How to create an AI agent in an SME (without picking the wrong target)

Process before the model

Before you talk about models, look at the process. If you cannot measure time lost, the agent will not help.

An agent in practice: an LLM that generates tokens in a tooled loop (read, write, call APIs). As the talk goes, tokens and tool results fill the context window. Without designing that context and rights, quality drops (see context engineering).

Three bricks: model, tools, control policy. Without the third brick, you just have a POC / demo.

It is not a chatbot glued on the marketing site. It is a digital teammate bounded to a flow, with logs and escalation.

Pick the first flow that pays

Filter: frequent (several times a week), time-expensive, fairly clear rules, acceptable error or simple human validation.

Cases that often work in service SMEs: inbound qualification (form + email) with reply/quote draft; quote/invoice follow-ups; tier-1 support on a closed knowledge base. Measurable volume, natural escalation.

Avoid day one: open sales strategy, automatic SEO, undocumented multilingual support, payments/HR without a legal frame.

Scope in 40 minutes: current path (who, tool, duration), exceptions, weekly volume, definition of done, one KPI (minutes saved, leads under 2h, escalation under 20%). If you cannot name the KPI in one sentence, you are still exploring.

Minimum architecture

Five layers: input (webhook, email, message); context normalization; generation (LLM); tools (API, MCP if needed, bounded index); controlled output (draft, action, log).

Tools: short allowlist (read ticket, search docs, draft, create task). Destructive writes behind confirmation. 3 to 7 tested tools beat a catalog of 40.

MCP standardizes wiring; it is not magic without a stable API. Detail in the Model Context Protocol article.

Observability from sprint 1: log input, tools, versioned prompt, output, human decision.

HITL and measure before multi-agent

Modes: suggestion, assisted (one click), bounded autonomy (green cases only). Raise the dial after calibration on 50-100 cases. Detail in the human in the loop article.

Minimum KPIs: volume/week, median time before/after, escalation, post-validation reject, token cost per ticket, incidents.

3-4 week pilot: real corpus, 1-2 power users, weekly amber/red review, feature freeze. Go/no-go before stacking multi-agent.

Multi-agent only if a single agent becomes unreadable. Otherwise you stack complexity without ROI.

If you want to scope a first agent on a real process, we can do it in 20-40 minutes.

Common mistakes

Site chatbot with no goal. Wire everything day one. No business sponsor. Unversioned prompts. Scale before quality. Ignore compliance (PII, hosting, log retention).