The Economics Era of AI

Northshore Mutual is a regional insurance carrier in the Midwest. Their VP of Underwriting watched a competitor publish a press release about "AI-assisted underwriting." Two weeks later she had a budget. Two months after that, her team had a working prototype.

Three questions you should be uncomfortable with before this lesson tells you anything. Click each to reveal a starter framing.

1. Why did Northshore's project stall after the demo worked?

   Hint: ask where the output goes. If "the output" lives in a Slack channel and "the work" lives in a policy admin system, the work has not changed.
2. Why is the same kind of model becoming more expensive to train each year while becoming more valuable to deploy?

   Hint: when capability is concentrated in a few well-capitalized labs, the rarest skill in the chain shifts from producing capability to translating it.
3. Who, on Northshore's team, is responsible for closing that gap — and what would you have to teach them?

   Hint: every named talent role exists because a specific bottleneck exists. ML scientist solves "can it work?" ML engineer solves "can we deploy it?" Something else solves "is it producing value?"

By the end of this lesson, you will be able to:

Backward design (Wiggins & McTighe, 1998): every lesson is built around its assessment. The independent task in §7 assesses exactly these three outcomes; everything between here and §7 exists to make that task possible.

Bring to mind one AI project you have personally seen — at your company, at a customer's, or in the press — that started with energy and ended in a slide deck. Take 60 seconds and answer the three prompts below in your journal. You'll use this example again in §10.

The three eras of enterprise AI adoption

Enterprise AI has moved through three eras. The eras are not strict chronology — every organization, and every project, sits somewhere on the map — but the field as a whole has shifted its center of gravity.

The driver: the cost curve of frontier training

The economics framing is not rhetoric — it is forced by numbers. Epoch AI's analysis of 45 frontier-model training runs since 2016 found that the amortized hardware-and-energy cost for the final training run of a frontier model has grown by approximately 2.4× per year (95% CI: 2.0×–3.1×). At that rate, the largest publicly announced training runs are projected to exceed $1 billion by 2027 (Cottier et al., 2024).

Why the curve forces an Economics Era

If a frontier training run costs ten times what it cost three years ago, frontier capability is becoming concentrated in a small number of well-capitalized vendors. The supply of frontier capability is therefore abundant from the buyer's perspective — Claude Opus 4.7 (Anthropic) and GPT-5.4 (OpenAI), the leading commercial frontier models in May 2026, are both available by API to anyone with a credit card. The bottleneck has moved.

• In the Research Era, the bottleneck was capability.
• In the Experimentation Era, the bottleneck was access.
• In the Economics Era, the bottleneck is the translation of abundant capability into measured operational value inside a customer's actual systems, under the customer's actual constraints.

The FDE is the named role for that bottleneck

Every era has its named talent role. Research had the ML scientist. Experimentation had the ML engineer. Economics has the Forward Deployed Engineer — a hybrid of software engineering, product strategy, and implementation consulting whose accountability is the customer outcome end-to-end. That is the thesis of the program. Every other lesson assumes this one is in place.

Return to the hook. Northshore built a working underwriting-summary prototype on Claude Opus 4.7. The prototype produces good summaries. Six months later, it has zero users. Walk through the diagnostic with me — click each move to reveal the expert reasoning.

Brightspire Lending is a consumer-lending fintech with about 200 employees. Their data team built an LLM-assisted risk classifier that reviews unsecured loan applications and flags likely defaults. The classifier runs in a sandbox alongside the loan origination system. It is not yet integrated. The Head of Risk has not signed off on production criteria. The CFO has asked twice what the dollar value of the project is; both times the data team answered with model accuracy.

Scaffold drop. Questions 1–4 were answered for you; Q5 demands synthesis — combining era classification with the SoR / metric / Day-2 owner pattern. That is the upper edge of your Zone of Proximal Development for this lesson (Vygotsky). Productive struggle there is the whole point.

Brief

You are the newly hired Forward Deployed Engineer for a private equity firm that has just acquired stakes in three companies, each running an AI project. The managing partner has asked you for a one-page memo classifying each project by era and naming the one change that would advance it to the next era. The memo will go to the board; she does not have time for nuance.

Three project descriptions — classify each

Write the memo

For each company, write three things: era classification (one word), the single determinative piece of evidence (one sentence), and the one change that would advance it by one era (verb-first sentence). Lead with the answer to the managing partner's question (Pyramid Principle), not with context.

Retrieval, not summative grading. The three items below queue spaced review and flag what to revisit before Lesson 1.2.

Refer back to the project you brought to mind in §3. Write a short paragraph for each prompt below. The point is not to feel introspective; it is to give your future self a hook for retrieval when the framework is needed in the field.

One flashcard, three retrieval visits — at 1 day, 7 days, and 28 days. Drop from rotation only after three successful recalls at the 28-day interval.

Success criterion: you can articulate the abundance-vs-translation asymmetry — frontier capability is concentrated and abundant via API; translating it into operational value is scarce — and give one concrete example from your own domain.