Here’s where I think all this AI stuff is heading: imagine a company that gets a little smarter every day, on its own. Every meeting, every sales call, every decision, every number flows into one place, and a layer of intelligence sits on top of all of it, always reading. Monday morning, before you’ve poured your coffee, it has already flagged the client who’s quietly drifting, the offer converting better than anyone noticed, and the process costing more than it should. Nobody had to go digging. The noticing is just done.
Walk through a company like that and the people aren’t buried in the work that used to eat their weeks. The research, the first drafts, the follow-up, the endless piecing-together of who said what across a dozen threads, the layer carries all of it. What’s left for the people is the judgment, the relationships, the creative leaps, the real thinking. And because the layer never stops learning from everything the company does, it compounds. The business improves itself a little more with every conversation it has.
That isn’t science fiction, and it isn’t ten years out either. Parts of it are buildable right now. The full version, where the whole business quietly improves itself, is the direction to build toward, not a place any company has fully arrived, ours included. But it’s where I think everyone should be heading.
The honest part, and the reason I’m writing this instead of just selling you the dream, is that you don’t get there by buying a platform and flipping a switch. You get there one deliberate step at a time, and it’s easy to start in the wrong place.
At this point, I’d say most mid-market owners aren’t afraid of AI. They’re afraid of falling behind while they figure it out, and of spending real money on tools nobody ends up using. That second fear is the well-founded one. The tools are sitting there, mostly unused, in companies that already pay for them.
So the real question isn’t whether AI matters. It’s where to start, and what “integrated” even means for a company your size, which is bigger than a startup that can rebuild itself overnight and smaller than an enterprise with a data-science department. This is a guide from the operator’s chair. We integrated AI into our own business first, made most of the mistakes below ourselves, and only then started advising other companies on it. Everything here is what actually worked.
The short version
Start with literacy, not systems. Teach your team to use AI as a thinking partner, find your first problems by watching what eats their time, build small skills instead of autonomous agents, and keep everything read-only until you trust it. Let the early wins compound into an operating layer that runs through the whole business.
That progression matters more than any single tool, and it’s easy to invert it. A lot of companies buy a platform, announce a rollout, and wait for transformation that never arrives, because they skipped the part where people actually learn to use the thing. Literacy first, then small useful skills, then a layer that runs through the business, then a company that starts to optimize itself. The rest of this piece is that path in detail, including the honest parts about cost and what it really takes.
Fluency before integration
The trap isn’t picking the wrong tool. It’s that the tools already exist and nobody uses them well. We watched a private-equity operator describe his own firm, which had an enterprise AI license and a bespoke tool trained on tens of thousands of internal files, and admit that nearly all of the real usage was basic cleanup and email. The capability was paid for. The adoption wasn’t there.
That gap is a literacy problem, not a technology problem. Before you integrate AI into anything, the people in your business have to understand what it is and what it’s for. And the single most useful reframe is this: AI is a thinking partner, not a search engine and not a magic button. You can use it the way you’d use a sharp colleague, not a vending machine where you put in a request and expect a finished can to drop out. The people who get value talk to it, push back on it, and iterate. The people who don’t treat it like Google and walk away disappointed.
Getting AI into a company is an org-wide, person-by-person effort, and everyone comes on board eventually. But if you’re the one driving it, start with your directors and managers, because they have the judgment to tell good output from bad, and that judgment is the whole game. They also become the champions who make it real for the people under them. Someone less experienced can’t always tell when AI is wrong. An experienced manager can, and their standards rub off on everything they build with it.
Finding your first problems doesn’t take a strategy offsite, and it isn’t only the boss’s job. It’s the same mindset up and down the company: do your normal work, notice the parts that are tedious or repetitive or eat your whole afternoon, and ask whether AI could help with that specific thing. Three questions surface it fast, whether you’re asking them of your team or of yourself. What do you do all day? What part sucks? Show me. The tedious, repetitive, judgment-light work is exactly where AI earns its first wins, and those early wins buy the credibility for the harder integrations later.
