Strategy
Buy, build, or assemble: the real AI decision facing the SME
The old procurement question was buy or build. For AI in a mid-sized business, both answers are usually wrong.
20 July 2026 · AI Labs
Every SME leadership team eventually has the meeting. Someone has seen what the new AI systems can do. Someone else has a quote from a vendor. And the conversation settles into the shape it has had for thirty years: do we buy something off the shelf, or do we build our own?
For AI, this framing quietly fails, because both classic answers carry assumptions that no longer hold.
Why pure "buy" disappoints
Off-the-shelf AI products are genuinely good now, and for well-bounded functions the right product is the right answer. The disappointment comes at the seams. An AI tool that does not know your pricing rules, your approval chains or your customer history produces generic work, and generic work is precisely what AI has made worthless. The value of AI in a business is concentrated in its contact with that business's specifics. A shelf product, by definition, ships without them.
Why pure "build" breaks
The opposite instinct fails differently. Training or heavily fine-tuning your own models, standing up your own inference, hiring the team to tend it: for an SME this is a capital project with the depreciation curve of a fishmonger's stock. The frontier moves every few months. Whatever you build at the model layer is competing against the best-funded research organisations on earth, and it will lose on a schedule you can predict.
The third answer
The posture that actually works is assembly. Rent the intelligence; own the workflow.
The intelligence layer, the models themselves, is becoming a commodity you buy by the token, and it improves every quarter without you lifting a finger. The workflow layer is the opposite: your processes, your data, your approval rules, your tone with customers. Nobody sells that, because nobody else has it.
Assembly means building the thin, durable layer that connects the two:
- Your data, organised for machines. Not a data lake project. The practical version: the documents, records and rules an agent needs, in a place it can read, with the sensitive parts governed.
- Your workflows, made explicit. Agents execute what can be described. The act of describing a process precisely enough for an agent is itself where much of the value appears.
- Your controls. Which actions need a human signature. What gets logged. Who reviews. This layer is yours to define and it is the part that makes AI safe to point at real work.
- Swappable intelligence underneath. The assembly layer should treat the model as a component. When a better one ships, and it will, you change a configuration, not an architecture.
Notice what this does to the economics. The expensive, fast-depreciating part is rented. The cheap, slow-depreciating part, the encoding of how your business works, is owned. That is the correct side of the ledger for each.
What assembly looks like in practice
In our consultancy work the successful pattern is consistent. Start with one workflow that hurts: quoting, onboarding, reconciliation, inbound triage. Assemble an agentic solution around it with the controls visible from day one. Put it in production with a human reviewing the output. Measure honestly. Then, and only then, widen.
The failed pattern is equally consistent: a platform decision made before a workflow decision, a large licence bought before anyone has specified what the software should actually do, or a moonshot internal build that treats the model as the product.
The SMEs that win with AI over the next three years will not be the ones that bought the most or built the deepest. They will be the ones that understood which layer to own. Own the workflow. Rent the rest.
Related work
Saint AGI
Agentic AI platform
Role-specific AI agent teammates for finance, sales, recruiting, marketing, support and operations, plugged into the tools a business already runs on, with approval gates and audit trails built in.
Kumo
HR technology
An HR platform with an agentic AI colleague that executes HR work rather than suggesting it: hiring, payroll, leave, performance and learning, with a full execution log behind every action.
Talking beats reading
If any of this maps onto a decision you are facing, a short conversation with the team that wrote it is free.
Start a conversation