95% of organizations are getting zero return on their generative AI investments, according to MIT's review of enterprise AI projects — and marketing teams are still rushing to build automation in-house rather than buy it.

That finding sits inside an estimated $30 billion to $40 billion in enterprise genAI spending. It anchors a build-versus-buy playbook from Kevin and Amanda, the co-authors behind the Growth Memo newsletter, writing on Search Engine Land.

Where the AI work actually comes from

The pair lay out three sourcing options for any AI workflow:

• Vendor-built AI software or tooling
• Customization and implementation from a consultant who has already shipped the same thing
• Internal team time to build it in-house

"Teams aren't sorting their AI work hours in a targeted manner," they wrote. "Instead, they're going straight to automating as much as they can in-house."

The authors flag that pattern as expensive, not clever. Skipping a sorting exercise that takes 30 minutes or less, they argue, is the easy fix that marketing teams keep missing.

When to buy the tool

Common problems — rank tracking, citation monitoring, brand mention tracking, crawl diagnostics, content scoring — belong on a vendor's shelf. MIT's data backs the call: bought AI tools reached deployment twice as often as in-house builds in the sample studied.

The MIT report named the reason projects fail: "the tools do not learn or integrate with how people already work." A vendor who has fitted that workflow for 4,000 customers has already absorbed the integration cost.

"Buy the tool when the problem is common," the authors wrote. "Your team also gets two things you can't build for yourself: the vendor maintains the software, and somebody on the other end fixes it when it breaks."

When to buy the know-how

Some workflows genuinely belong to one team: who signs off on work before it ships, how data is structured, when reporting runs, how SME input is gathered. The temptation is to build those internally because nobody else has them. The authors reject that framing.

"The workflow is yours, yes, but someone who has built 65 of these for 20 different teams already knows which steps usually break, which ones need input from your individual contributors, which ones are worth automating, and which ones look automatable but aren't," they wrote.

Their test: if the hours about to be spent produce knowledge the team will not need weekly, contract out the knowledge instead.

The verification problem nobody budgets for

Only 13% of marketers fully trust AI output without a human reading it, according to data the authors cite. The authors reframe that figure: it is a staffing requirement, not an adoption hurdle.

The numbers get worse at the top. Validity's State of CRM Data Report 2026 found that 78% of C-suite respondents and 92% of SVP/VP-level respondents have acted on an AI recommendation they later suspected was wrong because of bad underlying data. Among individual contributors, that figure was 41%.

The rule the authors draw: never buy or build a tool for a job nobody on the team can verify for accuracy by hand. "The tool did not remove the skill required to do the work," they wrote. "Judging the work is the harder, less-replicable of the tasks."

Why one-step swaps beat full-job automation

The most actionable section is a framework for choosing what to automate. A job is a bundle of steps with judgment distributed across all of them. A step has one input, one output, and a check that takes seconds.

Teams that name a slow process and hand the entire thing to AI often end up "with in-house software with no real tests, no documentation, and one person who understands how to maintain it," the authors warned. A one-step swap has one input to validate and one thing to fix when the model changes. A 15-step workflow has 15 places to break.

Steps worth automating share four traits: slow, repetitive, tightly defined, and checkable at a glance. Anything requiring experienced taste, or anything where being wrong could go unnoticed for three months, should stay with a human.

What marketing leaders should do next

The MIT failure rate and the Validity numbers on bad AI recommendations land at the same place. AI does not eliminate the skill required to judge its output; it relocates that skill. Marketing leaders planning AI budgets for the rest of the year will measure ROI differently once they accept that judgment sits upstream of the model, not downstream of it.