Optimize Your Fund's AI Spend

Most funds now spend real money on AI, and almost none of them can say what it buys. This is a short course, read in order, that fixes both halves of that problem: it teaches you how AI is actually billed, the handful of levers that cut the bill without cutting quality, and how to tell whether any of it is paying off. By the end you will be able to read an AI invoice, question any “AI for your fund” purchase, and set a spending policy that keeps cost tied to value.

It is the companion to Learn the 80x Method, which teaches how to build fund software. This one teaches how to pay for it well.

Who this course is for

Partners, principals, and operations leads who approve the AI budget, whether or not they write any code. No engineering background is assumed. Every term is defined in plain English the first time it appears, and no chapter expects you to have opened a terminal. If you already build with AI, you lose nothing but a few definitions you already knew.

The one reframe this course rests on

An AI bill is not a subscription. It is token volume multiplied by price, summed over every call your fund makes, and it arrives on three separate meters:

1. Model usage, billed by the token (the small unit of text a model reads and writes). This is the meter most people forget exists.
2. AI software seats, billed per person per month, and increasingly per person plus usage on top.
3. Engineering and human time to build, run, and check the work. This never appears on any invoice, and it is usually the largest cost of all.

The unit that makes all three legible is cost per fund workflow: what it costs to screen one deal, capture one meeting note, refresh one portfolio update, or prep one LP conversation. Hold that unit in your head and the rest of the course is five ways to lower it and one way to prove the lowering was worth it.

The five levers, and the ROI wrapper

Four short lessons on what you are buying, five on how to spend less for the same result, and two on how to know it worked.

1. 01What AI actually costs your fundHow AI is metered — per token, across three cost centers — and why the same task can cost 100x more depending on how you run it.8 min
2. 02Right-size the model to the taskThe biggest structural saving: route each job to the cheapest model that clears its quality bar, and escalate only the hard few.8 min
3. 03Cache the fixed part of every promptWhen you send the same thesis or rubric on every call, caching re-reads it at roughly a tenth of the price.7 min
4. 04Batch the work nobody is waiting forOvernight sweeps and bulk enrichment do not need instant answers, so a batch queue halves that line item.6 min
5. 05Shrink the input: retrieve, don’t stuffA million-token context is billed as a million tokens every call. Retrieve the few relevant pages instead.8 min
6. 06Put a ceiling on the agent loopAn autonomous agent can quietly spend 20x a single call. Hard caps and budgets keep one runaway job from eating the month.8 min
7. 07Measure ROI: hours returned vs true costMost AI pilots show no measurable return. The ones that pay off set a baseline, count full cost, and measure lift.9 min
8. 08Build your fund’s AI budget and policyTurn the levers into a one-page plan: audit seats, decide buy-versus-build, set enforced caps, review monthly.9 min

Prices and model names in this course were verified in mid-2026 and move quickly. Where a figure is a worked example rather than a published rate, it says so. The point is never the exact number; it is the mechanism, which lasts.

Prefer it in your inbox?

Get the 5-day email course

One short lesson each morning for five days — the whole course, in your inbox. No spam; unsubscribe in a click, anytime.

Start the course

See also

• Learn the 80x Method — the companion course on building the systems this one teaches you to pay for
• Context engineering — the reference page behind Chapter 5, on deciding what a model reads
• Automation safety — the discipline behind Chapter 6, on making a runaway job fail small