Subscribe now

TL · DR (The 30-second version)

AI fluency is not a skill. It is three builds an operator installs once and runs on for the rest of their career.

• Build One: The Research Architecture - Constrain Claude to a curated universe of Tier 1 publications, Tier 2 academic centers, and Tier 3 institutional research shops. Name the mode and output format before the question. Signal-to-noise ratio climbs from 20 percent to 90 percent in one move.

• Build Two: Charts and Dashboards - Drop a CSV, describe the question in plain English, iterate in 20-second cycles. The build-dashboard skill produces a self-contained HTML report. The analyst bottleneck collapses.

• Build Three: The Team of Agents - Install six core connectors. Write each agent as a one-page spec. Set the schedule. Direct, refine, add. Four to seven agents inside 30 days carry the rhythm work.

• The math - A senior operator carries roughly 50 hours of weekly admin load. The three builds compound to clear 12 to 18 hours inside 90 days.

• The window - Stanford finds 14 to 34 percent task-level productivity gains. NBER finds 89 percent of executives reporting zero enterprise productivity boost. HBR finds 91 percent of executives believe agentic AI will reshape work but only 13 percent have deployed agents. The space between is the operator opportunity.

If you only have 30 seconds, read this. If you have 20 minutes, read the rest.

Subscribe now

What is AI fluency for a senior operator?

There is a quiet thing happening inside companies right now, and the operators who are paying attention have already named it.

Something that took three days last year takes an afternoon now. A junior analyst with the right setup produces work that a senior used to own. The board deck that ate two evenings gets drafted before lunch.

Nobody is saying this out loud in the all-hands.

Everybody is noticing.

The research backs the pattern. A Stanford Institute for Economic Policy Research working paper of 5,179 customer support agents found access to a generative AI assistant raised productivity by 14 percent on average and by 34 percent for the lowest-skilled quartile. The 2025 Stanford AI Index reports the same pattern across drafting, customer communication, data preparation, and software development. The lift is real, and it concentrates where the work has been the heaviest.

The strange part is what the same research finds at the enterprise level. A National Bureau of Economic Research survey of 6,000 executives across the US, UK, Germany, and Australia reported 89 percent of them are seeing no productivity boost from AI inside their companies. The task-level data and the enterprise-level data are not the same data. The gap between them is the operator’s opportunity.

FIGURE 1 · The Productivity Gap

Source: Stanford SIEPR · Generative AI at Work (Brynjolfsson, Li, Raymond). NBER executive survey via Fortune. Chart by John Brewton.

The word for that opportunity is fluency. AI fluency is not knowing what AI can do. It is changing how your work actually gets done.

I am going to walk you through the three builds that take an operator from curious to fluent. I run all three of them within my own companies and within client engagements at 6AEP. Each solves a different administrative problem. Together, they replace about 30 hours of weekly desk work for the average senior operator.

You can build the first one this weekend.

You can build all three in three weeks.

How do you build an AI research workflow that pulls from elite sources?

Most people use AI for research the wrong way. They open a chat, type a question, and accept whatever comes back. The model pulls from wherever, the citations are inconsistent, and the output reads like a polished blog post written by a sophomore.

The fluent move is different. You build the research environment first. Then you ask the question.

FIGURE 2 · The Research Architecture

The architecture I use is called the elite-research skill. It lives in my Claude account as a reusable instruction set that is automatically invoked whenever I ask a research question. The structure has three layers.

Layer one is for source curation. The skill only searches inside a defined universe of top-tier publications, academic centers, and institutional research shops. The publication list includes the New York Times, Financial Times, Wall Street Journal, The Economist, Harvard Business Review, MIT Technology Review, and MIT Sloan Management Review. The academic list includes Harvard, Stanford, Chicago Booth, Princeton, Yale, and LSE. The institutional list includes Goldman Sachs research, Bridgewater Daily Observations, Blackstone insights, AQR, Andreessen Horowitz, and Sequoia Capital. The skill ignores everything else. No Medium posts, no LinkedIn essays, no SEO content marketing dressed up as analysis.

