Build, buy or get help
Build or buy: choosing how to bring it into the company

The short answer
Most companies now buy. Buy when your documents already live in one vendor's software and good-enough answers will do. Build, or have a partner build, when answers must be exact, sourced and checked, or your documents are spread across many systems.
Whichever you choose, nobody sells you clean documents. That work stays yours.
Think of feeding a company event
You can order ready meals: quick, cheap per head, and the same as everyone else's. You can hire a caterer to cook your menu in your kitchen. Or your own staff can cook, if you have a cook and the time.
Ready meal, caterer or your own kitchen: each is right for some events.
More companies buy each year
Menlo Ventures asks US companies every year how they get their AI. The share bought from vendors, rather than built in-house, jumped in a year.
Bought from vendors, 2024
The rest, 47%, were built in-house. 600 IT decision-makers.
Bought from vendors, 2025
Share of AI uses purchased rather than built. 495 decision-makers.
An MIT report in 2025 went further: projects bought from or built with an outside partner reached real use about twice as often as ones built in-house, roughly two in three against one in three. Treat that with care. It rests on interviews with 52 organizations, the authors say the link may not be cause and effect, and other researchers have criticised the report's methods (MIT NANDA, The GenAI Divide, July 2025).
What buying costs, per person
Ready-made products are mostly priced per person, per month. List prices, checked on the vendors' own pages on 16 November 2026:
Microsoft 365 Copilot
On top of a Microsoft 365 plan you already pay for.
$30 a person a monthGemini in Google Workspace
Included in Workspace Standard and up; the price is for the whole plan.
From $14 a person a monthChatGPT Business
Billed yearly; $25 billed monthly.
$20 a person a monthAmazon Q Business Pro
Plus a charge for the size of your document index.
$20 a person a month
Per-person prices add up fast. Copilot for all 800 people in a mid-sized company is $288,000 a year, whether a third of them use it or all of them do. A system built for one help desk is usually paid for once, plus upkeep. What it really costs works through an example.
Three ways to do it
Buy a ready-made product
An assistant already connected to where your documents live, often inside the office software you use.
Best for: general questions across everyday documents, when they already sit in that vendor's tools.
Gives up: control over how it finds and checks answers, and a price per person for everyone.
Build with an outside partner
A firm builds it around your documents and rules, then hands it over or keeps running it.
Best for: answers that must be exact and sourced, documents in many systems, and a first project nobody inside has done before.
Gives up: a bigger bill up front, and your people must learn enough to own it.
Build it yourselves
Your developers build it on an AI company's models and tools.
Best for: companies with developers who have time, and a reason to keep the know-how inside.
Gives up: speed. The first one teaches your team as it goes.
Getting locked in
Changing AI models is easier than it sounds, and rarer. In a 2025 survey of companies building with AI, only 11% had switched model companies; most upgraded to a newer model from the same one (Menlo Ventures, July 2025).
The real lock-in is everything around the model: your cleaned-up documents, the catalogs built from them, your test questions, and the rules about who may see what. If those exist only inside one vendor's product, leaving means doing them again.
Own your documents, your test questions and your rules, whoever runs the system.
What goes wrong
- Seats for everyone before proving one use. Hundreds of licenses, a handful of regular users.
- Building what a product already does. Months of work to answer "where's the holiday calendar?"
- Buying for a job that needs exact answers. A general assistant guesses where a contract question needs the clause.
- A partner who leaves no know-how. It works on handover day, and nobody inside can change it.
- Expecting the product to fix messy documents. It reads what's there. Is our information ready?
Questions to ask your team or vendor
- Where do our documents live today, and does this reach all of them?
- Does every answer show the passage it came from, and respect who may see what?
- What does it cost a year for everyone who'd have it, and what if only a third use it?
- If we leave, what do we take with us: our documents, our test questions, our settings?
- Who owns and maintains it after launch, and do they have the time?
Other names you may hear
- Build vs buy
- Making something yourself, or paying for a product someone else made.
- SaaS, software as a service
- A product you rent online, usually per person per month.
- Seat, license
- One person's paid access.
- Systems integrator, implementation partner
- A firm that builds and connects systems for you.
- Foundation model, LLM
- The AI model itself, made by companies such as OpenAI, Anthropic or Google, which products are built on.
- Vendor lock-in
- When leaving a supplier would cost so much that you stay.