As a CEO, where should we start with AI, and what are the highest-value use cases?

Most CEOs start in the most expensive place: a top-down strategy, an AI committee, a chatbot bolted onto the website. The highest-value move is smaller, closer to home, and you can start it this week. Here is where to begin and the use cases that actually pay.

You already pay for ChatGPT or Claude. You have used it a handful of times, it was fine, and some quiet part of you suspects you are missing something the people who keep talking about this are not. You are. But the thing you are missing is not a tool. It is a starting point.

The question in the title is the right one, and almost everyone answers it wrong. Where you start with AI matters more than which model you pick, and the default instinct, the one that feels responsible and strategic, is the one that wastes the most time and money. Let me tell you where not to start, then where to actually start, then exactly which use cases are worth your first month.

The most expensive way to start, and almost everyone picks it

The responsible-looking move is to treat AI as an IT project. Convene a committee. Write an AI strategy. Stand up a custom build or hire an "AI team." Put a chatbot on the website so the company has something to point at. It feels like leadership. It mostly produces a slide deck.

The data on this is brutal. MIT's 2025 study of enterprise AI found that 95% of corporate generative-AI pilots delivered no measurable impact on the P&L. The cause was not weak models. The models are extraordinary. The cause was that companies pointed them at the wrong place and ran them through the wrong process. The same research found that more than half of AI budgets went into sales and marketing tools while the biggest returns were sitting in unglamorous back-office work, and that internally-built custom AI projects succeeded about a third of the time, while buying a focused tool and actually using it worked about twice as often.

Read that twice, because it inverts the usual playbook. The expensive, custom, top-down, strategy-first approach is the one that fails. The thing that fails is starting with the technology instead of starting with a job you personally do every week.

I watched a version of this play out in a peer group last year. A CEO proudly demoed the AI assistant his team had spent a quarter building. The first thing it did on stage was schedule a meeting to talk about having fewer meetings. Everyone laughed. Nobody changed what they were doing. That is what starting in the wrong place looks like: motion, theater, a thing to point at, and no leverage you can feel on a Tuesday.

So before any of the use cases, the first principle: do not delegate your own AI education to a committee or a vendor. The leverage starts with you.

Start with yourself, on a task you already do every week

Here is the actual answer, and it is almost annoyingly simple. Start by using AI yourself, daily, on high-frequency tasks where you can check the output, before you mandate anything for anyone.

Ethan Mollick, the Wharton professor who has done more than anyone to make this practical for normal operators, has three lines worth taping to your monitor. First: pay for the good version, because "the free versions are demos, not tools." Spend the twenty dollars a month on a frontier model and use the most capable one for real work. Second: "always invite AI to the table," meaning try it on basically every task you do, because you cannot know where it is brilliant and where it is useless inside your specific job until you have thrown real work at it. Third: "assume this is the worst AI you will ever use," because it is, so the habits you build now only get more valuable.

The reason to start on tasks you can verify is the one fact that governs everything about these tools. An LLM, the engine inside ChatGPT and Claude, predicts plausible text; it does not check whether the text is true. It will hand you a confident wrong answer in the same voice as a confident right one. That sounds like a dealbreaker. It is not. It just tells you exactly where to point it: at work where you are the judge of whether the output is good. You have spent your whole career judging whether a draft, a number, or a plan is right. That judgment is the one skill the model does not have, which is why you stay in the loop, and why your own desk is the safest, highest-return place to begin.

So week one is not a strategy. It is a habit. Pick the single weekly task that annoys you most and run it through AI five times this week. Watch where it saves you twenty minutes and where it wastes five. That is your map. Everything below is just the highest-value territory on it.

The three use cases that actually pay

For an operator, the value concentrates. Most of the returns in your first quarter come from three places. They share a shape: high-frequency, context-rich, and easy for you to verify.

1. Turn your meetings into leverage (the fastest ROI there is)

If I could install only one thing for a CEO, it would be this one, and it is the one almost nobody starts with.

