The cost of failure is high. Think in terms of the 1-10-100 rule; $1 in prevention is $10 in correction and $100 in failure. Adapted for the AI era this gets worse:

$1 in design is $10 in engineering, $100 in tokens, and $1,000 in failure.

Catching flaws in design is cheap and easy to fix. Letting them slide into production can be catastrophic.

The Imprint

In The Future of Focus we talked about the Imprint: the point where human judgment and taste reign supreme. It’s the most critical concept for your AI implementation plan.

At Atmosphere we view AI as work augmentation. We’ve done the modeling on our actual hours spent and put sharp lines around where we expect high leverage from AI and where output is almost entirely human.

As a small firm, we need 2x blended AI leverage.

Our content, for example, is entirely human written. We use Claude for revision and proof edits, and Claude Design for visuals based off a hand curated brand kit.

We don’t use AI for any initial drafts or even concept suggestions. The quality is too low and doesn’t communicate in a voice we feel is authentic. That caps the AI leverage and we’re comfortable with the trade-off.

Three Core AI Use Cases

The AI implementations we’ve worked on all boil down to the following:

1. Personal productivity: research, writing, editing, coding, email etc.

2. Workplace productivity: augmenting specific workflows*

3. Product development: embedding AI into customer-facing products

* Many companies seek full workflow replacement. In our experience this almost never works.

Discovery

Map and quantify your workflows across the organization. Note that if you’re using AI to speed up a particular function (Sales), you need to ensure the rest of the organization stays in sync (SalesOps, Finance). If you don’t…

We call this the feral otter. The AI enhanced workflow starts to gobble up the rest of your company. Problems get worse and messier before any improvement.

The discovery work is painstaking but straightforward. Here’s a simple plan:

• Internal interviews with users on workflows (3-5 per area) depending on size

• Interview synthesis and workflow quantification

• Trade-off analysis of AI implementation vs. standard operating procedure

• Tool selection and platform choice

Your workflow scoring system needs to take several things into account at a minimum:

1. How many people touch it

2. Hours per week consumed

3. How repetitive the task is (score of 1 - 10)

4. How much human judgment is required (score of 1 - 10)

Where to apply AI

High-volume, high-repetition, low-judgment tasks pay back the fastest with AI. The graveyard for AI deployments is filled with projects that require human judgment and taste.

As a practical example, we worked with a company that needed to process ~200 leads per month to get 40% qualified with a close rate of 20%. That volume is ~2 SDRs going full tilt or ~$40k per quarter. It’s why SDR automation is such a hot topic.

TAM building and data enrichment is widely available and CRM systems have automated sequences (email etc.). That part is straightforward. We’ll cover the tools next time.

The problem was the upper funnel increase created a downstream mess in alerts, CRM data quality, human sales follow up and billing and revenue visibility. The solution was to step back and build AI enhancement into the other areas to calm the feral otter.

The lesson: own the system for your AI implementation

The end output is a structured quantification of your workflows (how many people, how repetitive a task, time spent on average etc.). The map is a spreadsheet and set of workflow diagrams. Every task in a row quantified with an assessment of AI lift. The top rows become your AI implementation plan for the quarter. The full list is your annual roadmap.

To give another practical examples for high-touch, BD type relationships. Lead sourcing gets high leverage, whereas follow up is quite specific and has a much higher judgment score. Repetition and judgment are coded scores, 1-10.

Change Management

Once you have your map, you need to spend several cycles deciding on the areas where AI can have maximum efficiency and ROI for your organization. But a plan that no one adopts is slideware. Which brings us to why most AI deployments are failing.

You need to socialize with your team. Your team members need to be bought in and frankly should be excited about removing tedious and repetitive work.

How to do this? Bring people in early. Make them part of the discovery process. Be transparent about where the judgment lines sit and where you believe that humans add the most value. You’ll be surprised during the discovery process about how nuanced work is and how hard it actually is to get a machine to repeatedly produce the right answer. It’s why you hired good people in the first place!