That single constraint changes the output more than anything else. The signal-to-noise ratio goes from roughly 20 percent to roughly 90 percent in one move.

Layer two is mode selection. The skill has four research modes, and the right one depends on what I am actually trying to do. Thesis Research is for stress-testing a position I already hold. Market Intelligence is for understanding a company or sector cold. Trend Scanning is for catching what is moving across an industry over 30 to 90 days. Deep Dive is for the questions where I need everything the institutional record contains on a narrow topic. I name the mode at the top of the prompt and the skill structures the search accordingly.

Layer three is the output format. The skill outputs in one of three shapes. A Research Brief is a 600 to 900-word memo with sourced bullets. An Advisory Intelligence Memo is the longer format I send to clients, around 2,000 words with a recommended action set. A Signal Pack is the lightest format, used for surfacing five to seven data points fast.

The build itself is straightforward. You write the skill as a markdown file with three sections: source list, mode definitions, and output templates. You install it in Claude as a custom skill. The model then automatically invokes the entire architecture every time you ask a research question.

The first time you run a real question through it, the difference is unmistakable. Yesterday I asked the skill to assess a sector trend across the last two years. The output came back in 11 minutes, with 14 sourced citations, all from inside the curated universe. The same question in an unstructured chat would have produced a 400-word summary with three of the citations made up.

That is the entire point of build one. You stop relying on the model to choose your sources and start treating them as part of your operating discipline.

Companion playbook: The Research Architecture · 8-page field guide (PDF). Curate, Mode, Format, Deploy. The full architecture with worked examples and a maintenance cadence.

How do you build charts and dashboards with AI?

The second build solves the analytics problem. Every operator I work with carries some version of the same complaint. They have data in five places, the BI tools their company pays for are too slow to iterate against, and the analyst who could run the query is booked through Friday.

Here is what fluent looks like. You drop a CSV into the chat. You describe the chart you want. Ninety seconds later you have a publication-quality visualization, with formatting you control, that you can drop into a board deck or a strategy memo.

Subscribe now

FIGURE 3 · Charts You Build Yourself

The Goldman Sachs research team has been blunt about where AI gains actually show up at the macro level. In a 2025 update to its productivity tracker, Goldman found no meaningful relationship between AI adoption and productivity at the economy-wide level, with a median 30 percent productivity gain confined to narrow, localized use cases. The chart-building loop is exactly the kind of localized use case the research describes. The lift does not show up in the firm’s headline numbers. It shows up in the work that one operator does on one Friday afternoon.

I want to show you exactly how this works, because the gap between people who know this is possible and people who actually do it is enormous.

Step one. You collect the data. This part is unchanged. You pull from Stripe or QuickBooks or your warehouse, and you save the result as a CSV. If the data is already inside a connected system, you can skip this step entirely and ask the model to fetch it through the connector.

Step two. You describe the chart in plain English. Not the type, the question. Something like: “Show me monthly revenue from 2024 to today, broken out by product line, with a 3-month moving average overlaid.” The model picks the chart type, builds the matplotlib or Plotly code, runs the code in its sandbox, and returns the chart as an image you can save.

Step three. You iterate. This is where the workflow earns its keep. You say “make the moving average dashed, drop the 2024 quarters, and change the title to fit a board format.” The new chart appears in 20 seconds. You repeat that loop until the chart says what you want it to say.

For anything bigger than a single chart, the build is a dashboard. There is a Claude skill called build-dashboard that constructs a self-contained HTML file with KPI cards, filterable charts, and sortable tables. You give it the data and the questions you want answered. It returns a single HTML file you can open in any browser, share with a colleague, or post on your internal wiki. The dashboards are static once built, which makes them easy to version and easy to trust.

The fluent operator runs through this loop two or three times a week. A weekly revenue review takes 14 minutes instead of two hours. A board prep cycle that used to involve four people coordinating across three tools collapses into a single afternoon at the desk.

The skill ceiling on this build is high. The space between you and having this part complete is one weekend.

