The Dispatch · Field Guide No 002 · HR AI Institute

There is a structural truth about enterprise AI that almost nobody has said out loud yet. Once you see it, you cannot unsee it.

This is not a conspiracy. It is not malicious.

It is a rational response to the business model they chose. Every major frontier provider like OpenAI, Anthropic, and Google prices inference per token. Every feature they ship, from longer context windows to multi-step reasoning, is marketed as a capability improvement. And every single one of them is also a token multiplier.

Your bill scales with motion. Not with results.

In the pilot phase, this does not matter. Budgets are experimental. Usage is low. The conversation inside the company is still focused on how fast we can adopt this technology. But we are now entering a different phase. The conversation is shifting. You have probably felt it already in your own organization.

The correction is coming. Here is a look at four predictions about what it is going to look like, and why the People function is going to feel it before the rest of the enterprise does.

Uber recently reported a massive AI budget overrun. It is not an outlier. It is a preview. Klarna spent two years loudly claiming AI had replaced its customer service function, then quietly walked that back and started rehiring humans. Enterprise renewals are coming in with severe sticker shock. Companies that deployed these tools at scale are beginning to ask pointed questions about what they are actually getting.

The pattern is forming in public. And if you have been watching enterprise technology cycles for long enough, you already know what comes next. We have seen this movie twice before.

Round one was the cloud. In the early 2010s, every engineering team spun up unlimited instances. It felt magical. You could build anything at any scale. Then the bills started arriving. By 2015, a new discipline was born: FinOps. The permissive early phase gave way to disciplined architecture. The cloud did not disappear. It matured.

Round two was SaaS. A few years later, every team bought every tool. Marketing had six different attribution platforms. Engineering had twelve observability tools. Then procurement came knocking. Contracts got renegotiated. Tool portfolios got rationalized. SaaS did not disappear either. It got organized.

Enterprise AI is now entering round three. The arc will look incredibly familiar.

CFOs will soon start walking into quarterly business reviews and asking the question almost no one has a clean answer for.

Show me the ROI per token.

They will not ask if people are using it. They will not ask how many seat licenses were deployed. Those metrics are survivors of the adoption phase. They will get set aside in favor of a different question: What did we produce with this spend that we could not have produced without it?

The "adopt everything" phase ends very soon. The "prove it" phase begins immediately after. Budgets do not necessarily shrink, but they get redirected. The organizations that can answer the ROI question keep spending. The ones that cannot face a meaningful reallocation. This is what every enterprise technology wave goes through when it crosses from novelty to infrastructure.

The quality gap between open-source models and the leading closed models is closing faster than most observers realize. For the overwhelming majority of enterprise inference workloads like summarization, classification, and routine reasoning, the gap is already effectively closed.

And the economics are not close.

Token-priced inference scales linearly with usage. There is no buyer-side economy of scale. A 5,000-person company putting everyone on frontier models is signing an open-ended bill that grows with every meeting summary and draft email. Self-hosted open-source inference scales sub-linearly. You pay for compute, which has a fixed floor but declining marginal cost. Once the finance team does this math, the conclusion is inescapable.

But here is the more interesting part. While open-source is closing the gap on general-purpose tasks, vertical specialized models are widening a different gap.

A small model fine-tuned on your company documentation will outperform a massive frontier model on questions about your company documentation. A small model trained on legal contracts in your jurisdiction will outperform generic models on your legal contracts.

The future of enterprise AI is not one model to rule them all. It is a portfolio. You will use open-source for the volume, vertical specialists for precision, and frontier models only for the 5% of workloads that actually justify the premium price.

This is where the People function sits directly in the path of the correction.

Right now, every CHRO dashboard has an adoption rate panel. Percentage of employees actively using tools. Training completion rates. Tokens consumed per business unit. These metrics will not survive the correction.

Usage tells you whether people are pushing the button. It tells you nothing about whether pushing the button made anything better. In a cost-disciplined phase, that gap between motion and outcome defines which initiatives survive and which get cut.

The management question is about to flip. Right now, People leaders are asked how to accelerate adoption. Six months from now, the question will be: what did people produce with the adoption we already bought?

Tickets resolved per hour. Code shipped per sprint. Sales cycles shortened. Manager effectiveness scores. Outcomes.

For the last three years, the implicit job of the CHRO in AI transformation has been adoption evangelist. The job for the next phase is performance architect. Which roles, using which tools, in which workflows, produced what outcomes, and at what cost per outcome?

The CHROs who see this shift coming will become some of the most strategically important leaders in their companies.

Prices on pure inference will drop. But the providers will not be fighting on that turf by then. They are already pivoting.

Watch what the big players have been shipping in the last six months. It is not just better models. It is agent workflows, coding environments, and vertical suites for sales and finance. The model is becoming a feature inside a larger product. The margin is moving from raw inference to orchestration and outcomes.

The frontier providers are going to retreat up the stack to where the margin actually is. The product becomes an operating system for AI-mediated work. The companies who architect for that reality now will have a multi-year head start over the ones still running everything through a single premium API endpoint.

Here are three questions to bring to your next executive meeting. How you answer them tells you where your organization sits on the curve.

If the answers are vague, or if the room goes quiet, that is highly useful information. It means you are still in the pre-correction phase. The companies that move first in this window will be the ones you do not hear from for eighteen months, and then suddenly cannot catch.

AI providers optimize for token consumption. They do not optimize for your business outcomes.

That is not a complaint. It is a description of the incentive structure. Understanding it is the beginning of designing around it.

The correction is coming. The question is not whether it will happen. The question is whether you are designing your response or waiting to react to someone else's.

What changes are you making to prepare your teams for this shift? Share your thoughts in the comments below, and follow along as we continue to break down the realities of building modern businesses.