You sit in meetings all day. Each one generates decisions, commitments, customer truths, and follow-ups, and most of that evaporates the moment the call ends. You half-remember, you scribble a note you never reread, you forget who owes whom what by Thursday. The single highest-ROI AI move for an executive is to stop losing that. An AI notetaker like Granola sits in your meetings, transcribes them, and turns each one into clean notes. That alone is worth it. The market agrees: Granola raised at a 1.5 billion dollar valuation in early 2026 precisely because capturing meetings turned out to be the wedge into an executive's entire working memory.

But the capture is just the start. Once your meetings land as plain text files in a folder on your laptop, you can point AI at the whole history and ask it to do real work. A weekly review that reads the last seven days of meetings and writes you a one-page summary of what shipped, what slipped, and who is waiting on you. A pre-meeting brief that reads everything you ever discussed with a person before you walk in. Follow-up drafts in your voice, sitting in your outbox before you have left the room. Andreessen Horowitz, watching where AI productivity actually shows up, points to exactly this category, the meeting notes and follow-ups and reminders that pile up on even senior people, as the natural place AI earns trust first.

This is the install I would do before anything else, and it is why we wrote a whole pillar on it. It touches every hour of your week, the output is trivially easy to verify (you were in the meeting), and it compounds: the longer it runs, the more your meeting history becomes a thing you can ask questions of.

2. Build the small tools you would never have paid to build

Here is a category that did not exist for a non-technical CEO eighteen months ago and now does. You can have AI build you working software by describing it.

Not a product. A small internal tool. A calculator your sales team needs. A page that turns a messy export into a clean one. A dashboard for the one number you check every morning. For twenty years, getting a piece of custom software meant a budget line, a hire, or a six-week agency quote, so these little tools never got built. Now you describe what you want to a coding agent like Claude Code, and it writes the code while you watch. The practice even has a name now, vibe coding, coined by Andrej Karpathy in early 2025, and the striking part is that most of the people doing it cannot write the code themselves. They do not need to. They describe the outcome and judge the result.

I walk through the whole thing, from never having opened a terminal to a working tool you can open in your browser, in build your first software tool. The example there is a deal-margin calculator a sales rep can open mid-call, the kind of thing that was never worth a budget line and takes forty minutes to build by talking. The one rule that keeps this safe: vibe-code the internal, disposable, verifiable stuff freely, and the moment a tool touches customer data, payments, or the open internet, bring in someone who can read the code. Inside that boundary, you have just removed the budget meeting from the path between "I wish we had a tool that..." and having it.

3. Move your sales motion from calls to follow-through

The third place the value concentrates is the gap between a good sales conversation and what actually happens after it.

Your reps have great calls and then drop the thread. The real objections, the ones costing you deals, are sitting uncaptured in your call recordings while your landing page answers objections someone imagined a year ago. AI closes that gap in three concrete ways. It can turn a customer call into a follow-up draft in your rep's voice before they have left the room. It can keep your CRM honest, enriching records and logging the activity that reps never log, so the pipeline data you make decisions on is actually real. And it can read every deal's call transcripts and surface what is slipping, so your pipeline review is grounded in what was actually said instead of what a rep optimistically typed into a stage field.

This is verifiable work on high-frequency events, which is why it pays. You can read the follow-up before it sends. You can spot-check the enriched record. The AI supplies the discipline your team never quite maintains; you supply the judgment.

A quick word on the rest. Finance and board prep pay too: AI is genuinely good at working through a spreadsheet with you and at turning a quarter of mess into a board or investor update. Hiring, content, support triage, all real. But if you try to start everywhere, you start nowhere. Pick from the three above, in that order, and let the wins fund your appetite for the next one.

What is overrated, so you don't burn the first quarter

Knowing where not to spend is half of starting well.

