AI is unbundling the hourly consulting pyramid, and forward-deployed engineers are the counter-move

The Financial Times reported on August 31 that companies are moving technology projects back in-house because AI has made coding, implementation, research and analysis cheap enough to do themselves. The pressure lands on one specific product: the large-team engagement billed by the hour. Consultancies are repositioning toward AI implementation, cybersecurity and organisational transformation, which is a polite way of saying the old product is not selling the way it did.

I run a one-person practice built on AI leverage, so none of this surprised me. The pyramid was always the first thing AI would hit.

Why the pyramid is the exposed part

The traditional model is a partner on top, a few managers, and a wide base of juniors producing research, slide decks and code. The base is where the margin is: bill the junior at a multiple of their cost, and the difference funds everything above them.

AI removes the reason the base exists. A model produces the first draft of the research, the deck and the code in minutes, at a cost that rounds to zero next to a junior's day rate. What is left is the judgement about whether that draft is right, whether it fits the client's constraints, and whether it should be built at all. That judgement has always lived at the top of the pyramid, and it was always the smallest part of the invoice. The FT's summary of where value is moving is exactly that list: governance, data architecture, security, and large-scale deployment. Senior judgement, not headcount.

Hourly billing priced the pyramid's base. AI deleted the base and left the invoice.

There is a second-order effect that IEEE Spectrum caught on September 8: the volume of AI-generated code is forcing companies to retrain junior engineers as reviewers who can spot subtle defects, rather than as authors. The client's own juniors are becoming the verification layer. That is the in-house capability that used to be outsourced.

The counter-move: forward-deployed engineers

The Wall Street Journal reported on September 8 that Google Cloud and Accenture launched a joint unit with 1,000 forward-deployed engineers who work on site at customers, moving Gemini Enterprise projects from pilot to production. This is the opposite of the pyramid. No juniors producing artefacts; senior engineers embedded in the client's environment, measured on whether the thing ships.

It is also a very specific bet. Those 1,000 engineers deploy Gemini. The unit exists to turn Google Cloud pilots into Google Cloud production, and its economics depend on the platform pull-through. A small practice cannot match it on volume. But it opens the gap a small practice can occupy: the client who wants the pilot to reach production without deciding, in the same contract, which vendor's stack they are married to.

What a small AI-augmented team sells instead of hours

• Outcomes with a number attached. Not an AI strategy but the token bill down by a measured percentage with quality gates that prove it, priced as a project. The audit method in 99 percent cost architecture is the kind of deliverable that survives an in-house review because the client can rerun it.
• Vendor-neutral routing. A forward-deployed team from one cloud will not tell you when the other cloud's model is cheaper for a task. A frontier model router built by someone with no platform quota is a different product.
• Governance and security as the engagement, not the appendix. The FT's list of where value moved is the list of things a client cannot yet do in-house with a model. Sell those.
• The verification layer. If the client's juniors are becoming reviewers, teach them the review. A senior who builds the eval harness and the deny rules is worth more than one who writes the code the model already wrote.

The honest gap

Both sources are reporting on a direction, not a measurement. The FT does not quantify how much hourly revenue has actually moved in-house, and the WSJ describes a unit that was just announced, not one with a track record of pilots converted. My own view is coloured by running the model I am recommending. The trend is visible in the pipeline I see; whether it holds at the scale of a Big Four practice is a question their next annual report answers, not this article.