Build vs buy AI: custom solution or off the shelf?

Somebody in every leadership team eventually says it: "Why are we paying a subscription for this? We could build our own." It sounds like ambition. Sometimes it is. More often it's the opening line of an expensive year.

The build vs buy AI question is older than AI. What's changed is that the gap between the two has filled with options, and most of the bad decisions I see come from treating it as a binary.

The short answer: buy off-the-shelf AI for anything that isn't your competitive edge, which for most small and mid-sized businesses is nearly everything. Build, or more accurately configure and integrate custom AI on top of a bought model, only when your own data or workflow is the advantage and you can afford to maintain it. Training your own model from scratch is rarely the right answer outside very large or specialised organisations.

Key points:

• In 2025, 76% of enterprise AI use cases were purchased rather than built, up from 53% in 2024 (Menlo Ventures).
• MIT researchers found externally sourced, customised tools reached deployment about 67% of the time, against about 33% for internal builds (MIT NANDA).
• There are at least five options between "buy" and "build", and their costs differ by orders of magnitude.
• Whatever you choose, the ongoing cost of ownership matters more than the launch cost.

Build vs buy AI: what are the real options?

Gartner's version of this is the most useful I've found, because it replaces two boxes with a spectrum. It splits generative AI deployment into five approaches and puts indicative costs on each (Gartner, July 2024).

A health warning: these are Gartner's 2024 estimates for large-organisation deployments. A 30-person firm switching on an AI assistant for its team won't spend $100,000. But the shape is what matters. The approaches involve different ownership and operating commitments. These example costs do not form a universal price ladder.

When most smaller businesses say "build", they mean the middle of this table: embed or extend. You buy access to someone else's model and build the plumbing, prompts and data connections around it. That's a perfectly good answer. It's just not the same decision as building your own AI.

What are other businesses choosing?

Increasingly, they're buying. Menlo Ventures' survey of 495 US enterprise AI decision-makers found the split moved from 47% built and 53% bought in 2024 to 76% purchased in 2025.

The success data points the same way. MIT's NANDA study found that strategic partnerships (buying from a vendor or co-developing with one) were about twice as likely to reach deployment as fully in-house builds, and that employees used the externally built tools nearly twice as much. The authors are candid that this is based on 52 organisations and may reflect the kind of companies that choose partners, rather than proving partnering causes success. Still, it's a striking gap.

Forrester predicted that three out of four companies building ambitious agentic AI architectures on their own would fail, and in its review of 2025 counted that among its accurate calls, noting that most agentic use cases in production run on vendor platforms (Forrester). If you're weighing AI agents for your business, that's worth sitting with.

When should you buy off-the-shelf AI?

Buy when the job is one that thousands of other businesses also need doing. Drafting, summarising, meeting notes, coding assistance, answering the phone, triaging email, searching documents. Check whether a maintained product already meets your requirements before funding an equivalent feature.

Buy when speed matters. A business seat on a mainstream assistant can be live this afternoon, at prices such as US$20 per seat a month billed annually on Claude's Team plan (Anthropic). Comparing the main business assistants is its own exercise; see ChatGPT vs Copilot vs Claude for business.

And buy when you don't have people to maintain the thing. That's the bit that gets skipped. An AI system isn't finished at launch. Models get updated or retired, behaviour drifts, staff find edge cases, regulations change. Someone has to own that, every month, indefinitely.

When do custom AI solutions make sense?

Custom AI solutions earn their keep in fewer situations than vendors of custom AI development would like. The real signals:

• Your data is the edge. You hold years of proprietary pricing, claims, case or production data, and the value comes from AI working on it in a way no general product does.
• Your workflow is unusual. The process is specific to your business, and bending it to fit a product would cost more than adapting the tech to fit you.
• Nothing on the market does it. You've actually tested the off-the-shelf options, ideally through a proper AI proof of concept, and they fall short on something that matters.
• Volume changes the maths. At high enough volumes, paying per call to an API can cost more than a tailored pipeline. Model it with your real numbers rather than a vendor's slide.
• Control is non-negotiable. Regulation, client contracts or data residency rules limit where data can go and what a vendor may do with it.

Even then, "build" almost always means building on top of a commercial or open model, not training your own. Fine-tuning and training from scratch belong to organisations with specialist teams and specialist budgets.

How much does custom AI development cost compared with buying?

Honest answer: it varies more than any published figure suggests, and most of the ranges you'll find online come from development firms selling the build. The Gartner table above is the most credible reference point I know, and it's pitched at enterprise scale.