Companion playbook: Charts and Dashboards · 8-page field guide (PDF). Collect, Describe, Iterate, Dashboard. The data-handoff discipline, the prompt shape, and the iteration loop.

How do you deploy AI agents that work on your behalf?

The third build is the one that changes the shape of the workweek. It is also the one that takes the most setup, which is why most operators stop at the first two and never get here.

An agent is a specialized AI assistant configured once, with a defined job, that runs on request. You direct it. It does the work. It reports back. The work happens 24 hours a day, 7 days a week, and you do not need to be present.

FIGURE 4 · Your Team of Agents

The institutional capital is voting on this build hard. Andreessen Horowitz’s 2025 survey of 100 enterprise CIOs found that average enterprise spend on large language models rose from $4.5 million to $7 million in 24 months, with another 65 percent jump expected to roughly $11.6 million the following year. The fastest-growing developer behavior in the same data is what a16z calls agentic inference, where the model acts across extended sequences rather than single prompts. Sequoia Capital went further. The firm’s 2026 essay Services: The New Software argues the next trillion-dollar company will not sell software tools but the actual work itself, delivered as a service powered by agents.

The adoption story is the inverse of the capital story. A 2026 Harvard Business Review survey reports 91 percent of executives believe agentic AI will reshape work, and 83 percent say effective adoption is essential to competitive position. Only 38 percent feel their organization is prepared. Just 13 percent have integrated agents into real workflows. 56 percent are still piloting under heavy human supervision. The operators building the agent stack right now are competing within a window that institutional data say is unusually wide.

FIGURE 5 · Belief vs Build

Source: HBR 2026 enterprise survey via HBS Working Knowledge. Chart by John Brewton.

You build agents through connectors. A connector is a software bridge that allows the model to communicate with one of your tools. Gmail, Google Calendar, Notion, Slack, HubSpot, QuickBooks, and Google Drive all have connectors that take about 90 seconds each to install. Once a connector is live, every agent you build can read from and write to that system on your behalf, with the permissions you set.

The plugin packages in Claude are what make the agent practical to build. A plugin is a bundle of connectors, skills, and pre-built templates organized around a function. The Sales plugin includes connectors for your CRM and email, plus pre-built skills for account research, call prep, pipeline review, and outreach drafting. The Operations plugin includes process documentation skills, runbook builders, and risk assessment templates. The Productivity plugin includes a memory system and a shared task file that the agents read and update.

Once the plugins are installed, the agent build looks like this.

You pick one piece of work that has been sitting on your desk too long. For most operators, that work is one of four things. The Monday briefing nobody has time to write. The pipeline review that gets done at 11 p.m. on Sunday. The customer follow-up sequence gets skipped half the time. The expense reconciliation that piles up for two weeks.

You write the agent as a one-page instruction. The instruction has three sections. What the agent does. What it has access to. What it sends back. The first version is short. Roughly 200 words.

You set the agent to run on a schedule. The schedule is the second thing that makes the agent practical. A Monday Briefing agent runs at 6 a.m. every Monday, reads your calendar and your inbox, summarizes what is coming, flags conflicts, and drops the result in your messages before you wake up. A Pipeline agent runs at 5 p.m. every Friday, reads HubSpot, flags deals that did not move, and drafts the follow-up emails.

You direct, refine, and add. Inside a month, a senior operator running this build has between 4 and 7 agents handling defined work. The agents are not magic. They do the things you would do if you had four hours a day you do not have.

The first agent feels like a trick. The fourth one feels like the floor.

Companion playbook: Your Team of Agents · 8-page field guide (PDF). Connect, Write, Schedule, Conduct. Connector install order, the one-page agent spec, four starter agents.

How much weekly time does an AI-fluent operator save?

I built all three of these inside the companies I run because the time math is brutal at the senior operator level. A Head of Operations carries roughly 50 hours of unbilled administrative load each week. Research, analysis, communication, scheduling, follow-ups, status updates, briefing prep, and the work of routing other people’s work.

1. Building one collapses the research load by about 70 percent.

2. Building two collapses the analysis load by about 80 percent.

3. Building three collapses the communication, scheduling, and follow-up load by 50 to 90 percent, depending on how aggressive you get with the agent set.

