The board of a 70-person financial services firm spent an hour debating whether to build a custom AI underwriting model. The CTO had scoped it at £180,000 and nine months. Halfway through, the head of operations asked whether anyone had tested the AI features now embedded in their CRM. Nobody had.
The CRM vendor’s AI handled 70% of the task at £40 per user per month. The remaining 30%, where the firm’s proprietary risk logic actually sat, was where the build maths started to make sense, and the scope dropped from £180,000 to a £40,000 thin custom layer. The board had been asking “should we build”. The right opening question was “how much of this does the vendor already solve, and where does our edge actually sit”.
The choice you’re facing
Buy means licensing an off-the-shelf AI product or API service from an established vendor. Build means contracting engineers, in-house or external, to construct a bespoke AI system from open-weight models or APIs. The 2026 shift is that buy now covers around 80% of common SME use cases at material savings on cost, time-to-value and ongoing maintenance, while build retains a sharper case in the remaining 20%.
Five product categories sit on the shelf for an SME owner. Embedded SaaS AI bundles capability into software you already pay for, like Microsoft Copilot at around £20 per user per month, or HubSpot AI at around £40. API services like OpenAI, Anthropic and Google Vertex let a developer wire third-party models into bespoke applications, paying per call. Open-weight self-hosted models like Llama and Mistral let you run inference on your own infrastructure. No-code AI builders like Zapier AI and Microsoft Power Platform sit between SaaS and custom development. Bespoke developer-led build is the highest-friction, highest-investment path.
Three structural shifts made this a genuine trade-off. Vertical SaaS vendors have pre-built domain features covering 70 to 85% of common use cases in days. Open-weight model access has eroded the proprietary moat frontier vendors once held. The UK AI engineering talent market has tightened, making custom builds slower and more expensive relative to vendor solutions.
When buy is the right answer
Buy wins where the use case is common across your industry vertical, time-to-value is critical, the feature set is stable, and your firm lacks the technical depth to maintain a bespoke system. These conditions are met more often than the build-curious version of the conversation suggests, which is why SaaS AI adoption has accelerated through 2025 and 2026.
The clearest pattern is embedded AI in business-critical software. A mid-market recruitment firm running HubSpot can activate HubSpot AI for around £40 per user per month and start getting auto-generated candidate summaries, outreach emails and high-probability matches within hours. A custom build of an equivalent recruitment-assistant application would take six to nine months and £150,000 to £300,000, and would still likely underperform the off-the-shelf product.
Octopus Energy is the canonical UK case for the buy path on customer-facing AI. Rather than building proprietary chatbots, Octopus integrated with embedded customer service platforms to handle tier-one inquiries. Customer service automation is not Octopus’s differentiator, energy procurement and pricing strategy are. By buying, Octopus avoided 12 to 18 months of engineering work and kept its capacity on forecasting and pricing optimisation, where domain knowledge creates advantage.
The trade-off is the 75 to 85% capability fit. The win is two-week deployment, the vendor handling model updates and security patches, and a support organisation when something breaks.
When build is the right answer
Build is justified by one of five conditions, any of which can carry the case on its own. Proprietary business logic that no vendor encodes. Competitive data advantage worth preserving. Data sovereignty or residency constraints that rule out US-based SaaS. Vendor dependency carrying material business risk. Specialist capability not yet productised. These conditions are met less often than the buy case, but where they apply the stakes are usually higher.
Monzo Bank is the working example of build winning on proprietary logic. Monzo’s transaction profile, predominantly small real-time retail payments from digitally native customers, differs fundamentally from the profiles vendor fraud detection tools were trained on. By building custom detection, Monzo could update models as new attack patterns emerged and optimise for its own false positive tolerance. The competitive advantage on Monzo’s specific customer base directly reduced fraud losses.
Sovereignty is the other common build trigger for UK firms. The FCA’s SS2/21 guidance requires firms to retain adequate oversight, contractual control and exit options when they outsource regulated functions. The ICO has been clear that controllers cannot outsource their compliance obligations under UK GDPR. For firms handling health, financial or biometric data, this often forces in-house deployment of open-weight models on UK infrastructure.
Build also requires stable access to AI engineering talent, the most fragile assumption. The Stack Overflow 2025 Developer Survey shows AI specialists commanding a 15 to 25% premium over general engineers, and the Hays 2026 UK guide puts senior AI engineers at £100,000 to £150,000 fully-loaded. The decision should be contingent on having the team already and expecting to keep them, or a trusted contractor network. Otherwise the project will slip when talent moves.
What it costs to get wrong
Both paths carry material failure modes, and they look opposite. Buy fails through vendor pivots and pricing surprises. Build fails through budget overruns and obsolescence. Stress-testing the decision against both sets of failures is what separates a procurement choice that holds for three years from one that has to be unwound at 18 months.
On the buy side, the dominant failure is feature deprecation and surprise pricing. A vendor sunsets a feature or pivots after an acquisition, and workflows built on it have to be rewritten. Salesforce price rises of 20 to 30% year-on-year on multi-year renewals are documented across the trade press. For an SME with limited negotiating room, a 25% rise on core AI licensing compresses margins fast. Data lock-in is the slower version. By the time a firm wants to leave, extracting historical data and retraining elsewhere is expensive enough that the firm stays.
On the build side, the dominant failure is budget and timeline overrun. A documented UK financial services build started at £60,000 and three months, and ran to £180,000 and nine months as scope expanded. A UK insurtech build in 2021 to 2022 ran 12 months over and was beaten to market by vendor releases 70 to 80% as capable for 10% of the cost. Model drift, team turnover, and technology obsolescence are common follow-on failures.
What to ask before you decide
Five questions carry the bulk of the decision weight for an SME conversation. Asking them before signing anything will usually make the answer obvious, and the cases where they do not are the ones that warrant a deeper procurement exercise rather than a board debate.
How common is this use case, and how many vendors offer a credible solution. Three or more with reference customers means the buy path is real; zero or one means the market is immature or the use case too specialised. What percentage does the best vendor solve. A good vendor solves 75 to 85%; if your top three requirements are covered, the trade-off is usually acceptable for 90-day time-to-value.
How urgent is the problem, in pounds per month of delay. £50,000 a month of delay makes time-to-value critical. Are there compliance constraints that rule out SaaS. UK data residency, FCA outsourcing obligations under SS2/21, and ICO transfer-impact assessments are the three that change the maths most often. What is the three-year total cost of ownership. Buy looks cheaper at year one and catches up by year three through price escalation; build looks expensive at year one and reasonable by year three if maintenance is budgeted.
The hybrid pattern is the realistic answer for many SMEs. Buy the SaaS platform for the 70% the vendor solves well, build a thin proprietary layer for the 30% where competitive edge sits. Microsoft Power Platform on top of a Salesforce CRM. Custom APIs feeding matching logic into HubSpot. The pattern preserves time-to-value, reduces build risk, and keeps the strategic asset in-house.
Default to buy unless one of the five build conditions applies. Revisit annually; the market moves quickly enough that 2027’s path may differ from 2026’s.
If you want to talk through where your firm sits on these, book a conversation.