Build in review cycles where you sit together, review the roadmap, discuss tradeoffs, and most importantly, listen to concerns.

Fear will nuke your deployment. If the message gets delivered as ‘your job is going away’ adoption will be zero. AI without a human operator is essentially useless.

Next post: Which AI tools and platforms to actually use (or build your own)

Atmosphere (us) will do a multi-part series over the next 4 weeks to give you the language and use cases to make sense of it all, as well as some tutorials for implementing AI at your organization.

We will not cover everything. Niche experts go deeper in all of this. There is a plethora of click-bait content of ‘Be an AI expert in 30 days!’. We’ll do something different. A practical guide, direct from the operator’s seat. Tips you can actually use.

What is AI anyway?

It’s not a stupid question. Executives are all well informed, smart, read the AI news, but it’s still confusing. Today’s AI has morphed from the traditional textbooks of Machine Learning, Natural Language Processing, Computer Vision, and Generative AI. You’ll still hear these terms. But today’s AI in practice has coalesced around a smaller number of concepts.

Machine Learning

Traditional Machine Learning (ML) is very much alive. These are systems that learn from data to make predictions and decisions. They have wide adoption and are best when you have lots of historical data. For structured tasks, purpose-built ML models can be more accurate and consistent than throwing the problem at a general purpose LLM.

Dumping your data into a Frontier Model, like ChatGPT, and hoping for the best is a common (and bad) mistake. The systems, people, and techniques are different. There are other tools you should be using like R and Jupyter notebooks.

The operator question is whether the problem is better solved with ML or Frontier Models. Even that line is blurring because the models can run real analysis and code. The mistake is pasting data and trusting the answer rather than directing the model to run a specific ML analysis.

Frontier Models (LLMs)

All of the attention has shifted here. ChatGPT, Claude, Gemini. They generate new data, reason, produce text, create images and videos. They are increasingly bundled. When people say ‘AI’ in a meeting, 99% of the time they are referring to these models.

They are probabilistic, which is why they will confidently tell you the US was founded 357 years ago even as we limp toward our 250-year celebration. Then instantly backtrack when you point out the error. That’s hallucination.

They’ll also give you a different answer for the exact same question because they are non-deterministic (you might also hear the word stochastic). Again don’t worry so much about the definitions. We just want you to be aware of the vocabulary.

It’s why high-risk situations (legal review, accounting audits) almost always have a human in the loop. You shouldn’t trust AI to close your books autonomously or give you a precise legal opinion. I could make the same argument around software architecture and data, which is why people hate on vibe-coding.

Two concepts that come hand in hand with LLMs are training tokens and parameters. Training tokens are a measure of how much data was used to train the model. Order of magnitude trillions. Parameters are the resulting weights that give the model intelligence. A bunch of numbers organized into matrices. Order of magnitude hundreds of billions. Smaller models can be in the single (7b) or double digit billions (15b) if tuned to a specific domain.

Parameter count is still a good predictor of intelligence but it’s no longer the whole story. You’ll hear terms like Reinforcement Learning from Human Feedback (RLHF), Fine-Tuning, and Retrieval Augmented Generation (RAG). These are techniques that can help a well-trained model beat a bigger raw one.

Specialized Models

These models are trained to do a specific thing extremely well. Examples here are text to speech models (ElevenLabs), speech to text models (Deepgram) or OCR models that can parse a PDF or scan a receipt. Frontier models get the press. These models do half the actual work.

AI Inference

This is the act of actually serving the model to a user. Processing a prompt, generating a response. The labs combine this all. What you see on your monthly bill is the tokens consumed and the aggregate cost. Typically in cost per 1m tokens, segmented by output and input tokens (Google, OpenAI, Anthropic etc.). But if you’re building your own with an open-source model you need to think carefully.

Inference can be a huge driver of cost and latency. More simply put, you actually have to pay for the model and compute. If you forget everything else, just remember this is where your AI bill will actually sting.