One more sequencing rule that saves money and grief: skills before agents. Build small, narrow tools that each do one useful thing, instead of reaching for an autonomous agent that runs the whole job on its own. The small tool helps tomorrow. The autonomous one is a bigger build and a bigger risk, and you work up to it.
One thing nobody warns you about: the first version of anything you build with AI will be maybe sixty percent of the way there, and that is exactly how it is supposed to go. You don’t one-shot a good tool. You build a rough one, watch where it falls short, and then tell it what you were actually thinking. Why this case needs more nuance. Why that judgment call went the wrong way. Why the tone was off. You iterate it into existence over a few rounds. What you’re really doing in those rounds is encoding your own judgment, because not everything in a business is mechanical, and the parts that aren’t are the parts worth capturing. The mechanical work is the easy part. Teaching the tool to weigh a problem the way your best person would is the real work, and it pays off, as long as you keep that judgment current when your tools, your models, or your standards change.
The mistakes that waste the money
If literacy is where the value starts, a handful of predictable mistakes are where the money leaks out.
The first is expecting the vending machine. People type a vague request, get a mediocre answer, and conclude AI doesn’t work for their business. The fix is the reframe above, and it’s worth over-investing in, because it’s the difference between a team that quietly abandons the tools and one that gets better every week.
The second is automating too early. There’s a strong pull to jump straight to full automation, to wire AI directly into the systems that run the company. Resist it. The discipline we hold to, in our own shop and every client’s, is spreadsheet first, automation later. Prove the logic works in something cheap and visible before you build the automated version. Premature automation bakes in mistakes at scale and is brutally expensive to unwind.
The third is buying point solutions that die on the vine. The market is full of narrow AI features bolted onto software you already own, and many of them get used twice and forgotten. The value isn’t in any single tool. It’s in your people getting fluent enough to solve their own problems with whatever tool fits, which is why we tell clients to buy the capability, not the widget.
The most dangerous one is giving AI write access before you trust it, the uglier cousin of automating too early. Letting a system act on your data, your customers, or your money before you’ve watched it work for a while is how a small error becomes a real one. Keep it read-only until it has earned more.
What actually becomes possible
When literacy is real and you’ve built a few useful skills, the character of the work changes. Things that were too expensive, too slow, or simply out of reach start to become ordinary, and it almost always begins the same way: someone asks how AI could help with one specific, real problem, and then actually builds the thing.
Some of what becomes possible is disposable, and that took us a while to appreciate. You can spin up a small piece of software for a single problem, use it, and throw it away. I built a little budget-allocation app once, just to show how a marketing budget could flow across every channel at once, used it to make the point in a single meeting, and never opened it again. Building a tool used to be a project you protected for years. When the cost of building it collapses, it can be worth making even if you use it a single time. And some of it happens the same day you think of it, the way a plugin we built in an afternoon now writes structured data across our whole website.
Underneath these one-off wins is the thing that actually matters. Strung together and run as a daily habit, capabilities like these stop being separate tricks and start becoming the operating layer, the early and real version of the company I described at the top. That’s the difference between AI as a productivity hack and AI as infrastructure. One is a tool you go use when you remember to. The other is a layer the business quietly runs on.
What this looked like for us
Let me take one example all the way through, because a principle you can’t picture isn’t worth much.
Our buyer personas were about four years old. They were directionally right when we wrote them, but the company had changed and they hadn’t, and because those personas feed the system we use to write our own marketing, the drift showed up everywhere downstream. The copy felt a half-step off from the people we were actually talking to, and we could feel it before we could name it.
So we asked the question from the last section. How could AI help us hear what our buyers actually say, instead of what we assumed they said? We had years of our own sales calls and client conversations sitting in a pile nobody had ever mined. We built something to read through all of it and pull out the real language, the exact words people used for their problems, their goals, and their objections, and to organize it into evidence we could write from, and rebuild our personas on, this time grounded in what buyers actually said instead of what we’d assumed.
