The AI Ladder: What MSPs Should Actually Tell Their Clients About AI

Every MSP owner I talk to is getting the same question right now. Some version of “what AI should we be using?” And most of us don’t have a clean answer, because the honest one is complicated and the marketing has spent two years making it worse.

The same question is coming the other way from small and medium business owners. They’re asking it about their own operation, often without an MSP to translate. The advice I’d give either of them is the same.

So I’m going to lay out the answer I’d hand a client. It’s a four-rung ladder, and the most important thing about the ladder is where the line is between the rungs you should climb fast and the one you should walk up to slowly. Three rungs are pure upside for almost any business. The fourth is a cliff.

The reason the question keeps coming back unanswered is that your clients have been sold a story. The story is that there’s a general AI that will absorb their business and run it, and the vendors closest to the technology are quietly spending a year and a fortune to make narrow slices of it work while selling a demo that looks like the finished thing. Meanwhile the genuinely useful, genuinely safe AI is sitting right in front of them, and almost nobody is using even half of it.

That gap, between what people think AI can do and what it actually does well, is exactly the gap an MSP is supposed to close. Nobody else in the client’s world is positioned to. Not the vendor with something to sell. Not the employee who watched one viral demo. The advisor seat is yours if you take it.

Before the ladder makes sense, three reframes have to land. They’re what most “AI strategy” conversations skip over, and they’re the reason most client AI rollouts disappoint.

Reframe 1: It’s an engine, not an employee

The word “agent” is quietly costing people a fortune. It makes you imagine the wrong thing. (For a primer on what people actually mean when they say agentic AI, the category itself is real. What’s misleading is the office-manager picture the word puts in your head.)

When someone says AI agent, the picture in your head is an office manager. You hire them, give them access to everything, train them for a few weeks, and from then on they just know. They hold the institutional context. You hand them a vague task and they fill in the rest.

That picture is half right, which is what makes it dangerous.

A new office manager can’t read your mind either. You say “find the thread from the vendor about the renewal,” and you point them. You guide them the first few times. AI isn’t all that different. Give it your email and it can search, but you have to tell it what matters, who to look for, what you’re actually trying to do. Same as the new hire.

The difference is what happens next. Your office manager learns. After six months they know your vendors, your clients, your patterns, and they stop needing to be pointed. They accumulate. They carry the business around with them.

The engine never does that. It knows exactly what you put in front of it the moment you run it, and nothing more. Next task, blank slate. It needs pointing every single time, because it didn’t keep anything from last time.

So dumping all your email and documents into an AI doesn’t create an office manager who knows your business. It creates a pile of inputs sitting next to an engine that has to be aimed every run. The pile is not the intelligence. What you feed the engine, and at what moment, is the intelligence.

This is why the same AI is your sharpest analyst in one conversation and confidently wrong in the next, with nothing changed but the inputs. The real work, the part that takes a year instead of a week, is building the thing that feeds it. What context this task needs. Which tool, in what order. What to surface and what to hold back. That is not a setting you switch on. It’s a discipline you build.

I’m not knocking office managers. A great one does things AI will never do. They read the room. They catch the thing you didn’t think to ask for. They hold the relationship. A person isn’t a slower engine. They’re a different kind of thing, and mistaking one for the other is exactly the error that costs people a year and a fortune.

Stop picturing an employee you hire. Start picturing an engine you feed. It changes every decision about where AI fits and what it costs to get there.

Reframe 2: It’s not one brain, it’s a set of tiny ones

Once you accept the engine model, the mechanism underneath it explains the other half of the failure pattern.

Your brain is a marvel at something you never think about. Switching. You move from the part of you that knows a client’s network, to the part that remembers a conversation from Tuesday, to the part doing mental math, instantly, without loading anything. It’s all just there, and you cross-reference across it without effort.

AI doesn’t do that well. Think of it less as one brain and more as a set of separate micro-engines, and you have to deliberately tell it which one to be in. You point it at an area, you feed that area the right slice, and it works there. Ask it to fluidly cross-reference six different domains at once the way a person does, and it gets worse, not better.

Two things follow.

First, each engine holds a limited amount at once. That window is growing at an exciting rate, but it is finite. Take email. If you actually wanted the AI to weigh every message before it decided anything, you couldn’t just hand it the inbox. You’d have to chop years of mail into pieces small enough to fit, then figure out which pieces to load for this one question. That work is real and it is expensive. You cannot pour everything in and expect it to sort itself out.

Second, and this is the one people miss. When something is not in the engine’s head right now, it does not exist. There is no background recall. No “oh wait, that reminds me.” A person half-remembers something and goes digging. The engine doesn’t. If you didn’t put it in front of the engine for this run, it is gone, and the engine will answer confidently without it.

This is exactly why “just give it everything” fails. Everything can’t fit in the head at once, and the stuff outside the head might as well not exist. So the skill is not loading more. It’s choosing what’s in the head, deliberately, for the job in front of you.

