Operations & Systems

    Just Because You Can Automate It Doesn't Mean It Moves the Needle

    In a world where you can automate almost anything, here is the contrary advice: you probably should not. Just because you can do something with AI does not mean it ever would have made the list of things worth doing. That is the easiest procrastination ever invented.

    Tanner O'BrienSeptember 17, 202614 min read
    Just Because You Can Automate It Doesn't Mean It Moves the Needle

    In a world where you can automate just about anything, here is the contrary advice: you probably should not.

    That is an odd way to open an article about AI, and we mean it. Over the last year and a half we have built our own stack, wired agents into our daily operations, and pointed these tools at real constraints in our business. Along the way we learned something that matters more than any tool on the list.

    Just because you can do something with AI does not mean it ever would have made the list of things worth doing.

    If a task would not have moved the needle before these tools existed, automating it does not make it valuable. It just makes it faster to do something that never needed doing. That is one of the easiest ways to procrastinate that has ever been invented, and it feels like progress the entire time you are doing it.

    So before any of the practical material, start here. Ask whether the thing you are about to automate would make you money or free up time that you would then use to make money. If the answer is no, skip it.

    Where AI actually belongs in building a business

    We build businesses in six steps. Mastery first, then marketing, then systems, then team, then scale, then freedom. It is a pyramid, and each layer sits on the one beneath it.

    Mastery is the foundation and it has four parts. Destination, meaning where are we headed. Delivery, meaning how consistently do we produce the product or service we sell. Time, meaning how we use that resource. And financials, meaning whether we actually understand the numbers. Once that foundation is solid, you layer marketing and sales on top, which produces predictable cash flow. Then you build systems. Then you hire the team that runs the systems. Then you scale.

    Here is the part worth writing down. AI is technology, and technology sits squarely in the systems step. Step three.

    Which means if mastery and marketing are not handled, trying to systemize your way into AI is almost certainly not the best use of your time. You cannot automate your way out of not knowing where the business is headed. You cannot prompt your way past inconsistent delivery. You will just get to the wrong destination faster and with better documentation.

    Within systems there are four places you can apply leverage: systems and technology, people and education, delivery and distribution, and testing and measuring. AI lives in the first one. It is a lever you can point at any part of the business. The skill is choosing where to point it.

    The stack, in three buckets

    You do not need every tool that exists. You need one or two in each of three categories.

    Think

    These are the chat tools, used as a thought partner. We use ChatGPT, Claude, and Gemini nearly every day. There is nothing wrong with Copilot, Grok, or Perplexity. The reason we use three is that we will run the same question through different models to see the problem from more than one angle. Gemini earns its spot for a different reason, which is that it is connected to Gmail, Drive, and Calendar, so it can search everything already living in the Google workspace and handle deep research well.

    Capture

    This is the bucket most owners skip, and it quietly makes the other two work. Capture means getting information out of your head or out of a conversation and onto a page. We use Granola and Whisperflow, plus Zoom for the meetings themselves. Otter, Fireflies, and Fathom all do a version of this, and devices like Plaud or Pocket will capture conversations happening away from your computer.

    We prefer Granola for one specific reason. It lets you type your own notes inside the platform while the meeting runs, so the AI summary is built from three inputs, the transcript, the recording, and what you actually thought was important. Most tools only give you the first one. It also records offline from a phone, which means a coffee meeting gets captured the same way a Zoom call does.

    Whisperflow replaces typing. Hit a key in any text field, talk, and it cleans up the brain dump and drops the polished version in. Far faster than typing, and it has a second benefit we will come back to.

    Act

    These are the tools that do work rather than talk about it. Claude Code and Codex build things, and you do not need to know how to write code to direct them. Claude Cowork and ChatGPT Work sit between the chat tools and the coding tools, working with the files on your computer and using your browser to take actions on your behalf. And then there are agents, which is a fast-moving category. Right now we use OpenClaw, Hermes, and more recently GrokBot. These operate more like an employee than a tool. They run around the clock and report back.

