AI Won't Replace Your SAP Consultants. It Will Replace the Layer You Never Needed.

This article is the first of two in the series "What AI Actually Replaces" — a short arc on where AI lands in the SAP world and the IT department. This piece argues it doesn't come for expertise; it comes for the friction layer. The second asks what your IT department is actually for once execution stops being the constraint.

Here is something I noticed before I had words for it. When AI started landing on real projects, the people who panicked were not the experts. They were the middle.

The developers I work with, the ones who actually write integrated programs against a client's twisted landscape, looked at the AI code generation and said some version of the same thing. Impressive, genuinely. But not fast enough yet where it counts. It writes a clean function. It does not write the thing that has to survive contact with this client's data, this client's twelve years of accumulated exceptions, this client's one consultant who left in 2019 and took the context with him. The experts were curious. They were not afraid.

The people in the middle were afraid. They were right to be, too. They just had the reason wrong.

The consensus is half true, and the true half is the trap

Let me steelman the fear properly, because it deserves it.

The demos are real. SAP has shipped a foundation model trained specifically on ABAP, and it is meaningfully better at the language than a general model that hallucinates method signatures. The roadmap puts agentic ABAP development, custom-code migration agents, and code generation inside the tools developers already use. SAP itself is targeting something like a forty percent efficiency gain on custom-code transformation work, on its own early evaluations. And even its own tooling tells on the limit: the automated quick-fixes for an S/4HANA migration handle roughly forty to sixty percent of the common simplification cases, and a developer fixes the rest by hand.

So yes. If your mental model of a consultant is "a person who turns a written instruction into working code," AI is coming for that, and the timeline is short.

But that mental model was always a description of the wrong thing. It described the part of the job that was never really the job, the part that looked like expertise and was actually transcription. And the reason the middle layer panicked is that somewhere it knew it had been living on that description.

The mechanism: an IT project is a chain of friction

Walk a requirement through a real project and count the times it gets re-encoded by a human whose only job is to re-encode it.

It starts in a workshop. Someone speaks a need out loud. That gets transcribed. The transcription becomes a synthesis email. The email spawns a slide deck because someone needed it "clarified for the steering committee." The deck feeds a business process model. The process model feeds a textual requirements document. That feeds a functional specification. Every one of those artifacts is tracked as a ticket, in JIRA, at every stage, and the whole thing is paced by status meetings run off an Excel sheet that someone maintains by hand.

V-model or agile, it doesn't matter. The ceremony differs. The chain doesn't. On the rollouts I've seen, the information gets rewritten by hand five, six, seven times between the person who has the need and the person who changes a value in a configuration table.

Most of those rewrites add no judgment. They add format. They move the same meaning from one container to another and lose a little of it each time, like a photocopy of a photocopy. Each rewrite is also a job, or part of one. A whole stratum of roles exists not because the work is hard but because the information travels badly and someone has to carry it across each gap.

That stratum is the layer. And friction is exactly what a language model eats first.

What AI actually reveals

AI does not respect any of the categories we use to feel safe. It does not care whether you are junior or senior, internal or external, onshore or offshore. It does not read a title or a rate card. It sorts on one thing: whether what you do is judgment, or whether what you do is friction.

This is where it touches something I have argued before, about offshore. Two decades of offshore growth captured value on exactly this kind of work: low-judgment, high-repetition tasks that were hard to automate before and are not hard now. When a ticket says, in effect, "read this spec, put this value in this customizing table," that is not consulting. It is a relay.

An agent that costs a few dollars a month does that relay. It does it without the day rate, and without the management surcharge stacked on top of the operational resource, which is the part clients quietly pay for and rarely examine. I will say this plainly as a terrain judgment and not a forecast: the body-shopping end of offshore, the part that was always a relay wearing the costume of expertise, is the most exposed thing in this whole picture.

And the people whose economy was built on it know. India's own government policy body, NITI Aayog, put it in a report late in 2025: in a business-as-usual scenario, AI could pull the country's tech-services headcount down from roughly seven and a half to eight million toward six million by 2031. The same report insists the outcome is a choice, not a fate, and sketches up to four million new AI-related roles if the country reskills fast enough. But read which roles it flags as first in line for redundancy: routine QA engineers and level-one support agents. That is the friction layer, named. Nobody in that report is worried about the people who exercise judgment. They are worried about the relays.

The sharper Indian commentary has named the shift more precisely than most Western coverage. The threat was never that AI writes a script faster than a junior. It is that Western enterprises now run autonomous agents over entire workflows, moving from "outsource to India" to "insource to software." Same work, no relay.

None of that is the AI replacing expertise. It is the AI replacing the relay, and the relay was never expertise. It just billed like it.

What resists, and why it isn't sentiment

You read everywhere that "human tasks can't be automated." Fine. But what does that mean on an actual project? Said that way it is a comfort blanket, not an argument.

Here's the concrete version, the one every SAP finance consultant has lived. IT runs the business. It has to compose with the demands of the functions and of general management, and that means it runs on information nobody wrote down. An AI works on the explicit. The real project lives in the implicit.

Take something as mundane as a calendar. The IT project plan wants its testing window, its data migration cutover, its go-live weekend. The finance department has a monthly close, a quarterly close, an audit. These collide, constantly, and when they collide it is the project that adapts, because you do not tell a finance director to delay a statutory close so your migration can have its weekend. None of that constraint is in any specification. It lives in the knowledge that this controller is already underwater in week one of the month, that this CFO will not move on the quarter no matter what the steering committee decides, that the right week to push hard is the third one. You learn it by sitting across from people and reading what they haven't said. An AI cannot collect what was never emitted.

This isn't romance about the human spirit. It's a fact about where the information lives. The implicit can't be ingested because it was never written down, and that is the whole moat: narrow, unglamorous, made entirely of the things nobody put on paper.

So what the CIO should actually do

The wrong reflex is the obvious one. Take the productivity gain the vendor advertises, divide your headcount by it, and cut. That treats every role as the same substance, and the whole point of what AI reveals is that it isn't.

The better move is to map your own friction before the technology maps it for you. Walk your own chain. Who is translating? Who is recopying? Who is carrying meaning across a gap a tool can now close? Then the harder question: who is exercising judgment that lives on context no document holds? Those are not the same people, and they are very often not the people the rate card or the org chart would have you guess.

Do that audit yourself, on purpose, and you keep the judgment and shed the friction deliberately. Wait, and the technology does the audit for you, indiscriminately, and you find out what you were paying for only after it is gone.

That is the shape of this moment. AI is a revealer before it is a replacer. It is going to show every organization, in detail, what it was actually paying for all along, and the layer that was friction wearing the costume of expertise comes off first.

Which leaves the question this piece doesn't answer and the next one has to. If execution is becoming nearly free, what is your IT department actually for? That's where the next piece goes: The Question Isn't What AI Replaces. It's What Your IT Department Becomes When Execution Is Free.