The website customer-service chatbot is the most over-started and worst-returning AI project for most companies. Research on AI customer service has found it fails at several times the rate of other AI use cases, and only a tiny fraction of customers actually want to be stuck talking to a bot. It is visible, which is why executives love it, and it is exactly the place where a confident wrong answer reaches a customer with no one in the loop to catch it.

Replacing headcount on day one is the second trap. The companies that loudly cut service teams to "AI-first" have quietly rehired when quality cratered on anything non-routine. AI is leverage on your existing people long before it is a replacement for them, and treating it as a layoff plan poisons the adoption you actually need.

And the big custom data platform, the one the committee wants to build, is the third. Remember the MIT number: buying a focused tool and using it beats building your own roughly two to one. Start by buying and using. Earn the right to build custom later, if ever.

The real ceiling: stop chatting, start delegating

Everything so far lives mostly in a chat tab, and a chat tab is where you should start. But it is not where the highest value is, and I would be doing you a disservice not to point at the ceiling.

The chat window has you in the loop for every step. You ask, it suggests, you do the work. The leverage layer is the move from that to agents that do the work. You describe an outcome, and the AI reads your files, runs the commands, drafts the output, and reports back while you do something else. The cleanest way to feel this today is Claude Code in the terminal, Anthropic's agent that runs in a folder on your laptop, reads the context of your business, and acts instead of just advising. It is the same shift Andreessen Horowitz describes as trust escalation: once AI handles the small steps reliably, you let it run the whole workflow before it taps you on the shoulder.

That is the difference between a CEO whose AI resets to zero every Monday and one whose setup knows the business and gets sharper every week. The orchestration layer that wires all of this together, the connectors, the memory, the scheduled jobs, is called a harness, and it is where this goes once the basics are paying off. You do not need it in month one. You need to know it is the direction, so the foundations you lay now point the right way.

If you want the full strategic case for working this way, in a folder instead of a tab, it has its own pillar. For today, the only thing to take from this section is that chat is the on-ramp, not the destination.

The order to actually do it in

Here is the sequence I would hand a CEO who asked me this question over coffee.

1. This week: pay for one frontier model and use it daily on the single weekly task that annoys you most. Build the habit before you build anything else.
2. Next: install meeting capture and run one weekly review off it. This is your fastest, most verifiable win and it touches every part of your week.
3. Then: build one small internal tool by describing it. Feel what it is like to produce working software without a budget meeting.
4. Then: wire one sales follow-up workflow so the gap between a call and the follow-through closes itself.
5. Then, when the basics are paying off: move into the terminal and let agents do the work instead of just suggesting it.

Notice what is not on the list: a strategy offsite, an AI committee, a custom build, a chatbot. Those come later, if at all, and they come easily once you personally understand the tool. Start with one task, on your own desk, this week.

Do this now

Pick the one weekly task you most resent doing, the recap you rewrite, the update you dread, the follow-ups you let slide, and run it through AI before Friday. Verify the output yourself. That single loop, repeated, is the entire on-ramp; everything in this piece is just the highest-value places to point it next. And once you know where to start, here are ten concrete things you can hand AI this week, ordered from the cheapest to begin to the highest ceiling.

When you are ready for the fastest install, the meeting-capture pillar is the place to go. When you want the strategic case for the leverage layer, the terminal pillar is waiting. And if you would rather start with a guided walk from never-touched-a-terminal to running real workflows on your laptop, the free sample chapter of the CEO guide to Claude Code does exactly that.

Tell me in thirty days which task you started with and what it gave back. Those are my favorite messages to get.

Andrew

Related reading

• Granola for CEOs: the highest-ROI AI install of 2026 · the fastest first win, in depth
• Why CEOs should use Claude Code in the terminal · the leverage layer above the chat tab
• What is a large language model (LLM)? · the engine inside every chat tool, and its one limit
• Build your first software tool in Claude Code · the tool you would never have paid to build
• What is vibe coding? · describing software and letting AI write it
• Claude Code tips for non-technical CEOs · the operating habits that turn it into leverage