The people who are great at this aren’t feeding the machine more. They’re feeding it the right things at the right moment and switching it cleanly between tasks.

Reframe 3: Nobody is training the AI on your business

When a vendor says they’ll train the AI on your business, you picture the office manager again. The model studies your company, gets smarter about how you work, and keeps it. Learns you, the way a new hire eventually learns you.

That is almost never what’s happening when a vendor says it to you. Yes, the frontier labs train models, and a handful of companies fine-tune on their own data. But the overwhelming majority of businesses rolling out AI are not training a model at all. The word “training” is borrowed, and it’s hiding what the work actually is.

Two completely different things get called training, and neither one makes the model smarter about you.

The first is configuring the system. Giving it skills, writing the instructions that sit behind it, connecting it to a set of tools. This is the rig that aims the engine at the right inputs. Real work, but it’s plumbing, not learning. The model isn’t absorbing your business. Someone is building the apparatus that points it in the right direction.

The second is the one nobody says out loud. Most of the training is for the humans.

It’s teaching your people to feed it a few good documents instead of dumping in forty. To be explicit and efficient about the output they actually want. To build a reusable setup, a saved prompt, a project, an artifact, so they get the same quality next time with a fraction of the effort. That’s the skill. And it lives in the person, not the model.

The engine never retains. Next task, blank slate. So the model cannot be the thing that improves over time. The only durable improvement comes from two places. Your people get better at directing it. Your setup gets better at aiming it. The intelligence accumulates in the humans and the pipeline, never in the model.

Which flips what “rolling out AI” should mean. The valuable thing isn’t “we trained an AI for you.” It’s “we taught your team to use it well and built the setups that make good output repeatable.” One sounds more impressive. The other actually changes how a business runs. The same logic applies to AI copilots: the copilot doesn’t get smarter about your business. Your team gets smarter about driving the copilot.

When someone offers to train AI on your business, ask the real question. Are you making the model smarter about us, or making us smarter about the model? Only one of those is actually on the table, and it’s the one that sounds less impressive.

It’s not hallucinating, you onboarded it badly

People keep telling me AI “hallucinates.” Most of the time, it didn’t. You onboarded it badly.

I’ve hired and trained a hundred-plus helpdesk techs. When I started working with AI agents, the skill transferred almost completely, because working with an engine you feed turns out to be the same job as onboarding a new hire.

A new tech on day one doesn’t need a bigger brain. They need to know the situation, which systems to look at, and how to think about the problem. That’s the manager’s job, not the hire’s. Drop someone in with no context and then blame them for floundering, and the problem was never the person.

Here’s the part most people miss, and it separates someone who’s run a service desk from someone who hasn’t. Not all good data is good for the same thing.

Some systems hold data you trust for accuracy. The value is ground truth. Use it as fact. Other systems hold data that’s good for direction. It points you somewhere, but it’s a hint, not a truth. A senior tech knows in their bones that a device’s last-boot time is gospel and a ticket’s category field is a guess somebody clicked in a hurry. They weight those two sources completely differently, automatically, without being told to.

That weighting is the expertise. And it’s exactly what people forget to hand the engine.

When you feed an engine directional data and tell it to treat it as fact, it does what you asked. It builds a confident answer on a shaky foundation and hands it to you. Then you call it a hallucination. It’s not. You gave it low-quality input and told it to trust the low-quality input. You’d never let a brand-new tech treat the category field as truth. You just didn’t think to tell the machine the thing you’d never forget to tell a person.

The people who are great at AI aren’t better at prompting. They’re better at onboarding.

With the three reframes settled, the ladder makes sense.

The AI Ladder: From “Everyone Should” to “Almost Nobody Should”

Most people are stuck on the bottom rung of AI and don’t know there’s a ladder. A few are climbing toward a rung they have no business being on. Here’s the whole thing, bottom to top.

Rung 1. Just the chat. Nothing connected. Don’t discount it. From a blank conversation you can build a real financial model from a rough brief, think through a thorny problem, digest a dense document. Most people are using a sliver of what is right here. The catch: when the chat ends, that model is gone. Next time, you start over.

Rung 2. The chat plus Projects. Now the work persists. Save that model into a Project and it stops being a one-off. It becomes the seed. I open that Project and ask how this month’s cash flow is tracking against it, what changed, where I’m ahead or behind, and the engine works from my real baseline instead of a blank page. I think of each Project as a mode, a different hat. One for my financials, one for my LinkedIn voice and calendar, one for all things Junto. The engine already knows which version of me it’s working for. This is the rung almost everyone skips, and it’s the biggest jump on the ladder.

Rung 3. Connectors and the browser. It reaches into your actual tools. I have Claude wired to Granola via MCP, so I search every meeting I’ve had from inside the conversation. “What did we decide with that vendor in March?” and it pulls the answer from the transcript. No hunting, no tab-switching. Still reading, still assisting. The human stays in the driver’s seat. Honestly this rung is less used than Projects right now, and I think it’s the next big gap to get filled.