The first version was wrong, and how it was wrong is the useful part. It over-weighted whichever customers happened to have the most recorded conversations, so two talkative accounts could masquerade as a company-wide pattern. When we caught it, we rebuilt the scoring around one rule: count distinct customers, not conversations. A theme only earns its way into our marketing once several different customers have said it in their own words. The day we applied that honestly, a batch of our favorite in-house phrases fell apart, language we had coined that no actual customer had ever used, and we cut it. That alone was worth the build.
We never let the tool decide what was true. It proposes, a person decides. A finding has to clear the customer-count bar to become a candidate, and then it gets a human gut-check, because the model works only from the text and can’t see what I know about these people from years across the table from them. That review is there to catch the misreads. It once read a buyer’s self-deprecating joke as literal, when anyone who knew him would have heard it completely differently, and it’s also where I have to be careful not to just reconfirm my own assumptions. Our salespeople check it too, against what they actually hear on calls. It has a built-in blind spot: it only hears the buyers we already talk to, not the ones we never reached, so it sharpens who we serve more than it reveals who we’re missing. We treat it as one strong input, not gospel.
And it worked. It surfaced a whole kind of buyer we had been under-serving in our messaging, a revenue-and-sales leader who shows up in our deals over and over, confirmed across enough separate companies to be real rather than a hunch we’d half-sensed for years. Once the evidence was there, we rebuilt that part of our marketing around him. It was just as useful when it told us no, killing an expansion I had personally hoped for because the evidence wasn’t there.
It also corrected the buyer we thought we already knew. Our picture of our main marketing contact had quietly drifted upward over the years, more senior, more of a department head, than the people actually hiring us, who were usually running the whole function themselves with little or no team under them. We had even been picturing our ideal customer as a bigger company than the ones who actually became our best clients. Writing to the real person instead of the flattering version is what finally closed the half-step gap we’d been feeling.
The best part is that it keeps getting sharper without anyone running a project. The loop it runs now is simple enough to lay out:
- A meeting happens, and my daily routine picks up the transcript on its own.
- It checks one thing first: was this a customer conversation? If not, it stays out of this system entirely.
- If it was, a small skill reads the transcript and pulls out the signal, the exact words people used for their problems, their goals, the questions they asked, and their objections.
- That extraction gets written into its own note, one per conversation, with the quotes kept word for word.
- From there the signal is sorted into a handful of living documents, each with a job: the language we should be using, the questions buyers actually ask, the objections we have to answer, and the personas themselves.
- Everything is counted by distinct customers, not by conversations, and once enough different customers have said the same thing, it’s solid enough to build on.
All of that writing happens inside the system’s own notes, never out to our CRM or anywhere near a customer, and nothing becomes something we actually act on until it has cleared the bar and I’ve looked at it. Because that runs on every meeting, the personas are far less likely to quietly drift out of date, as long as we keep an eye on the system itself. And because we built it for ourselves first and it actually worked, it’s becoming something we now set up for clients.
Sharper personas were never the real goal, though. The payoff is downstream. When your marketing and sales material speaks in the actual language of the people you’re trying to reach, it resonates more, and more resonance is a lever on the number that actually matters, top-line revenue. I won’t pretend to draw a clean straight line from a persona doc to a closed deal. But the mechanism is real: talk to the real buyer in their own words, and more of them lean in. There’s a second bet riding along with it. As AI answer engines increasingly shape who gets found and recommended, the clearer and more honest your material is about who you actually serve, the better the mental model those systems build of you, and we would much rather they understand us accurately than guess.
Start to finish, getting it into good shape took somewhere around thirty hours, most of that the unglamorous part: feeding it conversations, watching where it got things wrong, and correcting it.
And this is just one loop. It happens to point at our personas, but the same shape, watch a real process, encode the judgment, let it run and keep improving, works just as well aimed at how you price, how you scope and deliver work, or where your marketing is quietly leaking money. Run one and you have a useful tool. Run a hundred of them, each sharpening its own corner of the business, and you are most of the way to the company I described at the top.
Keep a human in the loop
None of this works as replacement. It works as augmentation, and the distinction is the whole design.
