Friends of TableTalkAI!

Happy Sunday and I hope you had a great week.

Corporate America just discovered something fascinating, terrifying, and deeply relatable.

AI tools are not free.

Not even close.

In fact, companies reportedly burned through $500 million in AI tool usage in a single month, not because they were curing disease, reinventing supply chains, or building the next trillion-dollar company.

Nope.

A meaningful chunk of that spend came from employees using very expensive AI subscriptions to do very ordinary things.

Checking the weather.

Summarizing emails.

Asking a coding agent to think deeply about something that probably needed a calendar invite, not a $100-per-hour compute session.

Welcome to the AI expense report era.

We had the “AI is magic” phase.

We had the “AI will replace everyone by Tuesday” phase.

We had the “AI pilots everywhere, ROI nowhere” phase.

And now we have arrived at the phase every CFO knew was coming:

“Who approved this?”

The answer, apparently, was everyone and no one.

There is now a phrase for this behavior: tokenmaxxing.

That is when people burn absurd amounts of AI compute because they mistake usage for productivity.

It is the enterprise software version of ordering the most expensive bottle of wine at dinner because someone else has the company card.

The problem is not that employees are using AI.

They should be.

The problem is that too many companies gave everyone access to powerful tools without usage policies, workflow discipline, spending controls, or a clear definition of what “good” looks like.

One reported example had Uber’s CTO blowing through the company’s entire 2026 Claude Code budget by April.

Amazon reportedly pulled an internal AI usage leaderboard after employees started chasing token counts instead of meaningful output.

That is where things get interesting.

Because the lesson is not “use less AI.”

The lesson is “use AI better.”

Heavy Claude Code users reportedly burn about 10x more tokens than moderate users, while only producing about 2x the output.

That is not a productivity curve.

That is an all-you-can-eat buffet with very expensive shrimp.

And as someone who respects both buffets and shrimp, I say that with love.

Here is the mental model that separates casual AI users from serious ones.

AI is not deterministic.

It is probabilistic.

That means when you type a prompt and get an answer, you are not receiving “the answer.”

You are receiving one possible answer.

One roll.

One pull.

One version of millions of possible outputs.

Most people prompt once, read the first response, and either say “this is amazing” or “this is useless.”

The best users do something different.

They generate options.

They run the same idea multiple ways.

They compare outputs.

They pick the strongest one.

That is the slot machine reality of AI.

The first answer may be fine.

The fifth may be better.

The fiftieth may be the one that makes you look annoyingly smart in a meeting.

This is why some solo founders are now building five products at once using AI agents, while others are still complaining that ChatGPT “didn’t get it.”

It is not just the tool.

It is the operating style.

Josh Pigford, who sold his last startup for $4 million, has reportedly developed specific AI collaboration skills, including:

A /build skill for structured product development.

An /adversarial-code-review workflow where Claude and GPT critique each other.

A /but-for-real prompt that forces AI to catch its own overly confident nonsense.

That last one deserves a standing ovation and possibly its own office mug.

Anthropic’s Claude Opus 4.8 shipped with a feature that sounds small but may be very important: effort controls.

Four levels:

This lets you decide how deeply Claude should think before responding.

That matters.

Because not every task deserves the same cognitive budget.

Proofreading a sentence does not require the same depth as building a pricing strategy, reviewing a contract, or planning a product roadmap.

Before effort controls, AI often treated too many tasks like they were roughly equal.

Now you can tell it when to glance and when to grind.

That is a big shift.

It turns AI from a generic response machine into something closer to a configurable thought partner.

Which means the future of prompting may be less about writing longer prompts and more about setting the right thinking mode before the work begins.

Here is another quietly massive shift.

AI agents now represent 48% of visitors to documentation sites.

That means your documentation is no longer just being read by humans.

It is being read by Claude Code.

Cursor.

Coding agents.

Support agents.

Internal copilots.

Machine customers with infinite patience and zero tolerance for vague marketing fluff.

They read every endpoint.

They compare options.

They check schemas.

They look for examples.

They do not care that your homepage says your API is “seamless,” “robust,” or “enterprise-grade.”

They want to know what the thing does, how it behaves, what breaks, what the limits are, and whether the examples actually work.

This changes what good documentation means.

Good docs now need to serve two audiences at once:

Humans who need clarity.

AI agents who need structure.

That means businesses need better schema markup, clean examples, honest benchmarks, clear implementation paths, and less “future of work” confetti.

Because the agent is not impressed.

The agent is parsing.

Now let’s talk about the part that feels a little weird.

A manager recently sent an employee a performance review that was so thoughtful, detailed, and emotionally specific that the employee cried reading it.

Except the manager did not write it.

His HR software generated it from her file.

He skimmed it.

He hit send.

She never knew.

That is not science fiction.

Versions of this are already happening.