    The two inputs that decide whether any of it works

    Context and data. That is the whole thing.

    A year ago the conversation was about prompts and whether yours was written well enough. That matters far less now if you have given the tool enough context. Context is who you are, what you run, how you make decisions, and how you want to be spoken to. You write it once and put it in the settings of whichever tool you use, and it gets read on every single request.

    Think about it the way you would think about a new hire. You bring someone in and say go, with no training, no history, no explanation of how you run meetings or where the company is headed. They have general experience to draw on, so they will do something. It just will not be the thing you wanted. The tool is in exactly that position every time you open a blank chat without context.

    Here is the shape of one that works, which you are welcome to take and adapt.

    Start with identity. Who you are, what you do, where you are based, what time zone you work in, and any formatting preferences like using absolute dates. Then what you run, meaning the actual structure of the business, how you grow, and how your team is built. Then how you decide, which for us means profit and cash, client acquisition cost and payback period, time leverage, reversibility, and first principles over conventional wisdom. Then how you want it to talk to you. We ask for blunt, data first, lead with the answer rather than a wind-up, bullets over paragraphs, and risks flagged red, yellow, or green.

    Then, and this is the part most people leave out, tell it to challenge you. Ours says to name what we are missing, and if we are wrong to say so first and then fix it. Always land on a recommendation rather than hedging into it depends.

    Finally, list what it should never do. Never describe a deliverable instead of building it. Never hedge a recommendation into uselessness. Add whatever else you find yourself correcting over and over.

    If writing that from scratch feels like work, hand this framework to your AI tool and ask it to interview you until it can draft your version. Then paste the result into your settings.

    Data is the other half. Without it you get generic answers. Ask how do I improve retention with no data attached and you will get a list of things that worked for somebody else, which may or may not apply to you. Give it your numbers and the question changes entirely. Our average client tenure is 6.2 months across these 40 clients, it should be closer to 18, here is the data, what are the top three causes and what are we missing. That is a question worth asking, and it only exists if somebody has been capturing the data.

    Without data you are running on gut feel. Sometimes gut feel hits. More often it does not.

    The instruction that keeps you in the driver's seat

    One line changes how these tools behave more than any other:

    Before you answer, ask me the questions you need answered to do this well.

    That reversal does two things. It feeds the tool more context, so the output gets better. And it keeps the thinking on your side of the table, which matters more than it sounds like it does.

    These tools will work tirelessly, which is genuinely useful. They also have no stakes. They do not get fired. If they are wrong, they will apologize when asked and nothing about that changes anything. You are still the one who has to decide whether the answer holds up, how it got there, and whether it applies to your situation. Asking it to ask you keeps one hand on the wheel.

    Where we actually point it

    Three examples from the mastery layer, since that is where most owners need help first.

    Time

    We used to take notes in every meeting, which meant dividing attention between the person across from us and the page. Now a capture tool handles the record and we write down a few things that stand out. The summary comes back in whatever format is actually useful to the reader, longer and detailed for some people, bullets for others. From those notes it can pre-draft the follow-ups. It can triage an inbox too, so that after a week or so of correction you open your email and see only what needs you, with the rest already sorted.

    Delivery

    In a coaching business, consistency of preparation is the product. Every client now gets the same quality of prep, because an agent sends a brief before the session with the financial snapshot, the open loops from the last two meetings, and the patterns worth calling out. Ours flags things like a hiring decision that has been discussed repeatedly with no action taken. It will also tell us when it has no notes for a client past a certain date, which is a useful nudge that something did not get uploaded.

    Money

    These tools read numbers and find what you missed better than anything available at any previous point. Getting good at spreadsheets used to be a real competitive skill. Today the tool handles the extraction and structure, which means the time goes into the insight instead of the assembly.

    Two more places it earns its keep

    Content

    The wrong way is to open a chat, ask for a Facebook post on a topic, and publish what comes out. Everyone is doing that and it is producing exactly what you would expect.