What I can tell you is where the money goes, because the categories are the same whatever your size:

• Buying: per-seat or per-use fees, set-up and configuration, training your people, and the admin of managing access and data settings.
• Building on a bought model: development time, integration with your systems, API usage (charged per million tokens, such as US$2 in and US$10 out on Anthropic's Sonnet 5.5 at the time of writing), testing, security review, and then the ongoing maintenance most budgets forget.
• Training your own: specialist people, compute, data preparation and evaluation at a scale that puts it out of reach for most firms.

The trap is comparing a build quote with a year of subscriptions. The build quote is a down payment on a system you'll be paying to look after for years. Compare three-year totals, including the people who'll own it. The guide to AI implementation cost breaks the lines down further.

How do you avoid vendor lock-in if you buy?

Lock-in is the strongest argument for building, and it's a real one. In a Zapier survey of 500 US enterprise executives, only 6% said they could stop using their primary AI vendor without disruption, and 44% were already using several AI vendors to spread the risk (Zapier).

You can buy and still keep your options open:

• Keep your own data, documents and prompts in systems you control, not only inside the vendor's tool.
• Check the contract for data export, deletion on exit and notice periods for price changes.
• Where you build integrations, put a thin layer between your systems and the model so you can switch providers without rewriting everything.
• Write down which business processes depend on which tool. You can't plan an exit you can't see.

These belong in your AI governance framework, not in a drawer.

My take

I've spent 20-odd years on both sides of this decision, as a CTO building products and as the person signing purchase orders. My default for AI is to buy, and I'd want a good reason to move off it.

The risk I would look for runs the other way from the worry. A business commissions something custom to avoid a subscription, then discovers it now owns a small software product with nobody to look after it.

The middle of the spectrum is where the interesting work is. My own lab, Ascendancy IQ, is a good example of what I mean: AI agents doing real work, running on commercially available models. What I built is the layer around them: shared memory, approval gates and controls. That's the part specific to how I work. The models are someone else's problem, and I'm happy for them to stay that way.

One more thing. If you do decide to build, decide who builds it. A freelancer, an agency and an in-house hire each leave you with a different maintenance story. The guide on AI consultant vs AI agency vs in-house covers that choice.

How to decide whether to build or buy AI

1. Describe the job, not the technology. "Answer 80% of routine customer emails" is a job. "An AI chatbot" is a solution looking for one.
2. Ask whether it's your edge. If the capability is widely available, test a bought option first. Your advantage may lie in how you use it, rather than in owning the code.
3. Search the market properly. Look at the tools you already pay for first; many now include AI features you haven't switched on.
4. Test the best bought option on real work. A time-boxed proof of concept with agreed thresholds.
5. If it falls short, place yourself on the spectrum. Embed or extend before you even consider fine-tuning or training.
6. Cost it over three years. Include maintenance, monitoring, model changes and the person who owns it.
7. Plan the exit either way. Data portability if you buy; documentation and handover if you build.
8. Review annually. Today's custom feature is often next year's standard product feature.

Frequently asked questions

Is it better to build or buy AI for a small business? For most small businesses, buying is better. Off-the-shelf tools are cheaper to start, faster to deploy and maintained by someone else, and custom AI only pays off when your own data or workflow is a genuine competitive advantage.

What are custom AI solutions? Custom AI solutions are AI systems built or configured around a specific business's data and processes. In practice, most are built on top of commercial models through APIs, connected to the company's own systems and data, rather than trained from scratch.

How much does custom AI development cost? It varies widely with scope, integrations and volume. Gartner's 2024 enterprise estimates range from about US$750,000 upfront to embed AI models in custom apps to US$8 million or more to build a custom model from scratch, but smaller projects built on existing models cost far less.

Is building AI cheaper than buying in the long run? Sometimes, at high volumes, but only if you count the full cost of ownership. Maintenance, model updates, monitoring and the staff to run it often outweigh the subscription savings.

What is the hybrid approach to build vs buy AI? The hybrid approach means buying access to a commercial AI model and building your own layer on top: integrations, data connections, prompts and controls. It's how most businesses that "build" AI actually do it.

Sources

• Gartner predicts 30% of generative AI projects will be abandoned after proof of concept by end of 2025 (including Figure 1, costs of GenAI deployment approaches), Gartner, July 2024
• 2025: The State of Generative AI in the Enterprise, Menlo Ventures, December 2025
• The GenAI Divide: State of AI in Business 2025, MIT NANDA, July 2025 (PDF)
• How accurate were our predictions for 2025?, Forrester
• AI vendor lock-in survey, Zapier
• Claude pricing, Anthropic, checked October 2026

If you're weighing a custom AI build against an off-the-shelf tool and want an independent view before you sign either, let's talk.