Build vs. Buy for AI Tooling: A Decision Framework for Growing Companies

Apr 27
5 min read

By Jonathon Carlson | Atlas Thread Digital

For the last two years, every executive I've talked to has been asking the same question: "Should we be using AI?" That ship has sailed. Your competitors are using it, your customers expect it, and your board has already made up its mind. The interesting question, the one that actually determines whether you'll get value or burn cash for eighteen months, is the next one: should you build it, buy it, or partner for it?

Most companies skip past this fork. They sign up for a vendor's "AI suite," tack it onto an existing SaaS contract, and assume that's the strategy. Or they go the other direction and hire two ML engineers to build a chatbot in-house. Both paths have their place. What's striking is how rarely either path is chosen on purpose.

The cost of getting this wrong is now public. MIT's NANDA initiative published its State of AI in Business report in 2025 and found that 95% of enterprise generative AI pilots produce no measurable ROI. The same study found that buying from specialized vendors or partnering on AI succeeds about 67% of the time, while pure internal builds succeed at roughly half that rate. The model isn't the problem. The procurement decision is.

Three Options, Not Two

Build vs. buy is the framing every executive learns in business school, and it doesn't quite fit AI. The real menu has three items: build, buy, and partner. Most companies use all three, often without realizing it.

Buy means licensing a finished product from a vendor like Salesforce Einstein, ServiceNow's AI agents, or Microsoft's Copilot suite. You configure rather than code. Time to value is fast, usually weeks. Enterprise AI software licensing typically lands around $30,000 to $50,000 per user per year at production scale, which adds up faster than most CFOs expect.

Build means standing up your own ML or agent team and developing custom systems against your own data. A small to mid-sized AI team will run $500,000 to $1.5 million annually once you account for senior engineers, MLOps, and the supporting infrastructure. Time to value is months at minimum, sometimes a year or more.

Partner is the underrated middle path. You bring in a specialist firm or consultancy to build something custom against your data and workflows, then you maintain it (or hand it back to them to maintain). A typical partner engagement runs $150,000 to $500,000 and delivers in weeks to a few months. This is the option most decks ignore, which is curious because it's the one that produces the highest success rates in the field.

The interesting market signal: a recent Kellton survey found that 47% of enterprises are already running a hybrid mix, with only 21% on pure pre-built agents and 20% on fully custom builds. The hybrid majority didn't get there by design. Most stumbled into it. They bought one vendor for service tickets, built one agent for a niche workflow, and now they have two systems with no shared governance and no shared logs. That's not a strategy. That's an accident with a recurring invoice.

A Framework That Won't Lie to You

The decision comes down to four questions, and you have to answer them honestly.

Is this a competitive differentiator or a commodity? If the capability shows up in every vendor's roadmap, things like meeting summarization, basic chatbots, document classification, buy it. You will not out-innovate Microsoft on Outlook summaries. If the capability is genuinely tied to how you make money, that's where building or partnering earns its keep.

How much of your value is in the data? Generic models trained on the open web are good at generic tasks. They are not good at the contractual quirks of your supplier relationships, the historical patterns in your claims data, or the institutional knowledge buried in your service tickets. The more of your value lives in proprietary data, the more you need a system that can actually use it. Vendor APIs rarely give you that depth.

What's your honest internal capability? Building requires not just one good engineer but a team with ML, MLOps, and domain expertise. If you don't already have that bench you aren't building, you're hiring, and the timeline doubles. The companies that succeed at internal builds almost always already have the muscle. The ones that succeed at partnering don't, and they know it.

What does the five-year picture look like? SaaS pricing looks great in year one and ugly in year five. The break-even point between custom and SaaS is typically two to four years on total cost of ownership, and that math gets worse for buyers when the vendor raises prices, gets acquired, or sunsets the feature you depended on. Vendor lock-in is rarely visible at signing. It becomes very visible when you try to leave.

Run those four questions against any AI initiative on your roadmap and the right answer usually presents itself. The problem is that most companies don't run them. They default to "buy" because it feels safer or "build" because it feels more sophisticated, and neither default holds up under scrutiny.

What This Means for Smaller Companies

For Fortune 500 firms, all three paths are open. For a 50-to-1,000-person company, which is most of the work I do at Atlas Thread, the math is tighter. Industry benchmarks put mid-sized AI investment at $100,000 to $500,000 a year, and that ceiling forces real choices. You can't afford a six-engineer ML team. You probably also can't afford to license enterprise AI for every employee at $40,000 a head.

The pattern I've watched succeed is portfolio-style. Buy the commodity layer like writing assistants, meeting tools, and basic ticket classification. Partner for the one or two systems that touch your actual product or your proprietary data. Avoid building anything yourself unless you've already proven you can. Most of the engagements I take on at Atlas Thread fall into the partner bucket: a custom MCP server, a domain-specific agent, an integration layer between an old database and a new AI workflow. Those are the projects where off-the-shelf tools genuinely can't reach, and where a four-person internal team can't justify the hiring.

The companies that get this right tend to have one shared trait. They decided what category each AI initiative belonged to before they started buying anything. The ones that get it wrong almost always discover after the fact that they'd done all three options simultaneously, with no one connecting the dots.

The Boring Strategy Wins

There's a temptation to treat this as a binary choice and pick a side: "we're a buy shop" or "we're a build shop." That framing is flattering and wrong. The right answer for any company of meaningful size is a deliberate portfolio, with each AI initiative slotted into the path that matches its strategic weight, its data dependency, and your team's actual capacity.

The 95% failure rate isn't because AI doesn't work. It's because most organizations skipped the harder question and went straight to procurement. The companies in the 5% asked the boring question first: build, buy, or partner, and why. Then they made the call once and moved on.

If you're staring at an AI roadmap and the build/buy/partner column is blank for half the items, that's where to start this quarter. Model selection, vendor pitches, proof-of-concept slides: all of that is downstream of the decision you haven't made yet.

Jonathon Carlson is the founder of Atlas Thread Digital, where he builds custom AI solutions, MCP servers, and intelligent automation systems for organizations ready to move beyond the chatbot. Reach him at jcarlson@atlasthreaddigital.com.