Buy or build your AI? Choosing the platform for an SME

When an SME decides to adopt AI on a process, the question it hears most often is "which software do I buy?". But before it comes one that weighs more, and that almost no one asks: is it better to buy a ready-made platform or build something bespoke? The answer changes timelines, costs and risks far more than the choice between two competing tools. Here we try to answer with a criterion, not on instinct — and then we look at the landscape of the tools you actually "buy" with.

Buy or build: the question before the question

The most-used build-vs-buy frameworks in 2026 have you assess a use case on three axes:

• Uniqueness — how specific the flow is to your company, rather than generic and already solved by a thousand others.
• Data sensitivity — can an external vendor touch this data, or must it stay under your control?
• Strategic value — is it a competitive differentiator or a back-office activity you just want done well?

The reading is simple: high on all three → building makes sense. Two or more axes low → buying. Most of the flows an SME wants to automate first — ticket triage, sales follow-ups, reconciliations, reporting — are generic, low-sensitivity and back-office. In other words: to buy.

Buying wins on speed: managed platforms go live in days or a few weeks, and the return shows in 1–6 months. Building wins over the long term but is slow — 12–24 months for ROI — and requires a minimum team (an AI engineer, a backend developer, a product owner, and for an agent in production also someone covering DevOps and evaluation). Almost no SME has that team in-house.

There's also a useful volume threshold to keep in mind: below about a million interactions a year, buying almost always wins for integration speed; only beyond that volume does the per-token economics of building start to pay off. Very few SMEs cross that line — and those who do are no longer an SME on that process.

The middle way that in 2026 is almost the standard: the hybrid

The choice isn't binary. The model most serious implementations adopt today is hybrid: you buy the language model (OpenAI's or Anthropic's APIs) and you build only the orchestration and integrations on top of an open-source framework. It cuts timelines by 40–50% compared with doing everything from scratch, but it keeps control over the business logic where it should be: on your side. It's the shape on which, in our experience, almost every project should land: buy the model, graft in the logic.

The platforms you "buy" with

When the use case is one to buy, the market of automation and agent tools narrows to a few horizontal platforms. There's no "best" one in absolute terms: there's the right one for your stack and your volume. Here's how they differ.

Make — the best operations/price ratio

A visual scenario builder, over 3,000 connectable apps, with an agent builder in beta. Its strength is economic: the entry plan covers tens of thousands of operations for a modest spend, and it stays the cheapest of the three even as volume grows. It's the default choice when the flows are many and frequent and cost per operation matters.

Zapier — the widest coverage

The most widespread "zap" builder, with a natural-language copilot and autonomous agents that act across apps. Its merit is the integrations: over 8,000, the largest estate. It should be chosen when the client's stack is made of niche, long-tail software — not when the activity volume is high, because it's the model that scales worst on cost as operations grow.

n8n — the deepest agentic capability

Node-based and self-hostable; version 2.0 (January 2026) brings native LangChain integration, about 70 AI nodes, persistent memory, nodes for RAG and vector-DB and human-in-the-loop patterns. It's the right choice when you need a genuinely agentic flow — not just event-triggered — and when you can host the platform in-house, keeping the data under control. It's the point where "buying" blurs into the hybrid.

Microsoft Copilot Studio and Salesforce Agentforce — only if you're already inside

Copilot Studio (consumption-priced agents inside the Microsoft 365/Azure ecosystem) and Agentforce (licensed agents on top of the Salesforce CRM) make sense in one case only: if the company already lives on that platform. Data, runtime and model choice stay tied to the vendor's cloud, and the price threshold — per seat or per conversation — starts too high for the budget of an SME that isn't already a customer. Outside that stack, they rarely balance the books.

How you pay matters more than the list price

List prices change; the pricing model stays, and it's the real decision lever. Before looking at the figure, look at how you pay, because that's what determines whether the cost will explode with growth:

• Per operation or per task (Make, Zapier) — predictable at low volumes, to be modelled carefully if the automations multiply.
• Self-hosted (n8n) — the cost is the infrastructure and whoever manages it, not a consumption licence: it pays off at high volumes and when the data mustn't leave.
• Per message (Copilot Studio) — rewards short actions, punishes the long, elaborate conversations.
• Per conversation or per seat (Agentforce) — the opposite: it handles multi-step interactions well, large volumes of micro-requests less so.

The practical rule for an SME: Make and n8n are the sustainable default — a modest monthly spend, or just the cost of hosting — consistent with the idea of bringing AI into the company without a big consultancy's bill. Copilot Studio and Agentforce come into play only when the client is already inside that world.

When it's worth building instead

If a use case is high on uniqueness, data sensitivity and strategic value, buying isn't enough. But even then, "building" in 2026 doesn't mean starting from a blank page: it means taking the open-source scaffolding that's now standard for agent orchestration — LangChain/LangGraph, CrewAI — and resting it on a model bought via API, instead of training your own. It's the concrete form of the hybrid above, and it's also the only way an SME can afford the "built" option: if someone provides that team as a graft, instead of having to hire it.

Here data sensitivity becomes the constraint that decides everything. If the flow touches personal or regulated data, the choice of platform isn't just technical: it's a compliance choice. It's worth passing it first through the compliance overlay — risk level, DPIA, where the data lives — and then deciding whether and what to self-host.

How to choose, in practice

Summarising the path, without shortcuts:

1. Assess the use case on the three axes (uniqueness, sensitivity, strategic value). Two or more low → buy.
2. If you buy, choose the platform from the pricing model and from your stack, not from the list price: Make for cost per operation, Zapier for coverage, n8n for agentic depth and data control.
3. If you build, think hybrid: buy the model, build only the orchestration — and consider who provides you the team.
4. In both cases, pass the sensitive data through the compliance controls first, not after.

It's exactly the method we work with: first understand where you are and what's worth automating, then choose the right platform — bought, built or half-grafted — with the controls around it. Not a tool imposed from above, but a graft that takes on your stack.

Don't know which process to start from, even before choosing the tool? The AI-readiness assessment gives you a first indication in two minutes; and on the Services page you'll find, department by department, what a concrete AI Workflow Design looks like.

The build-vs-buy criteria cited synthesise public 2026 frameworks (including Composio and Pharos); the platforms' features and list-price figures are self-reported by the respective vendors (Make, Zapier, n8n 2.0, Microsoft Copilot Studio, Salesforce Agentforce), vary over time and should be read as indicative — not as a quote. This article is for orientation and does not constitute legal advice or a compliance assessment.