Key takeaways
Building the agent is the small part of the bill.
Buy the plumbing and build what makes you different.
Decide by year one needs, not by what looks exciting.
Every team weighing AI agents reaches the same question: build it ourselves or buy a platform? The honest answer depends less on engineering talent than on what you want your team to maintain in two years.
What building really includes
The first agent is the easy part. Evaluations, permissions, audit logs, retries and model upgrades are what take the time, and none of them set your product apart.
When buying makes sense
If agents support your operations rather than define your product, a platform gets you to production faster and moves the maintenance elsewhere.
A simple way to decide
List the capabilities you need in the first year, then mark which ones are truly unique to your business. Build those. Buy everything else.
Count the people, not the licences
The real cost of an in-house platform is the team that keeps it running. Evaluations, permissions, audit logs, connectors and on call rotas all need owners, and those owners are not working on your product.
Before comparing prices, write down who would do this work in month eighteen, once the launch excitement has gone and the original builders have moved on.
Questions to ask any vendor
A polished demo tells you very little. Ask about the parts that only matter once you are live.
Can we export our prompts, tests and logs?
How do we test a change before it ships?
What does the audit trail show, action by action?
What happens to our data, and who can see it?
Clear answers are a good sign. Vague ones usually mean the feature is on a roadmap.
Run a small test before you commit
The best evidence is a real workflow, not a comparison table. Take one job, set a two week limit and try it on the platform you are considering, using your own data and your own edge cases.
Pay attention to the second week. Anyone can make a demo work on day one. What matters is how hard it is to change a rule, test the change and explain a bad result to a colleague. That experience is what you are really buying.
Keep your way out
Whichever path you choose, protect your exit. Keep your test cases, labelled data and written procedures in formats you own. Those assets hold most of the value, and they move with you.
Most teams end up somewhere in the middle: a bought platform underneath, with their own logic and data on top. Starting from there saves a year of work on problems that were never yours to solve.
About the author
Mara Lind
Mara leads product at Falak. Before that she ran operations teams and spent years cleaning up after tools that stopped halfway.
