Strategic AI Integration in Core Business Logic
Putting LLM pipelines inside legacy infrastructure without setting it on fire — patterns from agents running in production today.
by Yassine Emhamed // field notes
The AI agents I've shipped don't demo well. They route restaurant orders, chase dental no-shows, and assemble client onboarding paperwork. That's the point. The gap between an impressive AI demo and AI in core business logic is the part nobody live-streams: deciding what the model is allowed to touch, what happens when it's wrong, and where the humans stay in the loop.
02 // Give the model a job description, not a personality
Maison Lune's chatbot dispatcher works because its scope is brutally narrow: answer the questions staff were answering by hand, and route orders. Within that scope it now handles 80% of inquiries with zero staff intervention and cut order response time roughly 40%. The temptation is always to widen the mandate — let it discuss the wine list, let it improvise. Every widening multiplies the failure surface. Narrow agents compound; clever agents combust.
The same rule shaped SmilePro's reminder system. The AI doesn't 'manage patient relationships' — it runs SMS and email sequences against booking states. No-shows fell from 28% to 9% in six weeks, and the front desk got two hours a day back. Boring mandate, measurable outcome.
03 // Deterministic rails, probabilistic passengers
Every production AI system I've built follows the same shape: the workflow engine (n8n, mostly) is deterministic — triggers, branches, retries, audit logs — and the LLM rides inside it as one step among many. The onboarding pipeline that took ArcaneDot's client setup from 4 hours to 8 minutes is mostly not AI: it's contract generation, folder creation, CRM entry, Slack setup, scheduling. The model handles the parts that need language; the rails make sure a hallucination can't sign a contract.
Legacy infrastructure makes this non-negotiable. You don't rip out the system of record to add intelligence — you wrap it. The CRM stays the authority; the model proposes, the workflow disposes.
04 // Agents reviewing agents
The site you're reading was built with a daily multi-agent workflow — Claude Code, OpenAI Codex, and Gemini-based tooling, coordinated through shared state files in the repo, each agent logging handoffs the others read. The practical lesson generalizes past code: one agent's output is another agent's input to verify. Cross-checking between independent agents catches what a single model confidently misses.
That's also the honest answer to 'will AI replace the engineer?' Someone still has to design the rails, define the mandates, and own the failure modes. The leverage is real — the accountability doesn't transfer.
Ship narrow agents inside deterministic workflows, keep the legacy system as the authority, and measure in business numbers — response time, no-show rate, hours returned — or you're not integrating AI, you're decorating with it.