The build-versus-buy question is where most SMBs waste money on AI — either by over-building something they should have bought off the shelf or by stitching together five SaaS tools when a tailored build would have been cheaper and faster. The framework for deciding is actually fairly simple once you separate the layers.
The three layers
- Model layer. The underlying LLM. Never custom for an SMB. Use OpenAI, Anthropic, Azure OpenAI, Bedrock, or similar. Building your own model is almost never the right call at SMB scale.
- Product layer. The user-facing application. Almost always custom for vertical-specific workflows — voice agent for your specific practice, document automation for your specific firm's templates, intake flow tied to your specific system-of-record.
- Infrastructure layer. The deployment — public inference, private tenant, on-premise. Depends on data sensitivity. See On-premise vs private cloud AI.
Always buy
- The model itself.
- Core infrastructure — cloud tenancy, compute.
- Commodity horizontal tools — email marketing, CRM, scheduling SaaS.
- Off-the-shelf chatbots for pure-FAQ use cases without integration depth.
Usually build (with help)
- Voice agents tied to your specific phone system, scheduling stack, and practice-management system.
- Document automation working against your specific templates and workflows.
- Intake and scheduling wired to your CRM or practice management.
- Knowledge assistants indexed against your documentation.
- Any workflow where the data flow, integration, or compliance posture is specific to your business.
The expensive middle
The failure mode to avoid is the middle — stitching together three or four SaaS tools with manual handoffs between them because each tool covers 60% of the use case. The integration work to make them cohere is usually more expensive than building the tailored solution from the start, and the result is fragile.
Build-with-help vs build-from-scratch
"Build" here doesn't mean hiring a full dev team. It means working with a consulting partner — us or someone else — to build on top of commercially available models and infrastructure. Almost nothing at SMB scale justifies genuinely-from-scratch engineering. What justifies custom work is the product layer where your business is specific. See our custom and private AI engagement model.
The decision questions
- Is there a workflow that is specific to how your business operates? (If yes, lean custom.)
- Is the integration depth going to determine success? (If yes, lean custom.)
- Is the use case genuinely generic? (If yes, lean off-the-shelf.)
- Is data sensitivity driving the architecture? (Affects infrastructure, not product.)
Scope an engagement if you want help thinking through your specific build-vs-buy calls.