Your helpdesk vendor has just added an AI button. Half the office already uses ChatGPT or Copilot to write emails. And a developer tells you they could build exactly the assistant you need. So which is it: buy a tool or build your own?
For most companies the honest answer is “both, for different jobs”. Generic work such as drafting, summarising and translating is cheap to buy and hard to beat. Work that depends on your data, your systems and your rules is where bought tools tend to run out. This guide sets out the options, the criteria for choosing between them and the questions to ask a vendor before you sign anything.
Three options, not two
“Build or buy” sounds like a choice between two things. In practice there are three.
Buy a ready-made tool
- General assistants such as ChatGPT, Claude, Gemini or Microsoft Copilot. People type or paste, the assistant answers. Business plans usually add admin controls and keep your data out of model training.
- AI features inside software you already pay for: suggested replies in your helpdesk, call summaries in your CRM, receipt reading in your accounting tool.
- Vertical SaaS: tools built for one job or one industry, such as meeting note-takers, contract review for law firms or screening tools for recruiters.
You pay per seat, per use or per month, and you can start this week. You also use the tool the way the vendor designed it.
Hybrid: buy the model, build the workflow
You rent a model through an API from OpenAI, Anthropic or Google, or run an open-weight model such as Llama, Mistral or Qwen. Then you build the part that is specific to you: the workflow, the connections to your systems, the business rules and the screen where a person checks the result. When people say “custom AI”, this is almost always what they mean.
Build from scratch
Here you also own the model: you fine-tune an existing one on your examples (train it further) or train your own, then host it and keep it up to date. For an ordinary company this is rarely worth it. Better instructions and access to your documents solve most problems that look like they need a custom model. We explain the difference in prompting, RAG or fine-tuning.
Buy, hybrid or build from scratch
| Buy | Hybrid | From scratch | |
|---|---|---|---|
| Time to first result | Days | Weeks | Months |
| Upfront cost | Low | Medium | High |
| Fits your exact process | Partly | Yes | Yes |
| Uses data from all your systems | Partly | Yes | Yes |
| Cost as volume grows | Per seat or per use, can climb fast | Mostly model usage | Mostly servers and people |
| Control over data and changes | Partly | Yes | Yes |
| Moving to a newer, better model | When the vendor does | Yes | Partly |
When buying is enough
Buy when the task is generic, the tool sits where the work already happens and your process looks like everyone else’s.
Personal productivity. Drafting emails, summarising long documents, rewording a proposal, translating a contract for a first read. A business plan of a general assistant, plus clear rules on what data people may paste in, covers most of this. A 15-person accounting firm doesn’t need custom software to summarise letters from the tax office.
Features in your existing tools. If your support team works in a helpdesk and your help articles live there too, the vendor’s suggested replies may be good enough. The AI sees the same data your team uses, and nothing needs connecting.
Standard processes. Meeting notes, expense receipts, a simple chat on your website. Thousands of companies do these the same way, so vendors have already solved them well.
Two warnings. First, a licence doesn’t change how people work: someone has to show the team what to use the tool for and check that they do. Second, find out what the tool can actually see. The AI in your CRM knows your CRM. It doesn’t know the price list in your ERP or the contract in a shared folder.
When a custom build pays off
Build when one or more of these is true:
- The process is how you compete. Your pricing logic, the way you match candidates to clients, how you assess a claim. A tool everyone can buy gives you what everyone else has.
- The data is spread across systems. The answer needs the ERP, the mailbox and the document store at once, and no single vendor sees all three.
- Volume makes per-seat or per-use pricing expensive. A simple request to a model through an API often costs between a fraction of a cent and a few cents. At thousands of requests a day, the difference from per-use pricing adds up. We work through an example in what an AI project costs.
- You need control. Where data is processed, which model is used, a record of who approved what, and when things change.
- No tool fits without bending your process. If people keep a spreadsheet next to the tool to cover what it can’t do, the tool doesn’t fit.
An example. A wholesale distributor receives orders as emails, PDFs and photos of handwritten lists. An off-the-shelf document reader can pull the text out of a PDF. It doesn’t know the distributor’s product codes, the customer’s agreed prices or that “the usual blue ones” means one specific item for this customer. A custom workflow on a bought model can: it reads the order, matches the items against the catalogue and the customer’s history, and puts a draft order in the ERP for a person to confirm. Our guide to AI document processing shows how that works step by step.