    The better way runs on everything you have already captured. Point the tool at your notes, your session summaries, the recordings where you sat down and talked through an idea out loud. Ask it to pull one or two things that are timely, draft something, and then go back and forth until it is right. That iteration is a creative process and it still needs your taste. Once it is right, tell the tool that this is the output you want, and have it build a skill or rule set from that, so next time the first draft lands closer.

    One test for anything you are about to publish. If you would not say it out loud from a stage, do not post it. If you read it back and the words feel awkward coming out of your mouth, they are not your words. Change them until they are.

    Call reviews

    This is the one we would push hardest, because it turns a soft problem into a measurable one.

    Sales call reviews are not a new idea. Every organization that has ever produced great salespeople listens to calls and coaches reps on what they hear. The problem has always been that it takes enormous time and humans miss things.

    We built a simple setup, effectively a folder where every call transcript goes, and the tool grades each one against our script, measures rep talk time versus prospect talk time, and identifies why the call moved forward or did not.

    The first roll-up covered 44 calls, and one number stopped us. Across 13 value calls, zero of them quantified the cost of inaction. Not a few. None. In our business, understanding the investment required to move forward matters, and understanding the cost of not moving is just as important, because standing still has a price that most people never put a number on. We were not doing it at all, and we had no idea.

    Other things surfaced. Whether the rep actually asked for the money at the end. Average talk time, because if you are talking 80 percent of the call you have made the sale about you rather than the person in front of you.

    Then we did the human part. We took what we found back to the team, talked it through, and role-played it. Script compliance on the prior ten calls sat at 50 to 57 percent with 40 percent of next meetings booked live. Across the most recent stretch, compliance moved to 70 and then 74 and 75 percent, with booking live climbing to 60 percent and up.

    The tool did not fix anything. It showed us a constraint we could not see, and then people fixed it. That is the correct division of labor and it is worth being precise about it.

    A few practical answers

    Which tool should I use?

    Honestly it does not matter much. Pick one, get comfortable, and give it context. We lean on ChatGPT and Claude. We would not bet against Google or against Grok either, given the distribution behind both.

    Is it safe to put client information in?

    Consult your own counsel, and know that policies differ by tool and by account type. Team and business accounts generally do not train the model and retain less. Personal accounts typically retain longer, and there is usually a toggle in settings controlling whether your data trains the model, which is often on by default. Go turn it off. Be cautious with personal information regardless.

    What does it cost?

    For most owners, the paid plan at roughly 20 to 30 dollars a month per tool is the right call rather than the free tier. One tool in each of the three buckets lands you somewhere around 90 to 100 dollars a month total. The 100 and 200 dollar plans are only worth it if you are consistently hitting usage limits on one platform.

    Can I trust the numbers it gives me?

    Most of the time, but do not take it at face value. Put a rule in your instructions telling it to challenge assumptions, fact check itself, and cite sources. Then spot check. This is not that different from a team member. Someone can be excellent and still source something badly. Check the work.

    How do I make the output sound less like AI?

    Give it your voice. The more examples it has of how you actually talk, the closer it gets. This is the second reason we like a tool like Whisperflow, because talking into your work generates a record of your natural cadence. Then name the tells you want banned. Em-dashes are the famous one. Add repetition, double negatives, and choppy repetitive sentences. Build the rule set once and reuse it.

    One thing to do this week

    Pick one capture tool you are not using today. Granola, Fireflies, Fathom, Otter, or a voice tool like Whisperflow. Any of them.

    Start using it by Monday.

    Capture is the bucket that feeds the other two. Without it, your thinking tools have nothing but general knowledge to work with and your acting tools have no instructions worth following. It is also the easiest of the three to start, which is exactly why it is the right first move.

    And if what you took from this is that the constraint in your business is not the tooling, that is worth acting on too. Most of the leverage available to an owner is still in the fundamentals, and the fastest way to find out which fundamental is holding you back is to have somebody outside the business look at it with you.