Agents

Where the industry momentum is heading. It’s a lengthy topic that we’ll cover in a separate post. You will hear these words in conversations.

• Harness: the wrapping infrastructure and evaluations

• Skills: the specific capabilities and guidelines for what the agent can do

• Tools: functions the agent can call (data fetch or CRM record updates)

• Guardrails: the instructions of what an agent can/can’t do (no processing payment)

• Model Context Protocol (MCP): the common standard to connect AI to data/tools

Agents can do things instead of just generate content which is why the industry is leaning this way. The terminology can vary a bit across providers so treat this as directional.

The alphabet soup decoded

You may hear all of these terms spat out in a meeting. You don’t need to memorize any of this. Each has a different underlying technology. Don’t spend the time trying to understand the difference between transformers, neural networks, or diffusion models unless you’re a tech leader. You won’t need it in normal operating decisions. Worse, when you use it incorrectly you instantly lose credibility.

The state of the art changes every day so by the time you learn it you’ll already be behind.

• Large Language Models (LLMs): reasoning, code/content generation, ChatGPT

• Text to Speech (TTS): converting text into speech, ElevenLabs

• Speech to Text (STT/ASR): converting live speech into text, Deepgram

• Text to image (T2I): take a prompt and synthesize a digital asset, Midjourney

• Text to video (T2V): take a prompt and create a video, Veo

• Image to video (I2V): take a reference image and generate a video, Runway

• Video to video (V2V): adjust a video for lighting, character etc., Runway

• Multi-modal LLMs: unified models across text, audio, images, and video, Gemini

• Graphics Processing Unit (GPU): massively parallel chips for training/inference

• Tensor Processing Unit (TPU): Google’s chips optimized for neural network math

• Pre-training: the initial phase of training a model on tons of data

• Post-training: everything done to the model afterward to make it useful and safe

Next up: how to start on your AI implementation. Subscribe so the next one hits your inbox!

Subscribe now

We set out to understand which tools actually stick and why. So we went straight to the source, 30+ founders, executives, and investors at the leading edge of AI.

We expected to find smarter systems, better prompts, and a few useful tools. What we found was something else entirely.

A small group of operators has pulled so far ahead that their workflows bear almost no resemblance to those inside even the most established companies. They’re not using better tools, they’ve reorganized how they work entirely.

This report is a map of what the front line now looks like. Four discoveries from the leaders who have crossed to the other side, and the principles that separate them from everyone else.

Interviews included individuals from:

The Frontier Has Split

Across our interviews, operators fell into three distinct groups.

At the top: builders, operators who have constructed proprietary AI systems trained on their own data, customers, and workflows. Every interaction makes these systems harder to replicate.

Below them: power users, sophisticated operators running deep tool stacks but still manually bridging the gaps between them.

And then: adopters, experimenting with AI tools, but layering them onto existing ways of working. Inside some of the most prominent companies we spoke with, there is no unified stack, no shared best practices. Individual employees are often paying for their own subscriptions.

The distance between these groups is growing because the builders’ systems compound while everyone else restarts from zero with every new tool.

Across all three tiers, one pattern held: AI-native tools are largely companion tools. They augment existing workflows rather than replace them.

So which companions earn a permanent place?

The Operator Stack

We asked operators at the leading edge of AI a simple question: what tools do you actually use?

Across interviews, the same structure emerged.

This list reflects the preferences of the operators we interviewed, not a comprehensive market map.

Every operator we spoke to had strong opinions about tools. But the deeper we pushed, the conversation shifted, from what they use to how they act.

Subscribe now

Today’s advantage isn’t artificial. It’s knowing where AI ends, and you begin.

AI has given everyone the tools to produce good work. But when good becomes free, great becomes more valuable. We found the difference between the two isn’t more AI, it’s knowing when to put the human back into the machine.

We call this The Imprint.