Another example from our own work: for Shovels, which collects building permit data, we built Charlie, an assistant that answers questions from live database records. The model is bought. The connection to the data and the interface are built. That is the hybrid in practice.
The hybrid is usually the answer
Buy the commodity, build what makes you different. Language models, speech-to-text, text recognition and e-signature are commodities: many vendors sell them, and they get cheaper and better every year. Your workflow, your rules and the connections between your systems are not. Nobody can sell those to you.
From buying to building
- General assistant ChatGPT, Claude, Copilot 1
- AI feature in your CRM or helpdesk 2
- Vertical SaaS, configured 3
- Custom workflow on a model API The usual sweet spot 4
- Own fine-tuned model, self-hosted 5
← Buy as is Build everything →
The hybrid also keeps you from betting everything on one vendor. If the workflow is yours, you can switch models when a better or cheaper one appears, or move the sensitive steps to a self-hosted model without rebuilding everything else.
Seven questions that decide it
Go through these for each process you want to change, not for “AI” in general.
- Process fit. Does the tool follow your process, or would you change the process to suit the tool?
- Data access. Can the tool see all the data the task needs? How does the data get in, and how current is it?
- Integration. Does the result need to end up in another system? Is there an API, or will someone copy and paste?
- Control. Where is data processed? Can you choose the model, see logs and decide when things change?
- Cost at volume. What will this cost at today’s volume, and at ten times that?
- Lock-in. If you leave, can you take your data, prompts and settings with you?
- Differentiation. Is this how you win customers, or just something that has to get done?
Process fit and differentiation usually decide it. Place your candidate processes on those two axes and the answer is often obvious.
Where each process lands
- 1 Meeting notes
- 2 Email drafting
- 3 Helpdesk reply suggestions
- 4 Product recommendations in a web shop
- 5 Order entry into the ERP
- 6 Quote pricing logic
- 7 Leave requests with ten exceptions
The bottom-left corner deserves a word. If a generic process fits no ready-made tool, the process itself is often the problem. Ten exception rules for leave requests are usually cheaper to simplify than to automate.
Questions to ask a vendor before you sign
Demos look great on the vendor’s examples. Ask for a trial on yours, and get written answers to these:
Vendor questions
- Is our data used to train your models? Get the answer in the contract, not only in the FAQ
- Where is data processed and stored? EU hosting, subprocessors, how long data is kept
- Will you sign a data processing agreement? Required under GDPR when they process personal data for you
- Which model is underneath? Can we choose it, and will you tell us when it changes?
- Is there an API? So results can flow into our other systems
- How is pricing calculated? Per seat, per request or per resolution, and at ten times our volume
- Can we export everything if we leave? Data, prompts, settings and history
- Can we test on our own examples first? Before we sign, with 30 to 50 real cases
Mistakes we see
- Buying before you know the process. A tool picked from a demo, then a search for something to use it on. Start with the process and what “done” looks like.
- Building what a subscription already does. If a ready-made tool covers nearly all of a generic task, live with the gap. Custom work belongs where it makes a difference.
- Treating the model as the hard part. In custom projects the model is usually the easy part. Connections to old systems, messy data, permissions and the review screen take most of the work.
- Testing on different data. A vendor demo on its sample data against a custom pilot on your real data proves nothing. Test both on the same set of real cases.
- Locking your logic into someone else’s builder. Prompts and rules set up inside a vendor’s no-code tool often can’t be exported. Know what you would lose if you left.
What to do next
- Pick one process and write down what it takes today: steps, time, systems and who checks the result.
- If a ready-made option exists, try it first for two to four weeks on real work.
- Note where it breaks: missing data, no connection to your systems, the wrong output format, too much copying and pasting.
- If the gaps sit in the part of the process that sets you apart, scope a hybrid build: a bought model with your own workflow around it.
- Agree a fixed price for that first process before any build starts, and ask for an estimate of the monthly running costs too.
If you’d like a second opinion on what to buy and what to build, book a 30-minute call. You’ll leave knowing what to build first, or why you don’t need to build anything yet.
