--- title: "Custom AI systems vs off-the-shelf: when each makes sense" description: "A working framework for choosing between SaaS AI tools and custom systems built around your business. Three signals that tip the decision either way." slug: "custom-ai-systems" date: "2026-04-30" updated: "2026-04-30" author: "Nick Barnard" authorAvatar: "https://images.unsplash.com/photo-1535713875002-d1d0cf377fde?q=80&w=200&auto=format&fit=crop" image: "https://images.unsplash.com/photo-1531746790731-6c087fecd65a?q=80&w=2000&auto=format&fit=crop" category: "Strategy" tags: - AI - Custom systems - Buy vs build - Operations primaryKeyword: "custom AI systems" secondaryKeywords: - "build vs buy AI" - "AI for business operations" - "off-the-shelf AI tools" canonicalUrl: "https://tolly.ai/blog/custom-ai-systems" markdownUrl: "https://tolly.ai/blog/custom-ai-systems.md" featured: false published: true --- Buy off-the-shelf when the problem is generic, the data is standard, and your workflow can bend to fit the tool. Build custom when your data is unusual, your workflow already encodes hard-won expertise, or the SaaS option forces you to redesign your business around it. Most companies should buy more than they build — and the few exceptions matter a lot. ## Three signals that tip the decision The decision usually comes down to three signals. If two or three lean the same way, the call is obvious. If they split, the right move is almost always a hybrid. ### Signal one: how generic is the problem? If three companies in your industry would describe the work using the same words, an off-the-shelf tool probably exists and is probably good enough. Lead routing, calendar scheduling, basic document extraction — these are solved problems with mature vendors. Don't reinvent them. If your description of the problem makes peers in your industry raise eyebrows, custom is on the table. The eyebrow is the signal. It means the problem has an idiosyncratic shape that generic tools won't fit cleanly. ### Signal two: how unusual is your data? SaaS tools assume your data looks like everyone else's. They have a schema. If your CRM has six custom fields and a quirky tag taxonomy that took years to refine, an integration into a generic AI tool will lose most of the signal. The tool will see "deal stage" and miss the part of your tagging system that actually predicts close rate. Custom systems can model the schema you actually have, not the one a vendor wishes you had. ### Signal three: what's the cost of being wrong? For low-stakes outputs (drafted email, suggested follow-up, internal summary), off-the-shelf is fine — a human reviews before action. For high-stakes outputs (auto-sent invoices, customer-facing decisions, regulatory filings), the audit trail and the override path matter more than the model. Custom systems are easier to instrument with the controls compliance and operations teams actually need. ## The hybrid that wins most of the time The third option covers the majority of real engagements: buy the foundation models, buy the obvious tools, and build a thin custom layer that does the choreography between them. Your CRM is bought. Your inbox is bought. The model is bought (OpenAI, Anthropic, whatever). What's custom is the small bit of code that says: when a lead with these specific tags hits this specific stage, draft this specific email using this specific context, send it through this specific approval flow, and log it in this specific table. That bit is genuinely yours. It encodes the way your business handles the case, not the way the vendor's template handles it. And because it's thin, it stays cheap to change when your business changes. ## A worked example A 20-person services company asked us about replacing their lead-qualification process with AI. The off-the-shelf option (a SaaS qualifier) cost about $400/month and would have replaced their existing process wholesale. We pushed back: their qualification rubric had ten years of niche industry knowledge baked into it. The SaaS tool would have been *good*, but it would have been good at a generic version of the problem, not theirs. What we shipped instead: a custom workflow that read their inbound leads, ran an OpenAI structured-output call against *their actual rubric* (which lived in a Google doc the team owned), and dropped scored leads with a Slack approval into the same Pipedrive they already used. Total cost: about half the SaaS option, with the rubric portable and human-readable so any team member could adjust it without filing a ticket. The point isn't that custom always wins. It's that the rubric they'd built was the asset, and the SaaS tool would have asked them to throw it away. ## Frequently asked questions ### Isn't custom AI more expensive? Up front, sometimes. Over five years, almost never — the SaaS option's per-seat cost compounds, and the customization fees you pay to make it almost-fit usually exceed what custom would have cost in the first place. ### What about model lock-in? The model is the part you should *not* lock yourself into. We build systems where the model call is one well-defined function, swappable in an afternoon. If GPT gets cheaper or Claude gets better, you change one file. The rest of the system doesn't notice. ### When should I absolutely not build custom? When the problem is genuinely generic *and* you don't have the expertise on your team or in your vendor relationships to maintain it. A custom system you can't maintain is worse than the SaaS you grumbled about. ## The takeaway Buy generic when the problem is generic. Build custom around the parts of your business that are actually yours. And be suspicious of any vendor — including us — who tells you the answer without first asking what specifically about your work doesn't fit the template.