Every organization arriving at "we should do something with AI" faces the same three options — build, buy, or partner — and most pick one by temperament rather than analysis. The honest comparison is not difficult, and it usually points somewhere specific.
Buy: an existing product with AI in it
What it is: your helpdesk vendor ships summarization, your CRM ships lead scoring, your document platform ships search. You switch it on.
Cost: incremental subscription, often a tier upgrade. Predictable per seat.
Right when: the capability is generic, the vendor already holds your data, and the value is convenience rather than differentiation. Which is most of the time.
The trade: you accept their roadmap, their accuracy, and their data terms — read those, particularly on training. You cannot tune it to your process, and you cannot take it with you.
Build: a custom system on model APIs
What it is: software your business owns, built on Claude, GPT, or Gemini, connected to your systems and grounded in your data.
Cost: real up-front spend plus ongoing running and maintenance costs. Usage-based rather than per seat, which inverts the economics at scale.
Right when: the process is genuinely yours, the answer depends on data in your systems, permissions or compliance are in play, or no product does it. Also when headcount is large enough that per-seat pricing stops making sense.
The trade: you own it, including the maintenance and the model deprecations. Nobody else is going to patch it.
Partner: someone runs it for you
What it is: a firm builds and then operates the system — hosting, monitoring, accuracy management, model migrations — for a monthly fee.
Cost: lower up front than a pure build, higher over time. Somewhere between the other two.
Right when: you need a custom capability but have nobody internally who will own it after launch. This is the most under-considered option and it fits more organizations than choose it, because "nobody will maintain this" is the most common reason builds fail.
The trade: ongoing dependency. Mitigate it by owning the code, the prompts, and the evaluation set regardless of who operates the system — get that in writing at the start.
The three-year comparison
For a capability serving 50 staff, gather these line items from your own quotes for each of three years:
- Buy: the per-user add-on or tier upgrade times your seat count, rising with headcount and price increases. Zero engineering.
- Build: the build quote, then model usage, hosting, and a standing maintenance budget each year. Front-loaded, then largely independent of headcount, provided somebody owns it.
- Partner: a smaller build fee plus a monthly operating fee, with operations included. Ask exactly what the monthly fee covers and what it excludes.
At 50 seats, buying usually wins on cost if a product fits. As seats grow, the build line barely moves while the buy line climbs with every seat and every renewal. Headcount and fit decide this more than anything technical.
Four questions that settle it quickly
- Does a product already do this acceptably? If yes, buy it and adapt. Most conversations should end here.
- Does the answer depend on data in your systems? If yes, buying rarely works — products cannot see your records.
- Who owns it in eighteen months? If nobody can be named, do not build. Partner or buy.
- How many seats in three years? Large and growing pushes hard toward build, because per-seat pricing compounds and usage-based pricing does not.
Answering these honestly usually points somewhere specific, and it takes an afternoon rather than a procurement cycle.
The combination that usually wins
Buy for the commodity capabilities — summarization, transcription, search inside a product you already run. Build the one or two things that are genuinely specific to how you operate. Partner on that build if nobody internally will own it.
What almost never works is building a worse version of something you could subscribe to, or subscribing to eleven overlapping tools because no one made a vendor and license decision. The questions in our custom versus off-the-shelf guide help sort the first case, and ChatGPT versus a custom tool covers the line between them.
We do build and partner engagements under AI automation — including telling you when the answer is "buy the product."
Describe the capability and we will tell you which of the three it should be.
The Cyber-Insurance Questionnaire Gap Check
Covers MFA, endpoint protection, backups, email security, incident response and training. Each question comes with the evidence an underwriter expects to see.