Every AI initiative we've shipped this year sorts into one of four boxes. (I know what you're thinking... A former consultant busting out a 2x2 matrix? How unexpected!)

Two questions guide the framework.

1. Did the spend exist before?
2. Is AI the engineer or the employee?

The top boxes are familiar: If a vendor already sells it, AI can build you your own version in weeks. If someone's hired to do it, an agent does the work and a person reviews the queue. That's the leaner-company stuff, and it's worth doing.

The bottom boxes didn't exist three years ago: Software no vendor would ever build because you'd be the only customer. Work nobody ever staffed because the headcount would be absurd. (e.g., We parsed 23,000 hours of public-meeting audio for one client). Nobody was ever going to hire for that.

01 Build, Don't Buy

Your own software, in weeks

The vendor invoice you've paid for years, replaced by software built to fit your business exactly.

• Rebuilt ~90% of a market-leading real-estate data platform in two weeks for a multifamily investor
• Replaced a 12-app workflow with one platform for a national expert network
• Unlimited geocoding on a $42/mo server instead of per-call API pricing

The test: You can name the vendor you'd stop paying.

02 Digital Employees

Agents doing hired work

Jobs on the org chart, handled by agents. Workflows that needed a person now need a review queue.

• AI voice screening that interviews and qualifies inbound leads before a human ever joins the call
• Quotes drafted automatically from inbound RFQs for an industrial distributor
• Proposals drafted from discovery-call recordings for a marketing agency

The test: There's a job description for this work.

03 Markets of One

Software no vendor would build

Built around your exact ICP and the way your team actually works. The market was one company, so no product ever existed.

• An ICP model and territory design built around how one brokerage actually splits a market, their filters, their map
• A deal-sourcing engine tuned to one search fund's exact thesis, screens no off-the-shelf tool offers
• A market-expansion model built on the 20 filters one startup team actually uses to pick cities

The test: You'd be the only customer.

04 The Million-Person Team

An army you could never hire

Tasks a person could do once, done a million times. Headcount nobody ever budgeted, because the number was absurd.

• 23,000 hours of public-meeting audio parsed for expansion signals for a service provider
• Every county appraisal record ingested, classified, and scored for a CRE brokerage
• ~250 product-demo recordings distilled into a ranked bug register in a day

The test: One person could do one unit. You'd need a million people.

The top row makes you leaner. The bottom row makes you different.

What we've learned

1. Match the model to the box. We build with frontier models and run automations on open-weight models on our own hardware. You want a brilliant engineer and a nearly free army. A million-person team only pencils when each worker costs almost nothing. Pay frontier prices at runtime and the math dies.

Left Column = Frontier Intelligence / Right Column = Open Weight model.

2. Cost savings is the smaller prize. The value shows up when you grow 10-100x and the costs don't come with you. And in our experience, that growth comes from the bottom row. The top row just means you don't have to hire your way through it.

Real examples from client work in the graphic, anonymized.

Which box is your next initiative in?

Read the original post on LinkedIn.