A 25-staff legal services firm bought into Zapier in early 2024. By mid-2025 they were running 18 zaps across onboarding, document routing, invoicing and CRM updates. The bill had climbed from £30 a month to £1,200, and two zaps were business-critical and breaking under load. The IT lead proposed migrating those two to a custom Python service on a £30 a month server. The owner asked whether they could just stay put.
Yes, technically, but the per-task cost would keep climbing and the broken-zap problem would not go away. The firm migrated those two flows in three weeks, kept the other 16 on Zapier, and dropped the bill to £350 a month. The right opening question was never “no-code or custom.” It was which zaps had crossed the line.
The choice you’re facing
No-code AI means platforms where business users wire AI into workflows without writing code. Zapier, Make.com, n8n, Microsoft Power Platform and Copilot Studio sit here, priced £9 to £750 a month at SME tiers depending on volume. Custom build means contracted developers writing code against AI APIs from OpenAI, Anthropic or Vertex, or open-weight models you host yourself. UK mid-level developer day rates in 2026 run £450 to £700.
No-code platforms have absorbed enough of the integration burden by 2026 that scenarios needing five or six tool connections now run on off-the-shelf platforms. The question is rarely “can we do this in no-code”. It is usually “should we accept the constraints of the platform, or commission custom work to remove them.”
When no-code is the right answer
No-code is the right answer for the first 12 to 18 months of any new AI workflow, for high-velocity integration with standardised outcomes, for content generation and classification at moderate volume, for rapid prototyping where the workflow value is unproven, and for owner-operated maintenance in firms with no in-house developers. Below the migration thresholds, this is uncontested.
The integration case is the easiest. A firm receiving enquiries through a web form that needs a record in Salesforce, an invoice in Xero and a confirmation email has a use case Zapier solves natively. A Professional plan at £49 a month plus a few hours of configuration runs it indefinitely. Custom development runs £3,000 to £8,000 plus maintenance.
Prototyping is where no-code earns its place even when custom would be cheaper at scale. An operations lead testing whether automated lead scoring lifts conversion can stand up a Make.com workflow in a week and run it for a quarter. Custom prototyping runs £5,000 to £10,000 over six weeks, which kills the experiment before it starts.
Constrained technical resource is frequently the decisive case. By end-2026, around 80 percent of low-code and no-code users sit outside IT departments. For an operations manager automating their own workflow, the platform removes the developer-resource bottleneck. Autonomy is often worth more than unit-cost optimisation.
When custom build is the right answer
Custom build is the right answer when one of five conditions holds: proprietary business logic that no platform encodes, high-volume or latency-critical processing, deep integration with internal or legacy systems, regulatory data sovereignty that rules out third-party SaaS, or business-critical workflows where platform discontinuation creates unacceptable risk.
Proprietary logic is the cleanest case. A logistics firm with a bespoke routing algorithm cannot expose it through a no-code connector without losing the confidentiality that makes it a competitive asset. A professional services firm fine-tuning a model on past engagements also needs custom build. UK 2026 budgets run £20,000 to £80,000.
High volume is the financial threshold. A firm routing 10,000 queries a day through Zapier into Claude pays around £4,500 a month in Zapier fees alone before inference. A custom service running directly against the Claude API on serverless cloud functions costs a fraction of that. Above 50,000 requests a month, custom usually pencils.
Sovereignty is the regulatory threshold. UK financial services firms under FCA AI guidance, NHS Trusts, and any business with customer contracts that forbid US cloud processing land here. The ICO requires a Transfer Impact Assessment when personal data flows to US-based SaaS, and the EU AI Act in force from August 2026 raises the evidence bar for high-risk AI deployers. Where data cannot transit a third-party platform, custom infrastructure is often the only path.
Platform risk is the operational threshold. Zapier ran 14 hours of cumulative downtime in 2023 without SLA credits. Pricing changes, deprecations and roadmap pivots are normal vendor behaviour. For a workflow that generates revenue, that exposure is hard to live with. Custom shifts vendor risk onto components the firm can swap.
What it costs to get wrong
Both directions have a failure mode and they look opposite to each other. The no-code trap is tool sprawl with surprise bills and vendor lock-in. The custom-build trap is budget overrun, maintenance debt and developer dependency. A firm can land in either, or in both if the migration is done badly. Knowing which trap each path opens up is much of the work of avoiding it.
The no-code version starts simply. A firm runs Zapier, adds Make.com for content workflows, layers Notion AI and a few custom GPTs for specific tasks. Eighteen months in, no single person owns the integration strategy, workflows live in silos, and the combined bill exceeds a unified custom system. Research on AI vendor lock-in found only 42 percent of organisations that attempted a migration reported it went smoothly. Remediation typically runs £30,000 to £80,000 of pure rework.
The custom-build version starts with optimism and ends in maintenance. McKinsey research found software projects average 45 percent cost and schedule overruns. A £30,000 build often lands at £45,000 to £55,000 once data is messier than expected and the legacy integration is harder than scoped. Maintenance runs another 15 to 30 percent of the build cost a year. If the contractor moves on, the next developer has to read the system before they can change it.
Premature migration is the subtle version. A workflow validated for three months on Zapier rarely justifies a 24-month infrastructure commitment. Migrate only when volume has held above the threshold for three consecutive months.
What to ask before you decide
Five questions, in order, before signing anything. Who owns the workflow long-term, what happens when the platform changes, where does the data live, what is the exit cost if it does not work, and how will success be measured. The order matters. Maintenance ownership and platform risk surface the wrong answers before the budget and tooling questions ever come up.
One: who maintains the workflow long-term? An operations lead favours no-code. An in-house developer or retained agency favours custom. If the answer is “we will work that out later”, the answer is no-code.
Two: what happens when the platform changes? Ask vendors about deprecation policy, pricing notice periods and customer recourse. Get it in writing or assume worst case. For custom, ask whether the contractor offers a maintenance retainer covering dependency updates.
Three: where does the data live? For no-code, ask which jurisdictions store the data and whether UK or EU residency can be specified. For custom, ask which cloud provider hosts the system and whether on-premises is feasible if a regulator demands it.
Four: what is the exit cost? For no-code, ask whether workflows export in a standard format. For custom, ask whether the codebase is documented and portable. The hybrid pattern of no-code orchestration over a custom AI service is partly an exit-cost play, because the integration logic stays portable.
Five: how will success be measured? Define time saved, cost reduced or error rate improved before you commission anything. The common failure mode is deploying a system, finding the benefit is half what was predicted, and realising too late.
The honest answer for UK SMEs in 2026 is start on no-code, design for the migration from day one with portable data formats, and move the workloads that cross the thresholds. The right architecture is rarely all of one or the other.