The fluent operator I describe in client engagements clears 12 to 18 hours of desk work per week within 90 days. That is not a productivity story. That is the entire competitive position of the job.

FIGURE 6 · The Operator Time Math

Chart by John Brewton → Data from 6AEP work

The macro forecast points in the same direction. Goldman Sachs Research projects that generative AI will raise global GDP by roughly $7 trillion and lift productivity growth by 1.5 percentage points over ten years. The historical parallel is the Solow paradox of the late 1980s, when computers were visible everywhere except in the productivity statistics for nearly a decade before the late-1990s acceleration. The current gap between Stanford’s task-level findings and the NBER enterprise survey maps onto that lag with high precision. The operators who close the gap first will be running the work when the macro numbers finally show up.

What does an AI-fluent operator’s week actually look like?

• Monday: A deck the board needs by Wednesday. You draft it in an hour, with the research agent feeding sourced bullets and the chart builder generating the four visualizations. You spend the second hour writing the narrative.

• Tuesday: Three back-to-back calls. Each one ends with the summary, the decisions, and the action items already written and routed by the meeting agent. You took no notes.

• Wednesday: The question that meant an afternoon of reading returns a sourced, current answer before lunch. You make the call by 2 p.m.

• Thursday: Your team of agents pulled the numbers, flagged what moved, drafted the update, and built the spreadsheets. You edited. You did not build.

• Friday: You closed the week ahead and on time, with every task that was pressing on Monday handled, already moving on to next week.

That week is not aspirational. It is what the fluent operator’s calendar looks like now.

Subscribe now

Can you become AI fluent without taking a course?

You can build a version of each of these on your own. The skill files are documented. The connectors are publicly available. The plugin packages ship with Claude. None of this is proprietary.

What is hard is the sequencing. The order in which you build them, the way you scope the first agent, the prompts that work versus the prompts that look like they should work, and the discipline to keep the source list curated as you go. That is the gap between an operator who tinkers with AI and an operator who runs on it.

I am running a three-week sprint in June called Oper(AI)te that walks operators through these three builds, live, on their own work. Six sessions across three weeks, two a week, 90 minutes each. We connect your stack in week one. We build the research and analytics architecture in week two. We build your agent team in week three. You leave with the connected setup, the research workflow, the dashboard pattern, the agent team, and a 30-day plan to add more as the work evolves.

The sprint begins Tuesday, June 16, 2026. Enrollment closes June 15. Details and the seat link at unfazedfounder.com/operaite.

If the three builds make sense and you want to do them on your own, that is the right answer too.

The cost of waiting another quarter is the part nobody is pricing correctly yet.

Become AI-Fluent This Month

Frequently Asked Questions About Becoming AI Fluent

What is AI fluency?

AI fluency is the ability to work differently because AI is in the loop. It is not knowing what AI can do. It is a change in how work is actually produced. A fluent operator drafts decks in an hour, runs research in minutes, and delegates rhythm work to scheduled agents.

How long does it take to become fluent in AI?

The three builds take three weeks if you sequence them correctly. Build One ships in a weekend. Build two ships in a weekend. Build Three takes the remaining three weeks because the agent team requires the most iteration. The 90-day mark is when the time savings stabilize, and the fluent operator clears 12 to 18 hours of weekly desk work.

Do I need to be technical to build these workflows?

No. The three builds use natural-language instruction inside Claude or another major AI platform. Source curation is a markdown file. Chart-building is plain-English prompts. Agent specs are one-page documents. The technical layer is handled by the platform. The operator's skills are sequencing and prompt discipline.

Which AI platform should I use?

The article uses Claude as the worked example because that is the platform I run on. The same builds work on ChatGPT, Gemini, and Perplexity with minor adjustments to connector availability and skill format. Platform-agnostic fluency is the goal. The Oper(AI)te course covers all four.

What is the elite-research skill?

