Product engineering ~11 min read
Vertical SMB Software + AI: Ship the Workflow, Not the Demo
What changes when AI is added to a small vertical product for a real business: messy inventory, human confirmation, and a workflow that has to survive Tuesday.
Chapter guide 5 chapters
01 / Premise
A small business does not buy an AI demo
A restaurant needs a menu to stay current. A butcher needs cuts, units and availability to make sense in a message. A greengrocer needs to handle weight, quality and products that disappear during the day. These are not generic “content” problems; they are operational workflows.
Vertical SMB software starts with the sequence a person already has to complete: update an offer, receive an order, confirm the exceptions, and get the work to the next person. AI belongs where it removes friction from that sequence, not where it makes the screenshot look intelligent.
02 / Engineering
Model the awkward nouns before adding the model
- An item may be sold by unit, weight, bundle or portion.
- Availability is temporary and different from a product being deleted.
- A price can be current, indicative or awaiting confirmation.
- A customer message can contain a request, a preference and a delivery constraint.
- A draft generated by AI is not the same thing as a committed change.
These distinctions are product design, not bureaucracy. If the schema cannot represent them, the model will fill the gap with confident prose. The UI then has to teach the business to distrust its own software.
03 / Reliability
Let AI draft; let the workflow decide
A useful boundary is simple: AI can propose a structured action, while the application validates it and a person confirms the facts that matter. A voice note can become a draft menu update. A customer message can become a draft order. Neither should silently publish an unavailable item or invent a total.
The product should make uncertainty visible: show the source text, highlight missing fields, preserve the original message and provide an easy correction path. Observability is part of the feature when a small business cannot afford to debug a mysterious automation at closing time.
04 / Delivery
Ship the narrow loop before the platform
- Choose one role and one repeated workflow.
- Build the smallest end-to-end path without AI.
- Instrument where people correct, abandon or repeat work.
- Add AI to one expensive or annoying input boundary.
- Keep a deterministic fallback and test the edge cases with real operators.
Our own products point to this shape: Platario is a live restaurant-menu example, while MiCarni and MiVerdulería treat catalog and WhatsApp ordering as different vertical problems. The shared infrastructure is useful; the domain rules are not interchangeable.
05 / Fit
The last mile is adoption, not inference
A small business has existing habits, staff turnover, spotty connectivity and a finite amount of attention. A product that requires a new ceremony for every update is not saved by a better prompt. Setup, permissions, mobile ergonomics, export paths and support are part of the technical surface.
The durable advantage of a vertical product is not that it says “AI” on the landing page. It is that the workflow becomes easier to run without hiding the decisions the business still owns. That is a much smaller claim, and a much harder one to fake.