What businesses
actually get from AI

Last post I argued that AI is not intelligent yet. Some of you read that and concluded "great, I can ignore it for another year." Wrong conclusion.

The right conclusion is the opposite: precisely because the models aren't going to wake up next Thursday and run your business for you, the value left on the table is enormous, and almost all of it is operational rather than visionary. You don't need superintelligence to delete forty hours of copy-paste a week. You need a competent operator and the patience to wire things up properly.

Here's the practical map.

The hyperscaler reality (or: why you don't own this)

Before we talk value, let's talk who's actually selling it. Today's frontier models live on stacks of Nvidia H100 and B200 GPUs that retail for around USD $25–40k each, get bolted into eight-GPU servers that run six figures, and get racked in data centres that cost more than small countries' GDPs. The Stargate joint venture announced in early 2025 alone earmarked USD $500 billion for AI infrastructure. Microsoft, Google, Meta and Amazon collectively committed somewhere north of USD $300 billion in AI capex for 2025 alone.

You are not going to outspend them. Almost no one is. For the next one to three years, "doing AI" for 99% of businesses means renting frontier intelligence over an API.

That subscription model has consequences, and ignoring them is how you wake up two years from now realising your competitive moat got pasted into a training set.

The data-leakage tax

Every prompt you send to ChatGPT, Claude, Gemini, or any hosted model is a packet of data crossing a border into a data centre, usually American, owned by a company whose entire business model depends on getting smarter. Even when "training on your data" is officially off (and it is, on most enterprise tiers, see OpenAI's enterprise privacy policy and Anthropic's commercial terms), your queries shape product priorities, surface evaluation data, and exist as logs in someone else's jurisdiction.

When you give an agent control of your computer, this gets worse, not better. The whole point of an agent is to read whatever it needs to read to complete the task. That now includes your CRM, your shared drive, and that one folder labelled do_not_open_legal. LLMs were already a porous data boundary. Agents are a sieve.

Rule of thumb: assume any data your agent touches has, in some form, been observed by the platform. Build accordingly.

For most businesses this is an acceptable cost, the upside dwarfs the leakage. For a handful (regulated industries, IP-heavy R&D, anyone with genuinely defensible proprietary data), it's a five-alarm problem and the answer is local-first scaffolding: open-weight models, your hardware, your network, your rules. That's a different conversation. Have it with me if it applies to you.

The compute-cost glide path

Here's a data point worth tucking away: compute is getting roughly an order of magnitude cheaper every two to three years, before you even count algorithmic gains. Independent research (see Epoch AI's tracking of algorithmic progress in language models) shows the compute required to reach a given capability halves every eight months or so on top of hardware gains.

Compounded across hardware and software, a credible 100x improvement in cost-per-capability over the next three to five years is not science fiction. The training run that costs a hyperscaler $2M today plausibly fits on a $20k box in 2030.

What does that mean for you? Don't build your strategy on today's prices. Things you write off as "too expensive to AI-ify" right now will be trivial in 36 months. Things that look like permanent moats may not be.

Value Stream 01 · Custom software, suddenly affordable

This is the one that quietly changes everything, and almost nobody is talking about it correctly.

Until recently, building a piece of bespoke software for a specific quirk of your business (the weird invoice-reconciliation thing, the odd report your operations manager assembles every Monday, the integration between two SaaS products that almost-but-not-quite work together) meant hiring a developer at $100k+ a year, or paying an agency $30k for something fragile. So you didn't. You bought a SaaS subscription instead, and added it to the pile.

That calculus is broken. With a coding agent like Claude Code or OpenAI Codex, a competent operator with no formal CS background can ship working internal tools in days. Andrej Karpathy called this "vibe coding", describing what you want, watching the agent build it, course-correcting when it goes off the rails. It is genuinely a new mode of software production.

Two corollaries that founders consistently miss:

The deterministic discipline

Here is where I diverge from the breathless takes. Once your custom workflow exists, do not run it through an LLM every time.

If your inputs arrive in the same shape every day (a CSV at 9am, an invoice on the same template, an API response with a known schema), the right answer is plain old code. Hardcode it. If this, then that. Local CPU cycles cost effectively nothing; an agent run costs cents-to-dollars and produces slightly different output every time because the underlying model is, by design, probabilistic.

People complain about "context rot", agents losing the plot at the back end of a long task, and I have sympathy for them, because they're using the wrong tool. LLMs build the software. Deterministic code runs the software. When something breaks, you bring the agent back in to fix it.

This is the single biggest cost mistake I see businesses make: spinning up a frontier-model agent to do a job a Python script could do for free. It's the AI equivalent of taking a helicopter to the corner shop.

Value Stream 02 · Discovery, research, and irregular data

The other half of the value lives at the opposite end of the spectrum: situations where the input isn't uniform.

A résumé arrives in a different format every time. A customer email goes off on three tangents before getting to the point. A folder of contracts spans fifteen years of formatting drift. Probabilistic input is exactly where probabilistic models earn their pay. An agent can read that mess and turn it into structured data your deterministic pipelines can then digest cleanly. You're using the LLM as a translator into uniform shape, not as a worker on the uniform shape itself.

One layer up from that, and this is the part that surprises clients, is discovery. Pointing an agent at a business and saying "find me five processes that are wasting time" is not magic. It's directed inquiry. A capable operator who knows the business can lead an agent through the codebase, the SaaS stack, the meeting notes, the customer support tickets, and surface patterns that would take a human consultant a fortnight.

Note the word directed. The agent does not do this on its own. Refer back to post one: untethered, the model produces confident output regardless of whether the question made sense. Tethered to a human who knows what they're looking for, it's a force multiplier of the highest order.

The bottleneck, again

I'll keep saying it because it's the only thing that matters: the limiting factor is the human pilot.

Every other component of the AI stack is improving rapidly and getting cheaper. The chips will change. The cooling will change. The model architectures will change. The agentic harnesses will change, Codex, Claude Code, whatever ships next quarter. The languages will probably change. None of that touches the bottleneck.

What does touch the bottleneck is having at least one person in your organisation, internal hire, external partner, doesn't matter, who can:

• Tell deterministic problems from probabilistic ones, and route accordingly.
• Read AI output critically enough to catch when it's confidently wrong.
• Sequence the work, build the small obvious win first, prove the pattern, expand.
• Build local scaffolding when the data sensitivity demands it, and call frontier APIs when it doesn't.
• Keep the bill rational. (See "helicopter to the corner shop", above.)

If that person already exists in your team, brilliant, get out of their way and give them budget. If they don't, you have two options: hire and train one, or rent one for a focused engagement until you do.

What to do this week

Three concrete moves while the kettle boils:

• Audit your SaaS bill. Circle the line items that do something narrow. Those are candidates for in-housing once you have a pilot in place.
• Pick one repeating, deterministic task in your week (the Monday report, the daily reconciliation, whatever it is). That's your first build target.
• Pick one messy, irregular data flow (résumés, support tickets, vendor invoices). That's your first agent target.

If you want a sharper version of this list specific to your business, that's literally what my two-hour proving ground is for. Walk in with a problem; walk out with something working. No retainer, no PowerPoint, no charge for the chat.