“AI consultant” is, right now, one of the hottest titles in business. Putting “AI” in front of “consultant” opens doors: the budget is there, the demand is there, companies are actively hunting for someone who can walk in, look at their operation, and tell them what to actually do with all of this.

Take it. It works. But the title has an expiration date, and we have seen this movie before.

In the nineties, some people introduced themselves as “Excel accountants.” In the 2000s, “internet marketing” agencies sprang up everywhere. Today both sound odd: Excel is just Excel, and internet marketing is just marketing. AI is headed the same way, because it is going to seep into everything. In a few years the “AI” drops off and you are left with “consultant” again.

I have been programming since 1990. I lived through the Excel wave and the internet wave from the inside, writing the code, not reading about it in a report. And the part that did not change in 36 years is the part underneath the title: you walk into a company, you find the real constraint, and you prescribe the simplest solution that holds. The title is the arbitrage of the moment. The craft is the moat.

That craft has a concrete shape, and Nate Herk drew it a few days ago better than anyone: a pyramid. It is the thesis of this post, and the reason AI Maestro exists to tell you which tier of that pyramid your problem lives on, before you pay for the one above it.

The pyramid: three tiers, and almost everyone jumps to the top

The pyramid has three levels, and the gravity of AI projects always pulls toward the top, because the top is the one that sounds impressive in a meeting.

Tier	What it is	What it costs	When to climb
Base	Deterministic workflow, zero AI: a fixed rule, a better tool, a restructured database	Cheap, fast, predictable. Barely breaks.	When fixed rules no longer cover the real variation in the problem
Middle	AI workflow: a model inside steps you define (classify, extract, draft)	More capability, more cost, new failure modes	When you need the model to read messy language, but the path stays fixed
Top	Agent: the model directs its own path and uses tools on its own	Maximum capability, maximum cost, slowest to ship, compounding errors	When you cannot predict the steps and the decision changes every time

Read that last column top to bottom. Every tier you climb gives you more capability, yes, but in exchange for more money, more time to production, and more surfaces where it breaks. It is not a prestige ladder with the best thing at the top. It is a cost-and-risk ladder, and the most fragile thing is at the top.

And this is not coming from an AI skeptic. It is Anthropic, the lab that makes Claude, in its own guide to building agents: “finding the simplest solution possible, and only increasing complexity when needed.” The same document is blunt about the cost of the top tier: agents trade latency and money for capability, and their autonomy brings “higher costs, and the potential for compounding errors.” When the company selling the most capable model on the market tells you to start with the simplest thing, it is worth listening.

How to know which tier your problem belongs on

The discipline fits in one sentence: start at the base and climb only when the problem forces you to, not when the agent excites you.

In practice it is a short interrogation. Is the input structured and are the steps predictable? Then it lives at the base: a rule, an integration, a better tool out of the ones that already exist. Zero AI, and you will sleep fine. Do you need something to read messy human language, an email, an invoice, a ticket, but after that the path goes back to being fixed? That is the middle tier: a model placed inside steps you control. Can you genuinely not write the steps because they change in every case and the decision is different each time? Only there, in that case and not before, do you justify an agent.

The expensive mistake is not picking the wrong tool. It is climbing a tier with no evidence. “It would be more elegant with an agent” is not evidence. The only valid reason to climb is that you tried the tier below and it fell short against the real variation in the problem, not against the variation you imagined.

Starbucks paid that bill in plain sight. It deployed an AI inventory tool across 11,000 stores and pulled it nine months later. That was not a model problem. It was climbing to a tier the problem did not call for, at scale, without testing the one below first.

The title expires. The craft does not.

Back to the start, because the two halves of this post are the same idea.

Nate Herk closes it well: grab the label while the window is open, but do not confuse it with the job. The job, underneath any fashionable title, is the old one: walk through an operation, find the real constraint, and prescribe the simplest solution that resolves it. Sometimes that solution is an agent. Far more often it is a restructured database, a swapped tool, or a ten-line rule.

The “AI consultant” who does not understand the pyramid will prescribe agents for everything, because an agent is what the title promised the client. And they will lose the client’s trust the day the client realizes they paid for an industrial drill to hang a picture. The one who does understand it will recommend the boring tier, without flinching, when the boring tier is the right one. That honesty is the only thing that survives once everyone has “AI” on their card and the qualifier stops meaning anything.

What we do about this at IQ Source

AI Maestro is, in one line, the discovery that places each process in your operation on its tier of the pyramid. Two months mapping the real processes, not the ones on the org chart. We give each one an AI Opportunity Score, which at bottom measures one thing: how high up the pyramid that process justifies climbing. And at the end there is a Go/No-Go gate that decides, process by process, whether it stays at the deterministic base, moves to an AI workflow, or genuinely earns an agent.

The result that surprises clients most is that most processes stay at the base. Zero AI. A better tool, a missing integration, a rule. That is what a vendor who charges to install agents will never tell you, because their business lives at the top of the pyramid and yours lives at the bottom.

And when a process does earn the top tier, Tech Partner is the person who answers for how that agent was built, which model runs underneath, and what happens the day it breaks, because at the top it does break. That is the nature of the tier. This, by the way, is the opposite of deciding what not to automate: it is not just a yes or no; it is “and if it goes, at what level.”

Next time someone in your company says “let’s put an agent on it,” do not ask whether it is possible. It almost always is. Ask which tier they were standing on when they proposed it, and whether anyone tried the one below. That bottom tier, the deterministic, boring one, is where most of your return lives. The agent is the expensive exception you earn, not the place you start.

Place your processes on the pyramid