AI is sliding into workflows so quietly that people may not know when they are receiving human effort, AI-generated effort, or a lightly skimmed corporate casserole made from both.

And yes, AI can help managers write better.

That is good.

But there is a difference between using AI to express what you genuinely think and using AI to simulate care you did not actually give.

That line is going to matter.

A lot.

The question is not whether AI can write something thoughtful.

It can.

The question is whether the human behind it did the thinking, noticing, and caring first.

AI is also changing how product teams work.

The old flow looked something like this:

Have a meeting.

Write a spec.

Debate the spec.

Design a mockup.

Debate the mockup.

Build a prototype.

Discover the workflow breaks in real life.

Have another meeting.

Now teams are starting to prototype before the spec is fully baked.

FloQast reportedly tested a complex workflow by building a code-backed prototype with real customer data, which helped catch interaction problems that only show up when logic is live.

Merkle used AI to generate layout variations and realistic content while staying grounded in its design system.

That is the shift.

AI is compressing the distance between idea and test.

The best teams are not using AI to avoid thinking.

They are using it to think in motion.

While corporate teams are trying to stop employees from spending $100 an hour asking AI to write meeting summaries, researchers are doing something extraordinary.

AI has now mapped 1 billion proteins, expanding the known protein universe by more than 800 million entries beyond the previous record of 200 million.

That is not a small improvement.

That is a new map of biological possibility.

A free tool called ESMFold2 reportedly beats Google DeepMind’s AlphaFold3 at predicting how proteins interact, which could accelerate drug discovery, biology research, and the understanding of diseases.

This is the duality of AI in 2026.

In one room, someone is using it to model protein interactions.

In another room, someone is using it to ask whether they should bring a sweater.

Both may technically be valid use cases.

Only one should probably require enterprise approval.

The AI infrastructure boom is not slowing down.

Dell raised its AI server forecast to $60 billion.

Big Tech’s spending binge is still estimated around $700 billion.

And TSMC is warning that the next bottleneck may not be raw compute.

It may be energy efficiency.

That matters because if AI demand keeps growing, the winners may not simply be the companies with the biggest models.

They may be the companies that can deliver the most intelligence per watt.

The AI race is becoming an energy race, a hardware race, a cost-control race, and a workflow race all at once.

No pressure.

For businesses, one of the most important implications is this:

AI discoverability is becoming a real thing.

If AI agents are reading documentation, comparing vendors, summarizing options, and recommending tools, then companies need to think about how they appear to machines.

Not just humans.

Not just Google.

Not just social media.

Machines.

That means your website, docs, FAQs, pricing pages, product descriptions, benchmarks, and implementation guides need to be clear enough for AI systems to interpret correctly.

As one expert put it:

AI is the new internet.

Every business needs to be findable on it.

That may sound dramatic.

But so did “every business needs a website.”

Until it didn’t.

As AI costs rise, companies are being forced to answer the question they should have asked before buying every tool in sight:

What are we actually getting from this?

The best leaders are starting to measure AI ROI at three levels.

Company-wide efficiency gains.

Team-specific productivity improvements.

Project-level ROI.

That is the right direction.

But the measurement needs to be smarter than “tokens used” or “AI seats purchased.”

The right metrics are things like:

How much time did we save?

Did quality improve?

Did cycle times shrink?

Did we unlock work we could not do before?

Did we reduce rework?

Did employees make better decisions faster?

Did customers get better outcomes?

AI should be treated as a capability multiplier, not a magic replacement wand.

The companies that understand that will win.

The companies that do not will discover that “AI transformation” is a very expensive way to produce more Slack summaries.

Claude is Anthropic’s AI assistant and the latest Claude Opus 4.8 release includes effort controls, allowing users to choose how deeply the model should think before responding. This is especially useful for complex strategy work, product planning, technical reviews, research synthesis, and multi-step reasoning. Use lower effort for simple edits and higher effort when you need real analysis, edge cases, trade-offs, and implementation thinking.

Best for: deep thinking, writing, research, product strategy, code review, document analysis, and complex planning.

Wispr Flow is a voice-to-text tool built for people who think better out loud than they type. It captures messy spoken thoughts and turns them into clean, structured text that can be used in ChatGPT, Claude, Gemini, or any other AI tool. This is particularly useful for founders, operators, product leaders, and anyone who has strong ideas while walking, driving, pacing, or pretending they are not talking to themselves.

Best for: turning spoken ideas into prompts, memos, outlines, emails, strategy notes, and first drafts.

Adapt is a universal AI agent designed to connect with business systems like data warehouses, CRMs, and internal tools. It helps teams ask questions in plain English and get answers based on live company data, often directly inside Slack or similar workflows. This is where AI starts moving from “write me something” into “help me understand what is happening in the business right now.”

Best for: revenue operations, analytics, sales teams, customer success, and internal decision support.

