Three experts, many tokens
AI transformation is inverting the staffing model — fewer subject matter experts, heavy token use, human-in-the-loop QA. A field note on delivery economics from Big 4 to solo practice.
Something interesting is happening in how AI transformation work gets staffed.
Not everywhere. Not uniformly. But often enough that I would treat it as a signal, not a one-off.
The old shape was familiar. A programme needed nine people. Three seniors to steer, four mid-levels to draft, two juniors to format slides and chase references. The client paid for the pyramid. The firm billed for utilisation. Everyone understood the choreography, even when the output was slow.
The emerging shape looks different. Three subject matter experts. A large token budget. A human-in-the-loop discipline that treats review as the real work.
Headcount is not the unit of capacity anymore
When companies lean into AI-driven change, they are not always asking how many consultants can we put on this? They are asking who actually knows the domain, and how much AI leverage can we put behind them?
That is a different procurement conversation.
The expert does not disappear. If anything, the expert matters more. Someone has to frame the problem. Someone has to know what good looks like in risk, operations, architecture, or regulation. Someone has to catch the confident nonsense before it reaches a steering committee.
What changes is the middle of the pyramid. The drafting layer, the research layer, the first-pass analysis layer, much of it is shifting to agents, models, and orchestrated workflows. Not as a demo. As production plumbing behind a small senior team.
I am seeing programmes that would once have been sold as a nine-person workstream scoped instead as three SMEs plus platform access plus token spend. The economics are hard to ignore once you have run it once.
Generate, review, assure
The useful mental model is not AI replaces the consultant. It is AI generates the artifact; the expert applies judgment on top.
A practical loop looks like this.
Generate. Models and agents produce drafts, options, comparisons, code scaffolds, policy mappings, test cases, workshop materials. Fast. Wide. Sometimes shallow. That is fine, if you know what comes next.
Review. The SME reads like an editor with domain authority. What is wrong? What is missing? What would fail in our environment? What sounds plausible but is not true for us?
Assure. Quality assurance becomes explicit. Not a polite read-through at the end. A named step. Sign-off with accountability. The human in the loop is not decoration. They are the reason the client can trust the output.
This is where transformation programmes either mature or stall. Teams that treat AI output as finished get burned. Teams that treat it as raw material for expert judgment move faster than a nine-person deck factory ever did.
Same value, different cost structure
From the client’s side, the proposition is attractive.
You still want the outcome. The architecture. The operating model. The governance design. The production path. The board-ready narrative. But you no longer want to fund six people whose primary job was to translate the expert’s intent into documents.
If three SMEs, well tooled, can deliver comparable artifacts with AI doing the first mile, the client receives similar value at a lower price. Or, increasingly, the same budget buys more scope in the same quarter.
That is disruptive for traditional professional services economics. Utilisation-based staffing models assume bodies on the bench. Token-heavy, expert-led delivery breaks that assumption. The margin moves from headcount arbitrage to platform fluency, domain depth, and review discipline.
Big firms feel this first in transformation and advisory work, the places where deliverables are knowledge products anyway. Implementation partners feel it next, wherever specification and code generation compress cycle time.
From Big 4 pyramid to one-person practice
The inversion runs in both directions.
At the top end, a Big 4 or high-end consulting brand still brings trust, method, and access. But even there, the internal staffing story is changing. The partner pitch may still mention a team. The delivery reality is increasingly a few heavy hitters orchestrating a lot of machine output, with AI doing the volume work under firm QA standards.
At the other end, a one-person practice, or a two-person boutique, can now produce artifacts that used to require a back office. Not because one human became superhuman. Because they can spin up agents, run structured prompts, chain tools, and spend tokens like a small studio spends render hours.
That does not remove the need for credibility. It raises the stakes for it. When everyone can generate a plausible strategy deck, the differentiator is whether someone who has actually shipped in production has reviewed it. Domain expertise scales differently than slide production.
I find that honest. It rewards people who have done the work, not only people who can staff the work.
What this means if you are buying or building
If you are a buyer of transformation services, ask sharper questions.
- Who is the named SME on the hook for judgment, not only for project management?
- What is the human-in-the-loop model for drafts, code, and recommendations?
- Where does QA live in the workflow, and who signs off?
- What token and platform costs are assumed, and who owns them after the engagement?
- Are you paying for headcount or outcomes? Both can be valid. They are not the same purchase.
If you are building an internal AI transformation function, the same logic applies. Do not recreate the old pyramid with chatbots attached. Hire for expertise and review capacity. Invest in orchestration, memory, and tools. Budget tokens like you once budgeted contractor months.
If you are a solo practitioner or small firm, this is an opening. Not to pretend to be a Big 4. To be unambiguously expert, visibly rigorous, and structurally lean. Clients will compare the artifact, then look for the person who will stand behind it.
The work that remains human
None of this removes politics, alignment, or accountability. If anything, it concentrates them.
Fewer people in the room means each person carries more weight. Review becomes more visible. Errors are less diluted across a large team. Expertise stops being a title and becomes a load-bearing function.
That is the pattern I see now. Not fewer humans in transformation. Fewer humans doing mechanical knowledge work, more humans doing judgment under pressure. A lot of tokens spent. A small number of people worth paying for.
The firms and practitioners who get this right will look oddly small on the org chart and oddly large in output. I think that is the point.
Field note from the build-in-public log. NDA-safe, no client names, rounded figures only. If this matches what you are seeing, get in touch.