Strategy for organizations

The lesson’s arithmetic

All of the shop’s figures are illustrative: they show how the calculation is put together, not numbers to copy. What matters is the shape of the formulas and that every assumption stays in view.

Value at stake. With $f$ occurrences a month and $m$ minutes saved each time, the hours of work at stake per month are

Warranty questions: $2{,}000 \times 6 / 60 = 200$ h; diagnoses: $300 \times 30 / 60 = 150$ h; résumés: $20 \times 6 / 60 = 2$ h. If the value is not time (sales, quality, fewer errors), use whatever unit makes sense, but always multiplied by frequency.

Cost per task (lesson 26). With $t_i$ input tokens at $p_i$ dollars per million and $t_o$ output tokens at $p_o$:

Even though output is less than a tenth of the tokens, it accounts for 30% of each question’s cost: each output token costs five times as much.

First-year total cost of ownership. With an upfront cost $I$ (building and integrating, 4,000; preparing data and gold cases, 2,000), a fixed monthly cost $F$ (hosting and monitoring, 70; people, 40 h × 10 = 400), and $V$ questions a month:

Tokens are $12 \times 30 = 360$, or 3%.

Benefit and return. With a cost of $h$ dollars per hour of work:

With $V = 2{,}000$, $m = 6$, and $h = 10$: $B_1 = 24{,}000$, $R_1 = 12{,}000$, and $\text{ROI}_1 = 100\%$. The upfront investment is paid back in

Break-even points. Solving $R_1 = 0$:

That is why the pilot’s goal is “at least three minutes”: below that, the first year loses money. With the four minutes measured, $R_1 = 16{,}000 - 12{,}000 = 4{,}000$.

What this calculation leaves out

It is a first-order model. It does not discount money over time, count exit costs (switching providers), include the residual risk from lesson 35 (what an incident costs times its likelihood), or allow for token prices changing. Nor does time saved turn into money on its own: if nobody uses those hours for something else, the benefit never shows up in the books. That is why the pilot measures, and the calculation is redone with what was measured.

The two-by-two grid

It combines three factors on two axes: the horizontal one is $f \times v$ (the value at stake) and the vertical one is the risk $P \times I$ from lesson 35. The positions in the lesson are ordinal (order matters, distance does not). An alternative is a score, for example $\text{priority} = f \cdot v \cdot (1 - r)$ with $r \in [0, 1]$, but a single number hides why a case ended up high or low; the grid forces you to state both reasons. A high-risk case can come down with controls (human approval, answers with sources, minimal permissions) and move to another zone, like the warranty questions.

In the European Union, screening job applications is in Annex III of the AI Act (high risk); under Regulation (EU) 2026/1744, those obligations apply from December 2, 2027.

Build, buy, or fine-tune

Path	What it is	When
Buy	a ready-made product, paid per month or per user	common tasks: transcribing, drafting, summarizing
Build	your system on a provider’s model: prompt, RAG, tools	tasks that use your data or your processes
Fine-tune	on top of that, fine-tuning or LoRA with your examples (lesson 27)	a stable, highly repeated task with thousands of good examples, if the prompt and RAG fall short
From scratch	pretraining your own model (lessons 23 and 24)	almost never: tens of millions of dollars

Before buying or building, evaluate with your gold cases (lesson 34): compare the options on the same set and the same total cost.

What the studies say

• Brynjolfsson, Li, and Raymond (2025), Quarterly Journal of Economics: 5,172 customer support agents, with an AI assistant introduced in stages. Issues resolved per hour rose 15% on average and 30% among the less skilled and less experienced workers; among the most experienced, the productivity effect was small and their quality dipped slightly. (The 2023 working paper reported 14% and 34%.)
• Dell’Acqua and colleagues (2023), Harvard Business School and BCG: 758 consultants in a preregistered experiment. On 18 tasks within GPT-4’s reach, those who used it completed 12.2% more tasks, 25.1% faster, and with higher quality; on a task chosen to fall outside its reach, they were 19% less likely to reach the correct solution than those who did not use it. The authors call this a “jagged technological frontier.”
• Microsoft and LinkedIn (Work Trend Index, May 8, 2024): a survey of 31,000 people in 31 countries. 75% of knowledge workers were using generative AI at work; 78% of those users were bringing their own tools; only 39% had received training from their company.
• Gartner (July 29, 2024): predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, due to poor data quality, inadequate risk controls, escalating costs, or unclear business value. It is a consulting firm’s prediction, not a measurement.

In code

HOURLY = 10                          # dollars per hour of work (assumption)
TASK_COST = (3500 * 3 + 300 * 15) / 1e6   # dollars per question (lesson 26)
UPFRONT, FIXED = 6000, 470           # once; every month (assumptions)

def first_year(questions, minutes):
    benefit = 12 * questions * minutes / 60 * HOURLY
    total_cost = UPFRONT + 12 * (FIXED + questions * TASK_COST)
    return round(benefit), round(total_cost), round(benefit - total_cost)

print(TASK_COST)
for questions, minutes in [(2000, 6), (500, 6), (1000, 6), (2000, 4)]:
    print(questions, minutes, first_year(questions, minutes))

net_saving = 6 / 60 * HOURLY - TASK_COST          # dollars per question, after tokens
print(round((UPFRONT / 12 + FIXED) / net_saving))  # questions a month to break even
print(UPFRONT / (2000 * net_saving - FIXED))       # months to pay back the investment

Output:

0.015
2000 6 (24000, 12000, 12000)
500 6 (6000, 11730, -5730)
1000 6 (12000, 11820, 180)
2000 4 (16000, 12000, 4000)
985
4.0

What comes next

Lesson 37 brings the whole course together on a map: from the neuron to the chatbot, and from there to the questions worth asking any AI team or vendor.