Everyone Skipped the Middle

September 13, 2026

Read Time: 8 minutes

McKinsey put out their State of AI survey a couple of weeks ago, and there was one line in it I keep coming back to. (McKinsey, The State of AI, August 2026)

The share of billion dollar companies scaling AI agents went from twenty seven percent to forty percent in a year.

Smaller organisations stayed flat at twenty two percent.

One end of the market moved thirteen points. The other end did not move at all.

I have been saying for over a year that the mid-market is wide open because the labs and the Big Four are all chasing enterprise. This is the first time I have seen the size of that gap printed in a number.

If you have been wondering where the next eighteen months of work comes from, it is right there.

Why the Middle Is Still Standing Still

It is not the money and it is not the technology. Both of those are available to a forty person business today.

The blockers everyone lists are the same everywhere. Integration with existing systems at forty six percent, implementation cost at forty three, messy data at forty two. Those are the obvious ones and they look roughly the same at every size of company.

Here is the one that does not. Small and mid-sized businesses are significantly more likely than large enterprises to get stuck on the human side of it, on employee resistance and on nobody knowing how to use any of it. Fifty one percent of them say so.

Which is the same conclusion I reached last week from a completely different direction. The thing standing in the way is people, not architecture.

Then there is the part that never shows up in a survey. Large enterprises now have people whose actual job is owning this. No forty person business has that role and none of them are ever going to create it.

So nobody inside the business owns it, and nobody outside the business has offered.

Jos closed a ten thousand pound audit with a forty-person construction firm earlier this year. Not because he outbid anyone. Because he was the only person who turned up.

That is the market right now.

Start With Something Boring

Here is where I would push back on your instincts, because I have the same ones.

You will want the first thing you build for a client to be impressive. Impressive is what gets talked about, screenshotted, put on a case study page.

Build the dull one instead. Three reasons.

The failure is cheap. If something that sorts and qualifies enquiries gets one wrong, a person reads it and fixes it in thirty seconds. If your first build touches invoicing, or anything a customer pays money against, your first mistake is a phone call from the owner on a Saturday morning.

The result is provable. Response time before, response time after. One line on a report and nobody argues with it. Try proving that a clever content system improved anything and you will be in a meeting about attribution for an hour.

It buys you the second project. Nobody hands over their accounts payable process because of a slide. They hand it over because the boring thing you built in March has quietly run for five months and nobody has had to think about it once.

That is how this compounds. The first build is not the win. The permission it earns is the win.

Where I Actually Look First

Back in July I wrote that follow-up speed is probably the sharpest fine margin in this business. I still believe that, and the obvious move is to point the first build straight at it. Something that watches the enquiry inbox, sorts it, drafts the reply and books the call.

Let me put a limit on that before you go and sell it, because I overstate this one sometimes.

It only matters where there is volume. Local and service businesses with a genuine flow of enquiries coming in every day. If you are getting one lead a day, you should already be excellent at responding to it, and if you are not, an agent is not your problem.

Where it does apply, it applies hard. Service businesses of every size, big or small, need to get to their leads faster and on a more personal level than they currently manage. That has been true of almost every one we have worked with.

But the place I actually look first is broader than that, and it is the thing I think has the most meaningful impact on a business.

Human error.

Not mistakes exactly. Forgetting. The enquiry that sat in a shared inbox because everyone assumed someone else had it. The invoice that went out eleven days late because the person who sends them was on holiday. The call that happened on Tuesday and never made it into the CRM, so the next person to speak to that customer starts from nothing and asks them the same three questions again.

None of that is anybody being bad at their job. It is just work that depends on somebody remembering, and people do not remember reliably. They never have.

That is what I am looking for when I walk a business. Not the impressive automation opportunity. The places where the process only works if a human holds it in their head.

Support and enquiry inboxes. Invoices going out. CRM notes after calls, so the sales team and the support team both actually know what is going on. None of these are glamorous and none of them are hard to build, and they take a real cost out of the business immediately.

Nobody puts a number on forgetting. It never appears on a P&L. It is one of the most expensive things in most mid-sized companies and it is completely invisible, which is exactly why it is still sitting there.

And it is not one agent. Worth being straight about this.

What you are actually building is four or five small, separate workflows that each do one thing. It is tempting to sell it as a single centralised system that handles all of it, because that sounds more impressive and it is much easier to put a big number on.

Do not. The centralised version is precisely the thing that breaks, and I will come back to why in a minute.

One last practical note. Do not use the word agent when you are in the room. It means four different things depending on who you ask and half of them are things your prospect read a frightening article about. Describe the job instead. Nothing gets forgotten in the enquiry inbox. Every call ends up in the CRM. Invoices go out the same day.

