Build or buy your AI automation? How to decide
For most operators, buying a focused AI tool beats building one from scratch, but not for the reason vendors give. MIT's 2025 study found purchased and partnered tools reached production about 67% of the time and internal builds about a third as often. The variable that actually separated the winners from the 95% of pilots that stalled was not model quality or who wrote the code, but whether the tool was embedded in a real workflow, learned from feedback, and had an owner. So the real decision is embedded-and-owned versus generic-and-orphaned, pointed at the back office instead of the customer-facing demo.
For most operators, buying a focused AI tool beats building one from scratch. But not for the reason vendors give. MIT's 2025 study of corporate AI found that purchased tools and vendor partnerships reached production about 67% of the time, while internal builds succeeded about a third as often. The variable that actually separated those winners from the 95% of pilots that delivered no measurable return was not model quality, and it was not who wrote the code. It was whether the tool was embedded in a real workflow, learned from feedback, and had someone responsible for it. That reframes the whole question. You are not really choosing between build and buy. You are choosing between embedded-and-owned and generic-and-orphaned.
What the MIT data actually says, and what people misread
In August 2025, MIT's NANDA initiative published "The GenAI Divide: State of AI in Business 2025," based on more than 300 public AI deployments plus interviews and surveys of company leaders. The headline number that spread everywhere was that 95% of organizations saw no measurable impact on profit and loss from their generative AI pilots, despite an estimated 30 to 40 billion dollars poured in.
That number got misread immediately. It does not say AI does not work. It says 95% of the way companies deployed it could not pay off. The 5% that did succeed were narrow and embedded in a single process. The lead researcher put the build instinct plainly: "Almost everywhere we went, enterprises were trying to build their own tool," and the data showed the bought tools delivered more reliably.
The reason matters more than the percentage. The report pins the failures on a learning gap, not a model gap. Tools failed because they did not retain feedback, did not adapt to how the work was actually done, and had no memory of prior runs. A generic assistant is great for one person drafting an email and useless as a workflow because it forgets everything the moment the tab closes. That is the same failure whether you bought the generic tool or built it.
Build vs buy is the wrong axis. Embedded vs orphaned is the right one
Here is the part the standard build-vs-buy guides miss. They sort the decision on cost and control: build for control and lower long-run cost, buy for speed and less maintenance. Useful, but it is measuring the wrong thing. The MIT data did not find that buying is magic. It found that bought tools were more often the ones that ended up embedded in a specific workflow with a clear owner, because a vendor sells a finished thing aimed at one job. Internal builds more often became broad, half-configured platforms that nobody owned end to end.
Flip the cases and the advantage flips with them. A generic chatbot you bought, dropped on the website, and never trained on your actual process will stall exactly like a failed internal build. A narrow automation you wired yourself, that reads a real queue, takes corrections, and has a named owner, will outperform a polished tool sitting outside the work. The thing that predicts success is not the purchase order. It is three properties:
- Embedded. It lives inside one real workflow and touches the real system of record, not a demo or a side sandbox.
- Learns. There is a feedback loop. When it gets something wrong, someone corrects it and the correction sticks, through examples, rules, or a record it reads next time.
- Owned. One person is responsible for whether it works this week. Not a committee, not "IT," not the vendor.
Score any AI project on those three before you score it on build or buy. A bought tool that fails all three is a worse bet than a small build that passes all three.
A decision rule you can actually use
Once you sort on embedded, learns, and owned, the build-or-buy call gets simple. It comes down to whether a finished tool already covers your exact workflow.
| Your situation | Choose | The failure mode it avoids |
|---|---|---|
| A vendor tool covers your workflow end to end, in your stack | Buy that focused tool | A half-built internal platform nobody finishes |
| No tool fits, but the workflow is well understood | Build narrow in the runner you already use | A broad "AI platform" purchase that goes unconfigured |
| Workflow spans your own database and custom apps | Build, or partner to build | A generic SaaS bot that cannot reach your data |
| You want speed and have no one to own it | Buy, and assign an owner anyway | An orphaned tool that drifts and gets switched off |
| The job is broad, vague, and company-wide | Do not start here | The exact pilot that lands in the failed 95% |
The bottom row is the one operators ignore. The instinct after reading about AI is to buy or build something big and horizontal. The data says start narrow. If you cannot name the single workflow and the single owner, you are not ready to build or buy yet. You are ready to scope.