ESMFold is a protein structure prediction tool from the Meta AI research ecosystem. It helps researchers predict protein structures at massive scale and has been connected to the mapping of hundreds of millions of protein structures. For most business users, this will feel very far from the day-to-day AI tools we normally discuss, but it is an important reminder that AI is not just writing emails. It is changing science.

Best for: biological research, protein structure prediction, drug discovery exploration, and scientific AI workflows.

Mintlify is a modern documentation platform designed to make technical documentation easier to create, maintain, and navigate. As AI agents increasingly become readers of documentation, tools like Mintlify matter because they help companies structure content clearly for both humans and machines. If your product relies on APIs, integrations, developer adoption, or technical onboarding, your docs are no longer just support material. They are distribution infrastructure.

Best for: developer documentation, API docs, technical onboarding, AI-readable docs, and product-led growth.

Norton Neo is an AI-powered browser designed to help users think, search, and act across the web with more assistance built into the browsing experience. The idea is to reduce the constant tab-juggling, copy-pasting, and context-switching that defines modern knowledge work. It is part of a broader trend where browsers are becoming agentic, not just passive windows into the internet.

Best for: AI-assisted browsing, research, web tasks, summarization, and reducing tab chaos.

Julius AI is a data analysis platform that lets users upload datasets, ask questions, create charts, build dashboards, and compare analytical outputs from different models. It is especially useful for non-technical users who want to analyze spreadsheets, business data, survey results, or financial information without needing to write code. It is also a good playground for testing how different models handle the same analytical task.

Best for: data analysis, dashboards, charts, spreadsheet review, business intelligence, and model comparison.

Use this when you want AI to reflect patterns back to you based on your prior conversations, writing, and working style.

Prompt:

“Based on every conversation we’ve ever had, build me a complete profile of who I am. Include my communication style, my values, what I seem to care most about, what I may be insecure about, what motivates me, what patterns you notice in how I make decisions, and what I may be secretly working toward. Be specific, practical, and honest, but do not invent anything you cannot reasonably infer.”

Use this when you want a sharper mirror and are willing to hear where your stated goals and actual behaviors may not line up.

Prompt:

“Tell me the contradictions you’ve noticed in me. Where do my words not match my actions? Where do I say I want one thing but behave in a way that suggests another? Be direct, specific, and constructive. Do not flatter me. Focus on patterns that could help me make better decisions.”

Use this to model where your current habits, decisions, and operating style may lead over time.

Prompt:

“Predict the exact version of me in five years if I keep operating the same way I currently am. Include my career, relationships, health, finances, reputation, mindset, and day-to-day life. Then give me a second version of the future if I make three high-leverage changes starting now. Be realistic, not motivational.”

Use this when you are trying to avoid accepting the first AI answer and want real variance.

Prompt:

“Give me 3 different angles on [your task] for [your success criteria]. Make each approach distinct. For each one, explain the strategy, the audience it works best for, the trade-offs, the risks, and what would make it succeed. Do not give me three versions of the same idea with different wording.”

Use this with Claude when the task deserves deeper thought.

Prompt:

“Set effort level to HIGH. I need you to thoroughly analyze [topic] with deep consideration of edge cases, alternative perspectives, second-order consequences, implementation challenges, risks, and practical next steps. Do not rush to a conclusion. Think like an experienced operator who has to make this work in the real world.”

Use this after a task to help the AI adapt to your preferences and improve future collaboration.

Prompt:

“After completing this task, update your understanding of my preferences and working style based on what worked well and what did not in this interaction. Tell me what you learned about how I like information structured, what tone I respond to, what details matter to me, and how you should improve the next version.”

Use this whenever accuracy matters and you want the AI to separate confidence from guesswork.

Prompt:

“Generate this response, then review it for accuracy. Flag any claims that might be uncertain, outdated, exaggerated, unsupported, or require verification. Separate confirmed facts from assumptions. If you are not sure about something, say so clearly and recommend what should be checked before this is published.”

AI has officially entered its expense report era.

Which, honestly, was inevitable.

Every revolutionary technology eventually has to meet accounting.

The magic is still real.

The breakthroughs are still massive.

The productivity gains are still there.

But the free-for-all phase is ending.

The companies that win from here will not be the ones buying the most AI seats, burning the most tokens, or letting every employee turn Claude into an overqualified weather app.

The winners will be the ones that learn how to use AI with judgment.

More options.

Better prompts.

Clearer workflows.

Spending controls.

Human taste.

And maybe, just maybe, a finance team that gets invited to the AI strategy meeting before the invoice arrives.

Stay thoughtful, my friends.

And please, for the love of the budget, stop asking the $100/hour model if it is going to rain.

🧠 If you enjoyed tonight’s deep dive, forward it to someone in your network who wants to fully grasp AI in 5 minutes per day. They’ll thank you later.

Your slightly self-deprecating, definitely human narrators,
Anicia & Shane