The moment where judgment overrides generation. Where the human hand re-enters the work and makes it unmistakably yours.

Every operator, team, and company has an Imprint. The question is where yours sits. Step in too early and you lose the leverage AI provides. Step in too late and the work becomes generic.

The people pulling ahead aren’t defined by their AI stack. They win by having the strongest sense of where their Imprint is.

Finding it isn’t intuitive, but across more than thirty interviews, four discoveries surfaced that reveal how the best operators have found theirs.

01

The Last 10% Is the Only 10% That Matters

AI compresses production, but expands judgment. The last 10% isn’t just the finishing touch. It’s becoming the most valuable work in the economy.

There’s a phrase in engineering called the ‘march of the nines’. It takes the same effort to go from 9% to 90% as it does to go from 90% to 99%. The closer you get to completion, the harder each improvement becomes.

AI has shifted work in a similar way. Until recently, the grind from blank page to draft was where most of the work lived. AI has collapsed that middle. A deck that took two weeks now takes two hours. Research that consumed an afternoon now takes minutes.

What AI hasn’t changed is the effort required to turn good work into great work. That effort lives at the end.

A creative agency founder, working with leading AI companies, uses Midjourney and ElevenLabs to get concepts in front of clients three times faster than the old model of scripts, stock imagery, and decks. But they won’t use AI for the finished work. “It doesn’t allow enough room for the human.” AI is “essentially a sketching tool, it gives us direction to where we’re going.” The final output? “The last 10% is always done in real life, with a human touch.”

An engineering lead at a major cloud platform described building infrastructure tools that reach 90% accuracy, then hitting a wall. “If you can’t trust it 100%, you’re not gonna give it full control. That last 10% is the hardest to bridge.”

But here’s what separated the builders from everyone else. They’re not guessing when to step in, they’ve engineered it. Confidence thresholds at every stage, back-testing on every output, and human override built into the system by design. The Imprint isn’t instinct for these operators; it’s infrastructure.

The CTO of a Series A company, whose entire product is generative AI, framed it most directly: “When execution becomes simple, judgment becomes essential.”

AI has made the first mile nearly free, but the last mile just got longer, and now it’s the only mile that differentiates you. One of the most advanced builders we interviewed put it plainly: “Are we going to crank it all out faster? Or are we going to keep the time we have and raise the quality instead?” That choice — speed or craft — is quietly becoming a dividing line.

The Human Signal

AI is very good at producing the right answer. But people rarely respond to the right answer. They respond to the human one, rougher at the edges, more raw, more emotional, harder to replicate. The part that can only come from the last 10%.

The evidence is everywhere. Vinyl hitting a billion dollars in U.S. revenue last year, driven by Gen Z. Film photography sales tripling in five years. Rosalía singing German opera at Berghain. Even the recent demise of the metaverse. In a frictionless world, friction becomes the unmistakable signal that a human was here.

Rosalia, a Spanish popstar, singing German Opera = Friction as a signal.

The same reflex showed up in our interviews. “AI waters down ideas.” “People are craving human touch.” “It breaks at the final layer.” One more round of prompt roulette rarely makes the output better. It makes it blander. That’s motion masquerading as progress.

The conventional explanation is that AI lacks taste. AI has taste, but its default taste is bad. The chatbot’s voice is earnest. The image generator’s output is emotionally vacant. People clock it immediately. This matters because when someone senses AI in the output, their trust decreases. Not because the information is wrong, but because the presence of a machine implies the absence of care. The top operators know this and treat AI’s default taste the way a director treats a first rehearsal - raw material, not the performance.

The last 10% isn’t just where the human adds value, it’s where the human removes the residue of the machine. The time AI frees up is being reinvested here.

AI closed the gap between a blank page and a competent draft. It did not close the gap between good and great.

02

The Discipline of Less

Your edge doesn’t come from what you add, it comes from what you remove. The best operators are using constraint as a forcing function.

