or frameworks that make up the topic. - Introduce any math, algorithms, or scientific models central to the field. - For each technical concept, pair the explanation with: a) A plain-language breakdown. b) A real-world illustration (e.g., from everyday life, business, nature, or technology).

3) Applications & Implications Section - Show how [TOPIC] is applied in real-world systems, industries, or technologies. - Include notable case studies or examples that demonstrate its impact. - Explain why understanding these concepts matters — what it enables or changes.

4) Integration & Broader Context Section - Connect this field to adjacent domains (e.g., how it interacts with math, physics, biology, economics, etc.). - If relevant, trace how the theory translates into practice (e.g., from code → circuits → behavior). - Highlight open questions or ongoing research frontiers.

5) Formatting & Accessibility Guidelines - Use clear headings, subheadings, and summaries at the end of major sections. - Define jargon inline, not in a glossary. - Use metaphors, analogies, or thought experiments liberally. - If helpful, include short “mental models” or “rules of thumb” to aid intuitive understanding.

Output Goal: A research-style explainer (typically 3,000–5,000 words) that is educational, accessible, and intellectually rigorous — something that helps a curious but non-specialist reader gain a working, conceptual mastery of [TOPIC].

background: Google rolled out a new File Search system in the Gemini API—essentially pushing Retrieval-Augmented Generation (RAG) into the mainstream by making it automatic instead of a developer chore.

in plain English: Less hallucinations. You know when you ask ChatGPT to reference something from a file you uploaded days ago and it responds with confidence—and gets it completely wrong?

That’s because large language models are brilliant pattern-matchers by definition. They are not trained in the ancient ways of the Dewey decimal system. RAG is the mechanism that gives them an actual reference system instead of hoping they “remember.” Now Gemini got the upgrade.

Think of RAG like this:

Your company already has answers—they're just scattered in:

Normally, when you ask AI a question, it guesses based on its training data. With RAG, the AI literally goes and reads your real files first—then answers.

So instead of:

“How many PTO days do we get?”

AI (guessing): “Most companies give 10–15 days.”

You get:

“How many PTO days do we get?”

AI (with RAG): “You get 20 days. Page 4 of your HR handbook. Here’s the line.”

the point: For a year, RAG lived in GitHub repos and machine-learning Ted Talks. Today, it’s moving into the “normal feature” phase.

apply it: Anywhere your team currently says “where’s that doc,” you probably have access to a RAG-assisted helper at this point.