The rule we run everywhere is simple. AI proposes, a human approves. The machine does the volume, the drafting, the first pass, the synthesis, and a person with judgment makes the call. In our own operation that approval step is literally built into how changes ship, so nothing goes live without a human signing off. The judgment layer isn’t a safety afterthought. It’s the point. AI is spectacular at producing a lot of pretty-good, and worthless at knowing which pretty-good is actually right for your business. That knowing is what you and your best people bring, and it’s what you’re freeing up when you let AI carry the volume.
Approval is also where the real risks get managed, and they deserve more than a shrug. Starting a tool read-only takes the most dangerous failures off the table, since it can’t change your systems, your customers, or your money on its own. It doesn’t make you safe, though. A read-only tool can still surface confidential information to someone who shouldn’t see it, or hand a person a confident, wrong recommendation they go and act on. So be deliberate about what the AI can reach, because the moment you connect it to everything, people can suddenly see information they never could before, and some of it is confidential, or a client’s and not yours to spread. Know what your vendors do with what you send them. And watch the quietest failure of all: as the volume climbs, “a human approves” turns into a human rubber-stamping. Real approval means someone with the authority and the time to actually say no, spot-checking against what good looks like, not clicking yes forty times a day. None of this is exotic. It’s the same discipline you already apply to money and to client trust, pointed at a new tool.
This is also the honest answer to the fear about jobs. Done well, integration doesn’t clear people out. It gives them their hours back on the research, the writing, the follow-up, the busywork, so they spend more time on the work that actually needs a human. For a sales team that means more time in front of decision-makers and less time on admin. For a marketing team it means more strategy and less production grind. The headcount question is real, and we won’t pretend otherwise. It can mean growing without adding as many people as you would have. But the version that works amplifies the people you already have instead of replacing them, and that’s a choice you make, not one the tool makes for you.
What it costs
The cost question is the one everyone actually wants answered, and few guides answer it straight. The honest version has three layers.
The first is the surprising one. Most of the value, and as a rough rule of thumb we’d put it around eighty percent, is available through the off-the-shelf tools you can subscribe to today, used well. You do not need custom software to begin, and any advisor who leads with a big bespoke build is optimizing for their invoice, not your result. The expensive, custom work has its place later, for the specific problems the general tools can’t reach, but it’s the last mile, not the first.
The running cost is easiest to understand next to what it replaces and improves. We spend well into four figures a month on AI usage now, and each month it’s going up. We expect to climb into five figures, probably before the end of the year. That figure covers everyone in the company leaning on AI across their normal work, not a separate team or a pilot. Set that against what it does across the whole company at once, every hour, without turnover. Part of the return is headcount you don’t have to add. Part of it is the quality of life you give the people already there, the ones who stop grinding on the tedious work. And part of it, the part that’s easy to miss, is that the output itself gets better. Take an experienced expert who lives in the AI world, and their work improves, because their judgment now reaches far more than it could alone.
The real investment isn’t the tokens. It’s the time to build fluency and the discipline to do it in the right order. To make this concrete, and this next figure is illustrative, if a manager spends ten hours a month on a review task that a well-built skill can do in a fraction of the time, the tool pays for itself many times over in the first month, and then keeps paying every month after. We built exactly that kind of skill for one of our own recurring review tasks, and it replaced the better part of a day each month while holding to the standard of the person it learned from. The math on that isn’t close.
Something bigger is happening underneath all of this, and it’s worth naming, because it’s reshaping our whole industry, not just us. The value of pure hourly execution is compressing. Work that used to be billed by the hour is getting cheaper to produce, everywhere, for everyone. A name-brand strategy firm’s six-figure report is no longer safe just because it’s expensive, when a lot of what used to justify that price, the research, the synthesis, the polished deliverable, can now be produced for a fraction of the cost. That’s uncomfortable for a lot of service businesses, and it’s better to see it coming than to pretend it isn’t happening.
Where this ends up
Everything up to here is available now. What makes it worth the effort is where it goes when you keep at it.