Norway is the most decorated Winter Olympic nation in history, despite having 60x fewer people than the USA. What explains this? The philosophy is counterintuitive. Athletes spend the majority of their time at low effort, only pushing when it counts. They don’t chase new methods or fads. They only add something when they’re certain it improves the system.

The Norwegian Method in action: Klæbo, 6x golds in ‘26, most in Winter Olympic history.

The same pattern showed up in our interviews. In a market flooded with new tools, the most disciplined operators rejected most of them. They don’t reach for new, they reach for useful.

Constraint as Strategy

An AI-native founder described his philosophy as “delete and eliminate as many unnecessary things as possible.” He’s deliberately focused on constraints to force sharper thinking. His company runs what would traditionally take over a hundred people with fewer than thirty. Not because he couldn’t hire, because he chose not to. This deliberate friction didn’t slow him down; it made him innovate.

A founder building enterprise AI tools took the same approach, “to use the least amount of tools possible,” so that everything the company learns compounds in one place. Notion for documentation, Linear for product, Cursor for code, Claude Projects for everything from sales coaching to recruiting. Fragmentation kills the knowledge that makes a company smarter over time.

They chose Claude over ChatGPT because of editorial sensibility. They chose Granola because it disappears into the workflow. The selection criteria wasn’t “what can this do?” It was “does this earn a place in my day?”

This is what separates builders from adopters. Adopters chase the new tool. Builders go deeper on the system they already have.

The best operators are treating tool selection the way the top investors treat their portfolio: conviction about what to hold, discipline about what to cut.

03

Control is the Currency

Why is it that in every major technology wave, the early leaders rarely win?

From PCs to the internet to mobile, the pattern repeats. Google didn’t win the web by building websites. It won by controlling discovery. Apple didn’t win mobile by inventing the smartphone. It won by controlling the ecosystem. The technology commoditizes. The tools converge. And value shifts to the systems built on top, to the people who control how the technology is used.

The same shift is now underway in AI. The main LLMs can do roughly the same things at the same level. The tools built on top are converging too. If your AI stack looks like everyone else’s, it gives you the same advantage as everyone else. Which is no advantage at all.

It’s no longer about which tool to use. It’s which systems to build.

A founder scaling a global AI marketplace built multiple custom voice agents using ElevenLabs. One onboards over a thousand customers a month, 65% of all interactions. A second makes one thousand outbound sales calls daily. Within two months, they outperformed every team member except the founder. These aren’t off-the-shelf products, they’re proprietary systems trained on his business, his data, his customers. Every interaction makes them smarter.

“We tear up ideas, poke them,” said a veteran at one of the world’s leading AI platforms, describing how his agent teams work. Confidence scoring at every stage. Humans plugged in at every threshold. The system isn’t autonomous; it’s adversarial by design.

A creative technologist took it further: three different LLMs critiquing each other’s code before any human sees the output. “By themselves, they’re not good enough.” He built a cultural monitoring engine on the same principle, scraping sport, social media, and internet culture, delivering a daily Slack digest so analysts spend their time thinking, not searching.

Then there’s the CEO who turned Claude Projects into a hiring screener, a sales coach, and a deal copilot. “A second brain that is not biased.” She makes every final call. The system doesn’t replace her judgment, it sharpens it. And competitors can’t reverse-engineer it by buying the same tools.

The gap isn’t between companies that use AI and those that don’t. It’s between those that rent AI tools and those that own AI systems. The first group gets efficiency. The second gets the moat.

04

The Missing Layer

Every operator we spoke with described the same friction. Not a missing tool. A missing architecture. Right now, what’s filling that gap is you.

The tools work, they just don’t talk to each other. There is no shared context between apps, no continuity between sessions, and no awareness of what you were doing five minutes earlier in a different tool.

So you became the integration layer. Every time you copy context from one tool to another or re-brief AI on a decision already made, you’re doing the work the architecture should be doing. The most capable people we interviewed are spending real cognitive load on this infrastructure.