A custom Claude skill that constrains the model to a defined universe of Tier 1 publications, Tier 2 academic centers, and Tier 3 institutional research shops. It enforces source-tier credibility tags on every finding and outputs in one of three formats: Research Brief, Advisory Intelligence Memo, or Signal Pack. The skill is a Markdown file you install once.

What is the difference between a Claude skill and an agent?

A skill is reusable operating logic that Claude invokes inside a chat or task. An agent is a scheduled or triggered task that runs on its own and reports back. Skills shape how Claude works on a question. Agents do the work without you asking each time. Agents commonly invoke skills as part of their job.

Which agent should I build first?

The Monday Briefing agent. It reads your calendar and your inbox each Monday at 6 a.m., summarizes the week ahead, flags conflicts, and lists the three highest-impact items. It is the easiest spec to write, the highest-frequency value, and the agent that most clearly demonstrates how the rest of the agent stack works.

What if my data lives in tools that do not have a Claude connector?

Two options. The first is CSV export, which works for any system that produces a downloadable report. The second is building a custom MCP connector, which takes about 90 minutes for a competent technical user. Most operators do not need the custom connector path. The core six connectors (Gmail, Calendar, CRM, Notion or Drive, Slack, QuickBooks) cover the work of most senior operators.

How is this different from buying an off-the-shelf AI product?

The three builds compose a general-purpose AI infrastructure around your work. An off-the-shelf product compresses a single function (sales prospecting, transcription, scheduling) into a fixed workflow. The compounding advantage of the three builds is that the same connectors, skills, and agents serve multiple jobs and improve as you refine them. The product replaces a task. The builds replace the way you work.

What does it cost to run these workflows?

Platform subscriptions range from $20 to $200 per month, depending on the AI platform and tier you use. Most senior operators recover the cost of the saved desk time in the first week. The Oper(AI)te course is a one-time $4,500 fee and is structured to make sequencing easier rather than to gate access to the workflows themselves.

Become AI-Fluent This Month

What companion materials accompany this article?

Three resources accompany this piece. Take them in any order.

Field-guide playbooks · 8 pages each · PDF

• Playbook 01 · The Research Architecture - Curate the source universe. Pick the mode before the question. Three output shapes. Deploy and maintain the skill.

• Playbook 02 · Charts and Dashboards - Get the data out of the tool. Describe the question, not the chart. The iteration loop. The dashboard build.

• Playbook 03 · Your Team of Agents - Connectors are the bridge. Write the agent as a one-pager. Set the schedule. Conduct the team.

Elite-research companion brief · Markdown

• Research Brief · The AI Fluency Gap - Six findings, complicating evidence, historical parallels, narrative hooks, reading list.

Which institutions have studied AI productivity and adoption?

Every claim in this article is sourced from the elite-research universe. Tier 1 publications, Tier 2 academic centers, Tier 3 institutional research shops. The appendix below organizes the source material so you can pull threads yourself.

A · Six findings that anchor the article

1. Task-level productivity gains run 14 to 34 percent. Brynjolfsson, Li, and Raymond at Stanford SIEPR studied 5,179 customer support agents using a generative AI assistant. Average productivity rose 14 percent. The lowest-skilled quartile rose 34 percent. The highest-skilled quartile saw minimal lift.

Stanford SIEPR · Generative AI at Work · siepr.stanford.edu Stanford HAI · 2025 AI Index Report · hai.stanford.edu

2. Enterprise-level surveys report no productivity boost. A National Bureau of Economic Research survey of 6,000 executives across the US, UK, Germany, and Australia found 89 percent reporting no productivity impact from AI over the prior three years.

NBER survey · summarized in Fortune · fortune.com

3. Goldman projects 7 percent global GDP lift and 1.5 percentage points of productivity growth. Goldman Sachs Research’s headline AI forecast estimates roughly $7 trillion of additional global output over ten years, with 1.5 percentage points of additional productivity growth.