Point it at the back office, not the sales demo
The MIT report found that more than half of generative AI budgets went to sales and marketing, the visible, fun, board-deck functions. The measurable returns showed up somewhere else: back-office work like cutting external agency and outsourcing spend, clearing document and data drudgery, and streamlining internal operations. The money chased the demo. The payoff sat in the boring queue nobody wanted to show off.
For an operator this is good news, because the boring queue is where you have the most room and the cleanest data. Invoice intake, order status updates, dispatch and scheduling, permit and document checks, lead routing. These are high-volume, repetitive, and measurable, which means a narrow automation has somewhere to learn and an obvious before-and-after. If you are deciding where to aim a first build, the same instinct applies as picking the first workflows worth automating: start where the work is repetitive, structured, and already painful, not where it would look impressive.
What "owned" looks like in practice
Owned is the property people skip, and it is the one that quietly kills tools six weeks after launch. A tool with no owner does not get corrected when it drifts, so people stop trusting it, so they route around it, so it gets switched off. Ownership means a specific person can see what the automation did, change it without filing a ticket with a vendor, and is on the hook for whether it ran today.
That is also the honest line between buying a closed product and building or partnering on a system you control. When we build automation systems, the deliverable is not just the working flow, it is the ownership: a runbook, a feedback loop, and run records the operator can read, so the people doing the work can see and correct what the system did. Field to Flow is one example, an operations system embedded in the actual field workflow rather than a generic tool bolted on beside it. That is the shape the MIT data rewards, regardless of whether you bought it, built it, or built it with a partner. The thing to avoid is the closed tool that only the vendor can change and only the vendor understands.
If you do choose to build, the runner you pick is a smaller decision than it feels, and it follows the same embedded logic. We have walked through how to choose between n8n, Zapier, and Make by where your data of record lives. Pick the one closest to your real system, then keep the build narrow.
How to decide this week
Write down the candidate AI projects you are weighing. For each one, answer three questions before anything about cost or vendor. Can it live inside one real workflow that touches your real data? Is there a way for it to learn from its mistakes? Will one named person own whether it works? Drop anything that fails two of the three. It will join the 95%.
From what survives, pick the most boring, highest-volume, back-office workflow on the list. If a finished tool covers it end to end in your stack, buy that one and assign the owner. If nothing fits, build it narrow in the runner you already use. Either way, the win is the same: a tool inside the work, learning from it, with a person responsible. If you want a second read on which of your workflows is the right first target and whether to buy, build, or partner on it, tell us what the workflow is and we will tell you straight which one it is.
Frequently Asked Questions
SOURCES & CITATIONS
- MIT report: 95% of generative AI pilots at companies are failing — Fortunehttps://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- The GenAI Divide: State of AI in Business 2025 — MIT NANDAhttps://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
- An MIT report finding 95% of AI pilots fail spooked investors. It should have spooked the C-suite instead. — Fortunehttps://fortune.com/2025/08/21/an-mit-report-that-95-of-ai-pilots-fail-spooked-investors-but-the-reason-why-those-pilots-failed-is-what-should-make-the-c-suite-anxious/
About Alexey Yushkin
Alexey is the founder of GENERAL INFORMATICS LLC. He designs and ships AI and automation systems for businesses and operators across the US.
Related reading
Want this kind of system in your business?
We build practical AI and automation systems for operators. Send us your current workflow and we will show you what to automate first.
Request a Workflow Review