The founder of one of the fastest-growing AI companies is using a custom platform as “duct tape to hold over ten tools together”. A technical lead at a major cloud platform put it simply: the goal is an AI-native operating system, “because we’re still using Mac and Windows from the nineties.”

What operators are asking for is remarkably consistent: “A single layer that sits across everything”, “holds context”, “understands priorities”, and “acts proactively without being prompted”.

But the builders in our interviews aren’t waiting for it to arrive, they’re creating it themselves. Custom orchestration layers, tools stitched into proprietary workflows, context held together by force of will and engineering. The solutions aren’t elegant, but they compound, and that’s the point. The operators who control their own systems aren’t just moving faster on today’s work; they’re building the connective tissue that makes tomorrow’s work better.

The market has started to respond. Several of the operators we interviewed are already using products like Highlight to build context layers across their desktop, and open protocols like MCP are creating infrastructure that lets AI models reach into external tools and data for the first time. But these are early moves. What operators are actually describing is more fundamental: one surface, one context, one point of control that makes the architecture invisible. It doesn’t fully exist yet.

One quiet signal points to where it’s heading. Several operators told us that voice has become their default input mode. They don’t type prompts, they talk. Removing the friction of writing allows ideas to move closer to the speed of thought. This is a signal of what the missing layer actually looks like: not another app, but persistent intelligence you interact with as naturally as a conversation.

The companies that crack this won’t win by building a better point solution. They’ll win by making every point solution work together, with the human at the centre.

The Advantage Isn't Artificial

We set out to understand what the best operators are actually doing with AI. The pattern was clear.

The operators ahead of the curve know exactly where AI ends and they begin. They’ve condensed their stacks, built proprietary systems that compound with every interaction, and refused to automate the last 10% that only a human can do. That’s the Imprint.

As AI makes competent output ubiquitous, the advantage comes from what only a human can add: judgment, taste, conviction.

Every operator in this report is building something that didn’t exist two years ago. So are we. Atmosphere partners with founders scaling AI, health, and climate companies. If you’re navigating the same territory, we’d like to hear from you.

hello@atmospherepartners.co

This report was researched, drafted, and produced using a custom AI system built across capture, transcription, analysis, and editing, with human judgment at every stage.

Created by Atmosphere

Atmosphere is a growth partner for companies at inflection points, from initial go-to-market to enterprise transformation. We’ve spent the last two decades at companies large and small. Google, Pinterest, Saatchi & Saatchi, our own startups, and over a dozen as growth partners. We exist so the most ambitious ideas in AI, health and climate scale the fastest.

atmospherepartners.co

Subscribe now

Views from Atmosphere

We’ve spent the last two decades building and scaling companies at Google, Pinterest, Saatchi & Saatchi, Character.AI, our own startups, and over a dozen more as growth partners. Exits. IPOs. Product launches across 50+ countries. Platforms scaled from early traction to hundreds of millions of users.

That experience gave us a specific vantage point.

We sit at the intersection of technology, brand, and commercial strategy, close enough to the frontier to see what’s coming, deep enough in the work to know what’s real.

We also spend our time talking to the founders and operators building what’s next. Views from Atmosphere is where we share that perspective.

What you'll get

Bi-weekly essays and analysis drawn from our work with founders across AI, health, and climate. What's scaling, what's stalling, and why. Plus longer research reports when a topic demands it.

Who we are

Scott Coleman led growth at Google, Pinterest, and Character.AI before its acquisition. He advised the New York Times on AI and has scaled platforms from zero to hundreds of millions of users across 50+ countries.

Henry Simonds is a three-time founder with two exits who started in advertising at Saatchi & Saatchi before moving to the operator side, leading go-to-market and growth across AI, climate-tech, and consumer brands.

Views from Atmosphere is published by Atmosphere, a growth operating partner for companies at inflection points. If you found this interesting, share it with someone navigating the same territory.

Get in touch: hello@atmospherepartners.co