Goldman Sachs Research · Generative AI could raise global GDP by 7% · goldmansachs.com Goldman Sachs Research · AI is showing very positive signs of boosting GDP · goldmansachs.com

4. Enterprise AI spend is rising 65 percent per year, concentrated in agentic workflows. a16z’s 2025 survey of 100 enterprise CIOs found average LLM spend rising from $4.5M to $7M to a forecast $11.6M, with agentic inference the fastest-growing developer behavior.

a16z · How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025 · a16z.com a16z · The Rise of Computer Use and Agentic Coworkers · a16z.com

5. 91 percent of executives expect agentic transformation. 13 percent have shipped it. Harvard Business Review’s 2026 agentic enterprise survey reports the gap between executive conviction and organizational readiness.

HBR · What Leadership Looks Like in an Agentic AI World · library.hbs.edu MIT Sloan Management Review · The Emerging Agentic Enterprise · sloanreview.mit.edu

6. Sequoia argues the next $1 trillion company will sell work, not software. Konstantine Buhler’s Agent Economy framework defines the three primitives of agent infrastructure: persistent identity, seamless communication protocols, and security and trust. Sequoia’s Services: The New Software essay extends the argument to a trillion-dollar profit pool.

Sequoia · Services: The New Software · quasa.io Sequoia · The Agent Economy · inferencebysequoia.substack.com

B · Complicating evidence

Three findings push back on the simple version of the story. They belong in any honest read of the institutional record.

METR finds AI tooling slows senior engineers by 20 percent on complex repositories. Model Evaluation and Threat Research’s 2025 study of working developers found AI tools added 20 percent to task time on average, inverting the Stanford pattern. The lift is not uniform across roles and complexity levels.

Thomson Reuters finds modest productivity gains for lawyers, accountants, and auditors. The 2025 report found targeted use of AI for document review and routine analysis produced gains that fell short of signaling widespread displacement.

HBR warns against treating agents as employees. A May 2026 HBR paper argues that the agent-as-employee analogy hides three things human workers bring that agents do not: stable context across the day, an instinct to escalate when something feels off, and accountability that survives a bad outcome. The recommendation is to run agent rollouts through systems-engineering templates with scoped permissions, observability, kill switches, and a clear owner per action.

C · Historical parallel

The current gap between task-level AI productivity studies and enterprise-level surveys looks remarkably like the Solow paradox of the late 1980s. Robert Solow, writing in the New York Times in 1987, observed that “you can see the computer age everywhere but in the productivity statistics.” The lag between visible technology and visible productivity stretched roughly a decade before the late-1990s acceleration. The operators who built the workflow during that lag captured the gains when the macro numbers finally turned.

D · Reading list

A starting point if you want to go deeper. Eight pieces, ordered by relevance to the three builds.

1. Stanford SIEPR · Generative AI at Work. The canonical task-level productivity study. Begin here.

2. Stanford HAI · 2025 AI Index Report. Comprehensive annual review. The productivity chapter is the relevant section.

3. a16z · How 100 Enterprise CIOs Are Building and Buying Gen AI in 2025. The cleanest read on enterprise spend and behavior.

4. a16z · The Rise of Computer Use and Agentic Coworkers. The agent thesis from the firm with the largest agentic portfolio.

5. HBR · What Leadership Looks Like in an Agentic AI World. The clearest executive-level framing of the adoption gap.

6. Sequoia · Services: The New Software. The trillion-dollar thesis.

7. Goldman Sachs Research · Generative AI could raise global GDP by 7%. The macro forecast.

8. MIT Sloan Management Review · The Emerging Agentic Enterprise. Four organizational tensions in agent adoption.

E · Source architecture note

The elite-research skill described in Build One is what produced this appendix. Every cited source sits inside the curated universe of Tier 1 publications, Tier 2 academic centers, and Tier 3 institutional research shops. The output format is a Research Brief. The full brief that anchors this appendix is published as a companion document.

About the author

John Brewton documents the history and future of operating companies at Operating by John Brewton. He is a graduate of Harvard University and began his career as a PhD student in economics at the University of Chicago. After selling his family’s B2B industrial distribution company in 2021, he has been helping business owners, founders, and investors optimize their operations ever since. He is the founder of 6A East Partners, a research and advisory firm asking the question: What is the future of companies