And stay away from the AI employee framing the software vendors use. They use it deliberately to move the buyer from a software budget to a labour budget, because a labour budget is a much bigger number. It is a bigger number because it is the replacement pitch, and you know exactly why I am not making that one.

Don't Sell Them Something They Could Have Bought

There is a temptation here to become a reseller. Find a tool, set it up, take a margin.

I would not build a business on it.

If the problem could genuinely be solved by a product they could subscribe to, they would already have subscribed to it. These owners are not short of software. They are drowning in it. What they are short of is anyone who understands their actual process well enough to make software do something useful inside it.

That is the job. It sits on their systems, their data, their workflow, their exceptions, the four things their operations manager does manually every Friday that nobody has ever written down.

It is custom and it is one-off, and that is not a weakness in the model. That is the model, at least for now. Every consulting market starts bespoke. The productised version comes later, and it comes from having done a hundred bespoke engagements that told you exactly what to build. That is route seven, and it is where Consultix came from. You do not get to skip to it.

Price It Like a New Market

Now the part I feel strongly about, because I see people get it wrong constantly.

Go and look up what an AI agent build costs and you will find a comfortable middle. A few thousand to build, a few hundred a month to run, everybody sitting politely in the same range.

Do not sit in that range.

In a brand new market there are only two positions that work. You are either productised and cheap, selling the same thing a thousand times at a low price, or you are bespoke and expensive, selling something genuinely custom at a premium. Those are the two ends and both of them are real businesses.

The middle is where you do custom work at commodity prices. You carry all the cost of bespoke delivery and none of the pricing that justifies it. It is the worst seat in the house and it is the one most consultants take, because a middling price feels safer than a high one.

You are not productised. You are in someone's building, mapping their process, building something that exists nowhere else. Price accordingly.

And yes, somebody will undercut you. Somebody always does. They will also be the one delivering a generic setup to a client who then decides the whole category does not work. That is not competition, that is a different business, and losing that deal is not a loss.

Ninety Five Percent Reliable Is Thirty Six Percent Reliable

Now the honest part, because this is where I think a lot of this year's work is going to go wrong.

Gartner predicted back in June 2025 that more than forty percent of agentic projects would be cancelled by the end of 2027, on escalating costs, unclear business value and inadequate risk controls. That number still gets quoted everywhere as though it landed last week. The reasoning behind it almost never does. (Gartner, June 2025)

Here is the arithmetic that explains it.

A step that works ninety five percent of the time, run twenty times in a row, gets you to the end correctly about thirty six percent of the time.

Read that again, because it is the single most useful thing in this issue.

Reliability does not decay gently as you add steps. It collapses. Which is why anything sold as running autonomously for days with nobody near it is going to fail, and why it fails for mathematical reasons rather than because the models are not clever enough yet.

It is also why I told you not to build the centralised version. Four small workflows that each do one thing, each with a person able to see the output, will survive. One long chain that does all four will not.

So be honest about the ceiling, with the client and with yourself. You are not automating a department. You are automating four steps of a fourteen step process and handing somebody back their morning.

That is a much smaller promise than the one in every vendor deck this month. It is also one you can actually keep, which over three years is worth considerably more than the big one.

Who Owns It After You Leave

There is nobody in a forty person business who is going to maintain what you built. There just isn't.

So one of two things happens. Either it breaks in month three and quietly gets abandoned, and you lose the client and the case study with it. Or it does not break, and you own it by default.

The second one is the business. Maintenance is still the most passive revenue in this entire model and still the thing most consultants leave on the table.

But you have to say it at the point of sale, not discover it in month four. Tell him plainly that somebody needs to watch this, that it will not be anybody in his building, and put a number on it there and then.

Consultants lose more recurring revenue to politeness than they ever lose to competition.

The Real Point

The big end of the market moved thirteen points this year and the middle moved nothing.

Not because they can't afford it. Because nobody owns it inside the building and nobody has walked in from outside and offered.

So walk in. Go looking for where the process only works if somebody remembers, build four small things rather than one clever one, build them custom because that is the only version worth buying, and charge like a bespoke operator rather than a commodity one.

The gap will close eventually. It is not closing this year.

See you next week,

– Andrew

💡 How I Can Help

AI Audit Workshop ($9): Get the complete step-by-step system for conducting professional AI audits. Includes all templates, interview scripts, opportunity matrix frameworks, ROI calculators, and presentation slides I use to charge $10K+ per audit.

Start Your AI Consulting Business: Join other entrepreneurs building their AI audit businesses. Get the 90-day roadmap, real-time feedback from me, all automation workflows, outbound campaign setups, and everything you need to start landing clients.