The machinery under the picture I opened with is more concrete than it sounds. Every business is heading toward living as a single, accurate digital copy of itself, one place that holds everything the company actually is: the services, the way work gets delivered, the sales pipeline, the people, the finances, all of it. It’s fed constantly from the digital record you already produce, the meeting transcripts, the analytics, the CRM, the documents, and it’s kept accurate by real work, a little more every day, so it stays true instead of drifting out of date. On top of it sits an AI layer that never stops crawling the whole thing, and it surfaces what matters to each person through their own lens. The finance lead sees one thing, the head of sales sees another, the owner sees the whole board. Same source of truth, different view, all of it live.
And it isn’t surfacing trivia, because it knows what the business is actually trying to move. A company like this has defined its North Star, the outcome that matters most, along with the handful of balancing measures that keep it honest so you’re not gaming one number at the expense of the rest, and mapped how every team’s and every person’s work ladders up to them. That part is the one that’s easiest to skip, and the one that matters most, because the AI doesn’t invent your priorities. It optimizes toward the ones you set. Give it a clearly defined goal and it has a direction to work toward instead of just holding up a mirror. From there it hunts for ways to move those numbers and points each person at the specific improvements they can actually affect, theirs, not someone else’s. The salesperson gets the lever that moves their pipeline. The delivery lead gets the one that moves margin. Every small win rolls up toward the same goal at the top.
That is what makes it compound. Each skill you build makes the next one easier, and each decision you write down feeds the layer that finds the next opportunity, until the company I opened with stops being a fantasy and becomes the far end of a path you’re already walking.
That’s why the timing matters. The companies quietly laying the foundation now, the literacy, the small skills, the habit of writing things down, are the ones who’ll be ready when the rest of it arrives. The ones waiting for it to become obvious will be starting from zero against competitors with a multi-year head start on their own institutional memory. This isn’t a far-off promise. It’s the natural end of the same path that starts with teaching one manager to use AI as a thinking partner.
What it actually takes
None of this is a switch you flip.
The hardest part isn’t the technology. It’s change management, every time. Getting people to actually change how they work is where integrations succeed or stall, and it’s slower and messier than any demo suggests. Budget for the human side more than the technical side.
The second reality check is a quiet one that trips up good companies. Most businesses don’t write things down enough. AI runs on the knowledge you can hand it, and if your real operating knowledge lives only in people’s heads and hallway conversations, the layer has little to work with. Getting into the habit of capturing decisions and processes isn’t glamorous, but it’s foundational, and starting it now pays off later regardless of how far you take the AI itself.
Third, there’s a right order, and skipping it stalls the whole thing. We’ve watched an otherwise promising effort lose momentum because it jumped straight from basic literacy toward connecting everything to everything before people could use the simple version. Hold the line on the sequence. Literacy, then small skills, then broader integration. Rushing the middle is how you end up with an expensive system nobody trusts.
And your size changes the weight of it. The bigger the company, the heavier the plumbing. With multiple divisions, real IT ownership, and compliance to answer to, the literacy and the early wins look the same, but who owns the tools, who can see what, and how it’s all governed becomes its own real work. The path doesn’t change. The scaffolding around it scales with you.
And finally, the honest disqualifier. This doesn’t make sense yet if your leadership won’t use it themselves, or if the company won’t write anything down. Those two are prerequisites, not nice-to-haves. If they aren’t there, fix them first, and the rest gets much easier.
Where to start this week
If you take one thing from this, let it be the order. Don’t buy a platform. Teach a few of your sharpest people to use AI as a thinking partner, go watch where their time actually gets wasted, and build one small read-only tool that helps with one real problem. Let that win earn the next one.
That’s how integration actually happens in a mid-market business. Not as a big-bang rollout, but as a compounding series of small, honest steps that add up to a company running on an intelligence layer, and eventually to one that helps optimize itself. It’s real work, and it’s worth it, and the best time to start the literacy is now while it’s still a source of advantage rather than a scramble to catch up.
If you’d like a partner who integrated AI into their own company first, made the mistakes so you don’t have to, and can build the marketing, brand, and web around that same layer, that’s the AI integration work we do inside our strategic advisory practice. We’re happy to tell you where to start, even the parts that don’t involve us.
