The TL;DR
- Buy off-the-shelf when the task is generic (drafting, summarising, transcription), low-stakes, and the value is in saving 1–2 hours per person per week.
- Build custom when AI is the product (or a strategic part of it), when proprietary data is the moat, when integration with your stack is non-trivial, or when the output sits in a regulated workflow.
- Hybrid wins more often than people realise — use the model APIs (Claude, OpenAI) as the engine, but build the orchestration, prompts, RAG and evals around them yourself.
Why this is harder than it was two years ago
In 2023, "buy" meant a chatbot widget on your website. In 2026, "buy" can mean: a vertical SaaS with AI features baked in, a horizontal copilot, a no-code agent platform, a model API, or 47 startups in your inbox. "Build" used to mean training a model from scratch. Now it usually means orchestrating an API smartly. The lines have moved.
And the cost ratios have shifted. In 2023, building anything bespoke was 10× the cost of off-the-shelf. Today it's often 2–3× — and the gap shrinks every quarter as tooling matures. That changes the math.
The 5 questions to answer before you decide
1. Is this AI capability a feature, a moat, or just an internal tool?
If it's an internal tool (your team uses it, customers don't care how it works), strongly favour buy. Cursor, ChatGPT Enterprise, Claude for Work — pick one and let your team get on with their job.
If it's a customer-facing feature in a product you sell, build at least the orchestration. Slapping a third-party chatbot on your homepage screams "we don't get it" to your most savvy customers.
If it's a moat (the AI is the product, or your AI is meaningfully better than what's available off-the-shelf because of your data), build. The vendor lock-in risk of "buy" is highest here.
2. How specific is your data and workflow?
If the AI needs to know your customers, your products, your tone of voice, your compliance rules, your internal terminology — buying generic AI is going to disappoint. The generic AI will be 70% there and the last 30% is where your customers form their opinion.
Custom RAG over your data + a tuned system prompt closes most of that gap, and that's usually a 4–8 week build, not a 6-month one.
3. What's the cost of being wrong?
For a marketing draft that a human reviews? Off-the-shelf is fine. For a quote that auto-sends to a customer with your brand name on it? Build. For a finance summary that influences trading decisions? Build, and over-engineer the eval suite. Like we did for Havéne — finance is unforgiving, and the cost of one wrong number is bigger than the build cost difference.
4. How integrated does it need to be?
If your AI needs to read from your CRM, write to your invoicing system, look up customer order history, and post to Slack — you're already building something custom even if you started with a SaaS. At a certain point you've built more glue than the SaaS provides, and you'd be better off owning the whole thing.
5. Do you have the team to maintain it?
An AI system you build is a system you maintain. Prompts drift, models deprecate, evals need updating. If your team can't take ownership in month two, either buy something with a maintenance contract, or hire an agency on a retainer (this is what most clients hire us for — building once is easy; maintaining well is the hard part).
Where buy wins (with examples)
- Generic writing assistance: ChatGPT Plus, Claude for Work, Notion AI. ~$20/user/month and your team writes faster.
- Meeting notes & CRM auto-update: Fireflies, Gong, tl;dv. Solid out-of-the-box.
- Coding assistance: Cursor, Claude Code, Copilot. There is no in-house build that beats these.
- Customer support bots for simple FAQs: Intercom Fin, Ada, Drift if your support is mostly tier-0. (Caveat: most companies overestimate how much of their support is tier-0.)
- Image / video generation: Midjourney, Runway. No reason to build.
Where build wins (with examples)
- AI features inside a product you sell. If you're a SaaS adding AI, this is in your product, not bolted on. See: industry · software.
- Workflow agents specific to your operation. No SaaS understands your specific support taxonomy, your sales playbook, your invoicing rules. Built workflow agents earn back their cost in 3–6 months. See: Operator.io.
- Regulated industries. Finance, healthcare-adjacent, legal. Off-the-shelf vendors won't cover your compliance posture without you doing the same work anyway.
- Multi-tool workflows. Anything that touches 3+ of your internal systems. The integration work IS the project.
- Brand-voice critical writing. Newsletters, customer comms, public content. Off-the-shelf AI in your brand voice is mediocre. RAG over your past content + tuned prompt gets you 90% there. See: Vigor Beans.
The hybrid path: build on top of bought engines
This is what most "build" projects actually are in 2026. You're not training a model. You're using Claude or GPT-4o via API (the "buy"), and building the orchestration, prompts, RAG, integrations, evals, observability and UI yourself (the "build"). It looks like:
- Model: rented from Anthropic / OpenAI / Google ($-per-token)
- Vector DB: pgvector inside your existing Postgres (free with your DB)
- Orchestration: TypeScript or Python you wrote
- Prompts: yours, tuned to your data
- Evals: yours, against real production traffic
- UI: in your product, or in Slack, or in your CRM
The advantage: you get the model quality of frontier labs without the moat risk of being a Salesforce-AI tenant. The model can be swapped (Claude → GPT → Gemini) without touching the rest of your stack.
The strategic mistake we see most often: companies "buying" an AI SaaS because building seems intimidating — and then six months later they've built a Frankenstein of integrations around the SaaS that is more work than the build would have been.
A simple decision tree
- Is this AI capability core to your product, brand, or competitive moat?
→ Yes: build (hybrid). No: continue. - Does it need to know your customers, products, or internal data?
→ Yes: build (hybrid). No: continue. - Does it need to integrate with 3+ internal systems?
→ Yes: build (hybrid). No: continue. - Is there a SaaS that does 90%+ of what you need today?
→ Yes: buy. No: build (hybrid).
One more honest take
Most "build vs buy" decisions are actually "build vs buy vs hire an agency to build then hand off to us". That third option exists because building your first AI system in-house has a learning curve, and the cost of a bad first build (in time and in trust) often exceeds the agency fee. We say this as an agency; we'd still tell you the same thing if we weren't one.
Not sure which path is right for your specific situation? Book a 30-minute audit. We'll go through your use case and give you an honest recommendation — including when the right answer is "buy this off-the-shelf product, don't hire us".