For most teams the answer is buy or blend. Build only when nothing on the market fits, the capability is core to your business, and you can staff it for years. Try a bought or blended version first. It will show you what a custom build would have to do.
The three paths
It is not a simple choice between two options. There is a middle path, and it is where many teams end up.
- Buy. Use an existing AI product, such as a SaaS tool or built-in assistant, and configure it rather than building anything.
- Blend. Buy the model or platform, then build your own workflow, prompts, and integrations around it. This is where many teams land.
- Build. Develop a custom solution, which may include training or fine-tuning your own models and running them on your own infrastructure.
Eight questions that decide it
Answer these for one project at a time. A support assistant and a pricing model can have different answers.
- Fit. How well do products you can buy today cover the task?
- Advantage. Does this capability set you apart from competitors, or is it a routine function?
- Data. How sensitive is the data it would use, and can it go to a vendor?
- Team. What technical skills do you have to build it and run it?
- Timing. How soon does it need to work?
- Integration. How much does it need to connect to your own systems?
- Scale. How much use do you expect, now and later?
- Upkeep. Who will maintain it after the launch?
The AI build vs buy decision tool scores your answers and shows which path fits best, why, and what to watch for.
The costs people miss
When you build
The model call is often the smallest part of the cost. Most of the money goes to work around it: preparing and cleaning data, building tests to measure quality, monitoring, security reviews, and the people who keep it running. These continue long after launch.
When you buy
The main risks are being tied to one vendor, prices that grow with use, limited control over how the product behaves, and questions about how your data is stored and used. Read the contract and the data terms before you commit, and make sure you can take your data with you.
How to decide, step by step
1. Pick one project
Decide for a single use case, not for your whole AI strategy.
2. Answer the eight questions
Use the questions above or the decision tool. If two paths are close, take that as a sign to try the simpler one first.
3. Test with real examples
Don’t decide from a demo. Run two or three products, or two model APIs, against the same ten real examples from your work. Compare the quality and what each costs. The LLM model comparison helps with prices and context windows.
4. Cost it over time
Estimate what the option will cost at today’s use and at ten times today’s use, with the LLM API cost calculator. Then see whether it pays for itself with the AI ROI calculator.
5. Check the vendor
If you’re buying or blending, ask the vendor about security, data handling, and contract terms before a trial. The AI vendor questionnaire generator gives you a list of questions that fits what the vendor will touch.
6. Start small and review
Start with the simplest path that could work, with a named owner and a review date. If it falls short, you will know exactly what a custom build would need to do better.
Mistakes to avoid
- Deciding for the whole company at once. Different projects can call for different answers.
- Judging by a demo. A demo shows the best case. Test with your own examples.
- Leaving out the upkeep. A custom build needs a team for years. Count that cost.
- Building because it feels more in your control. Build when the capability is core to your business, not just because you can.
Common questions
- Should we build or buy AI?
- For most organizations, buy or blend. Buy when a product already does the job and the task isn’t what sets you apart. Build only when nothing fits, the capability is core to your edge, and you have a team to maintain it for years.
- Can we start with buy and move to build later?
- Yes, and it is often the safest route. A purchased or blended prototype shows what the solution really has to do, and what it costs at scale. That is better evidence for a build decision than a forecast made in advance.
- Is building cheaper at large scale?
- Sometimes, but check it. The cost of a custom build is mostly the team, the data work, the testing, and the upkeep, not the model itself. Add up the full cost over several years before you assume that building is cheaper.
I’m Courtney Stephenson, an AI strategy and implementation leader. If you’re weighing a build or buy decision, let’s talk it through.