Everything up to here is pure upside. Low risk, high return. This is where the vast majority of your clients should live, and almost none of them are there yet. That gap, rung 1 to rung 3, is the entire opportunity, and none of it is fancy.

Then there’s a line. It’s not a step up. It’s a different staircase.

Rung 4 is where the engine stops telling you things and starts doing them. Custom agents, vertical AI, a setup with the access to actually execute in the real world. If you’re an MSP evaluating whether AI agents acting on tickets is worth the risk for your service desk, the mechanism is laid out in how AI ticket triage automation works for MSPs and the outcomes are in the MSP AI ticket management ROI breakdown.

Here’s why that line matters more than any other. On rungs 1 through 3, the worst case is a wrong answer, and a wrong answer wastes your time. On rung 4, the worst case is a wrong action, and a wrong action already happened. You can’t un-send it.

So the rule I’d give any client is simple. Climb to rung 3 as fast as you can. At rung 4, stop and ask one question. Who is holding the guardrails? If a vendor built the safety in for your exact use case, fine. If the answer is “we are,” be honest about whether you know how. Human-in-the-loop AI isn’t a nice-to-have at rung 4. It’s the difference between an agent that helps you and an agent that destroys something. Agents that act don’t lower the skill you need to stay safe. They raise it, because now the mistakes execute.

I live near the top of this ladder all day, and knowing where to be careful took me years. Most people don’t need to be up here. They need rung 3, and nobody has shown them the way up.

Why rungs 1 through 3 are the real magic

For almost everyone, the right move is not a custom agent or some vertical AI platform. It’s to become an absolute power user of rungs 1 through 3 and push them to the limit before looking anywhere else. The magic there is real, and most people are using a fraction of it.

Concrete example. Claude built me a working financial model. Not an outline I finished myself. An actual model. It could not do this three months ago, so if you tried this a while back and gave up, try again. The ground has moved.

But here’s the honest part. It worked because I did a lot of the thinking. I fed it the exact frame I wanted and an example document to match. I used my brain. The engine didn’t replace my judgment. It executed on my judgment at a speed I couldn’t touch a year ago. That’s the real shape of the magic, and it’s the opposite of the “it just does it for you” pitch.

Something that sounds like it contradicts what I said earlier, so let me draw the line clearly. I told you not to dump everything into the engine. That’s true for reference data, the documents it has to search and weight. Pile in too much and you bury the signal. But there’s a different kind of input where more is almost always better. Direction. Your frame, your voice, your priorities, what good looks like, what to ignore.

Brain-dump your direction. Write long. Almost journal into it. You will never physically write enough to overwhelm it with your own thinking, but you can absolutely drown it in documents. Be stingy with data, generous with direction. That’s how you imprint your real voice onto the engine, and it’s why nobody can copy your setup. They don’t have your brain to dump.

One more technique that changes everything. Don’t lurch straight to “build me this.” Explore first. Think out loud with it, let it ask questions, let it push back, develop the frame together. Then flip the switch on purpose. Tell it: stop exploring, now build. Telling the engine which mode it’s in is itself part of the direction.

Do this and you can do things that genuinely felt impossible a year ago, with no API, no engineering, no risk. Just you, a good frame, and a lot of deliberate thinking poured in.

The cliff at rung 4

Only after you’ve pushed the safe rungs to their limit should you even think about rung 4, the one where an agent acts on its own. And you should walk up to that edge knowing it’s a cliff, not a step.

The economics alone will stop you. At rungs 1 through 3 you’re on a flat consumer plan, and for a power user it’s one of the best deals in technology right now. The second you cross into building an actual product on the API, the model flips to pay-per-use, and the bill for anything running at real scale will stop your heart. Same technology, completely different economics. The marketing never mentions that part.

It isn’t just cost. An MVP takes a week and feels like genius. A real product takes a year. The demo is easy. The thing you can trust in production, that handles the edge cases, that doesn’t do something destructive when reality gets messy, that’s the year nobody posts about. Build versus buy quietly died on this exact point. A lot of smart people built the week and discovered the year.

So here’s the whole answer, the one to hand a client.

Use the cheap, safe, powerful stuff aggressively. Put your people on the paid tier and actually learn it. That alone moves the needle for most of a team, and it’s rungs 1 through 3.

Be very careful at the edge where software starts acting on its own. That’s where cost, complexity, and risk all spike at once, and it’s where you need someone who’s been careful standing between the client and the blast radius.

And that last part is the conversation that matters for us as MSPs. “Should we let an AI act on our systems” is the question clients ask right after they fall in love with the safe stuff. It’s the question where they need an advisor who has actually been careful, not a vendor’s landing page. That’s our seat at the table. (For the broader landscape of AI helpdesk and service desk platforms clients are likely to ask about, that’s a separate read.)

The bargain is real. The shortcut isn’t. The person who can tell a client the difference is worth more than any tool.

So I’ll ask the question I opened with. What are your clients asking you about AI, and are you ready to give them the straight version now?