Two tabs open on a Sunday night. One is a vendor quote for an AI support platform that charges $0.99 for every customer ticket it resolves. The other is a job listing for a machine learning engineer, a role the Bureau of Labor Statistics prices at a median of $133,080 a year before you add stock, benefits, or the second engineer you will inevitably need. Close the wrong tab and you’ll pay for it for two years.

Most companies have already voted, and they voted buy. In 2025, enterprises purchased 76 percent of their AI use cases rather than building them in-house, up from 53 percent the year before, according to Menlo Ventures. Total enterprise spending on generative AI hit $37 billion that year, more than triple 2024. But the headline number hides the useful part. The companies getting this right never picked a side. They decide capability by capability, and they run the same three tests every time.

When buying wins

Buy when the thing you need is a commodity. If a hundred other companies need the exact same capability, answering password-reset tickets, transcribing sales calls, digging totals out of a pile of invoices, then somewhere a vendor has spent years and millions polishing exactly that, and spreads the cost across all hundred of you. You don’t build a power plant to run a toaster. Intercom’s Fin, for instance, charges $0.99 per resolved conversation. If your support desk sees 2,000 tickets a month and the AI handles half, that’s roughly $1,000 a month, live within a week or two. No hiring, no on-call rotation, no maintenance.

Speed is the second reason, and it’s usually undercounted. A vendor gets you to production in weeks. An internal build takes quarters, and the market doesn’t wait while you interview candidates. If a competitor ships the capability first, the money you saved building may cost you the category.

And then there’s compliance, the quietest argument for buying. A serious vendor already holds SOC 2 certification (an audited security standard), handles data-handling agreements, and employs people whose whole job is passing those audits. That burden lands on someone else’s payroll. It only counts, though, if you actually check. There’s a separate piece here on what to ask an AI vendor about security before you believe the badge on their homepage.

When building wins

The first case is blunt. If the AI is the product, the thing customers actually pay for, you cannot outsource it. A company whose entire value is a subscription to someone else’s intelligence isn’t a product company. It’s a reseller with worse margins and no control over its own roadmap. Investors have learned to check this in the first ten minutes of diligence.

The second case is data. If you sit on something proprietary, ten years of insurance claims decisions, a million labeled pathology slides, whatever your competitors can’t buy, then a model shaped by that data can do things no off-the-shelf product will ever match. And it compounds. Every new customer generates more data, which improves the model, which wins more customers. A vendor’s product improves for all of its customers at once, including the two competing with you.

The third case is arithmetic, and it deserves actual numbers. Say a resolved support conversation burns about 10,000 tokens (tokens are the word-fragments AI models bill by, and 10,000 is a generous estimate for a few back-and-forth messages). On a small model like GPT-5.4-mini, priced at $0.75 per million input tokens and $4.50 per million output, that conversation costs about a penny and a half in raw model fees. The platform charges $0.99 for the same outcome. At 2,000 resolutions a month, the 66x markup is irrelevant, you’re paying for the years of engineering wrapped around the model. At 100,000 resolutions a month, you’d be sending the vendor $99,000 monthly for roughly $1,500 of tokens plus work two engineers could maintain for about $22,000 a month in salary. The crossover in that example sits somewhere between 25,000 and 30,000 resolutions a month. Below it, buying is obviously right. Above it, every month of delay is a five-figure donation. The full cost side of that equation, salaries, evaluation, infrastructure, gets its own treatment in how much it costs to build an AI product in 2026.

The trap in the middle

Here’s where most of the wasted money actually goes. Not into bad builds or bad purchases, but into the hybrid nobody plans for: buying a platform, then spending build-level money customizing it.

It starts innocently. The platform covers 80 percent of what you need. So one engineer spends a sprint on custom logic inside the vendor’s workflow builder. Then integrations. Then a consultant. Eighteen months later you’ve spent $400,000 in salaries and fees constructing something intricate inside a product you rent. It’s remodeling a rented kitchen. The granite countertops are gorgeous, and they belong to the landlord.

The damage isn’t just the money. Everything you built lives in the vendor’s proprietary format, which means you’ve raised your own cost of leaving, which means the vendor knows it at renewal time. Price goes up 30 percent? You’ll grumble and sign, because the alternative is rebuilding eighteen months of work.

So do the switching-cost math before you sign anything, not after. Four numbers, added together: what it costs to rebuild your customizations somewhere else, what data migration costs, what retraining your team costs, and what it costs to run the old and new systems in parallel for the two or three months a cutover really takes. Then sort your assets into portable and stuck. Prompts and evaluation sets are usually portable, and so is your own data. Workflows drawn in a vendor’s visual builder are usually stuck. If the stuck pile is projected to exceed a quarter’s engineering budget, either negotiate exit terms into the contract or reconsider. A deeper comparison of the platforms themselves lives in AI agent platforms compared, before you commit.

What happened to Jasper

One more risk, and it cuts both builders and buyers. In October 2022, Jasper raised $125 million at a $1.5 billion valuation selling AI copywriting, marketing text generated by GPT-3 with a nice interface on top. About six weeks later, on November 30, OpenAI released ChatGPT. Free, and built on the same family of models Jasper was paying to access. By July 2023 Jasper was laying off staff, had cut its revenue forecast by at least 30 percent, and later reduced its internal valuation by roughly 20 percent.

That’s wrapper risk. When your product is a thin layer over someone else’s model, the model provider’s roadmap is your risk register, and every one of their launch events might delete a feature you charge for. Jasper survived by burrowing deeper into marketing team workflows, the part OpenAI had no interest in owning. The general lesson: capability rented from a model provider is not a moat, because the provider can hand it to everyone tomorrow. Workflow depth, distribution, and proprietary data are moats. This is the same reason narrow, focused AI agents keep beating general ones in regulated work: the value sits in the specifics, not the raw intelligence.

And notice the buyer’s version of the same risk. If the platform you’re about to purchase is itself a thin wrapper, its mortality becomes your migration project.

Four questions before you decide

  • Is this capability a commodity that a hundred other companies need in identical form? Buy it.
  • Is the intelligence itself what your customers are paying for? Build it, or admit you’re a reseller.
  • At your realistic 24-month volume, do per-unit vendor fees cross the cost of two engineers plus tokens? Run the actual numbers, not the vendor’s ROI calculator.
  • If the model provider shipped this feature natively next quarter, what would be left of the product? Whatever remains, data, distribution, workflow, is the only part worth building.

Buy to learn what you need. Build the thing you’d have to defend.