Custom AI vs Off-the-Shelf: The Decision Framework We Give Every Client

The "Year 2 Problem"

SaaS looks cheap in Year 1. A £200/mo tool sounds fine. But at 20 users it's £4,000/mo with per-seat pricing. Add a 20% annual price increase, two premium add-ons you need, and the API rate limits that force you to buy a higher tier — and that "affordable" tool is now your biggest software spend.

The crossover point where custom becomes cheaper than SaaS is typically Year 3 for teams over 15 people or processes running more than 10,000 operations per month. Below that threshold, SaaS almost always wins on pure economics.

The calculation changes entirely when you factor in the value of your data. If the SaaS vendor is training on your inputs, pricing your proprietary workflows into their pricing tiers, or making it difficult to export your data, the hidden cost is your competitive advantage. Our full cost breakdown covers the numbers most vendors don't share.

The Build Trap

We've seen this repeatedly: the pilot works in 4 weeks. Then production requirements hit — security review, error handling, monitoring, user authentication, audit logging, disaster recovery — and the scope triples. The £20k project becomes £60k and takes 8 months.

Custom builds also require ongoing maintenance. The team that built it needs to stay available. API providers deprecate versions. Models get updated and outputs change. Budget at least 20% of the build cost annually just for upkeep — before any feature additions.

Our 3-Question Test

We run every client through three questions before recommending a custom build:

• 1. Is your data or workflow genuinely unique — not just customised, but proprietary in a way that creates competitive advantage?
• 2. Will you need this for 3+ years? Custom builds amortise their cost over time; short-lived projects don't justify the investment.
• 3. Do you have the budget for a 2–6 month build and the patience to not ship until it's right?

If you answer no to any of these, start with SaaS. Revisit in 12 months once you understand your actual requirements.

The Hybrid Approach We Actually Recommend

Most of our clients end up here: off-the-shelf tools orchestrated by custom middleware. Make or n8n as the workflow layer, OpenAI or Claude as the intelligence layer, with custom code handling the unique business logic that no SaaS covers.

This approach typically costs £8–20k to build, uses existing tools at their SaaS pricing, and gives you full control of the logic that matters. You're not choosing between build and buy — you're buying the commodity parts and building the differentiating layer. Here's what a real custom build looks like from architecture to deployment.