Every SMB CEO in 2026 has been pitched an "AI agent." Most of those pitches are wrong for SMB economics — they assume you have a 10-person ML team and a clean data warehouse. You don't. Neither do most of our clients.
Here's the playbook we use to actually ship AI that pays for itself within a quarter.
The 5 questions
1. What's the cost of a wrong answer?
This is the single most important question, and 90% of AI sales pitches skip it.
If the cost of a wrong answer is high and undoable (legal contract, medical advice, financial trade), agents must be assist-only — they draft, a human approves. If the cost is low and reversible (categorise an incoming email, suggest a tag, summarise a meeting), agents can run autonomously.
We've never shipped an autonomous agent that touched money, legal documents, or patient records. That's not the LLM's fault — it's a risk-budget decision.
2. Is this a workflow or a question?
LLMs are excellent at questions ("what does this contract say about indemnity?"). They're mediocre at workflows ("draft the renewal, get it signed, file it"). Most "AI agent" projects fail because they try to do workflows when they should be answering questions.
Our heuristic: if the workflow has more than 3 deterministic branches, build it with traditional code + LLM at the question steps, not as an agent with tool calls.
3. What's already in the CRM/ERP?
The cheapest AI win is structured data you already own that nobody's reading. Sales call transcripts. Support ticket history. Old contracts in a SharePoint folder.
A simple RAG (retrieval-augmented generation) layer on top of this is usually a 2-week build, and it pays for itself in month 2.
4. What's the latency budget?
Real-time chat = sub-2-second response. Email triage = sub-30-second. Nightly report = "anything under an hour."
Latency budget determines whether you use a fast small model (Haiku, GPT-4o-mini), a powerful slow model (Sonnet, GPT-4o), or a fine-tuned local model. Don't pick the model first.
5. Who owns "the agent went wrong"?
Define this in writing before launch. Without a named owner, the project quietly dies the first time it produces a bad output, because nobody knows whose job it is to investigate.
A real example: MediConnect AI Triage
For MediConnect we built an AI triage layer that reviews patient symptom intake forms and suggests (never decides) a clinical priority.
- Cost of wrong answer: high → assist-only ✓
- Workflow vs question: question (rank these 12 symptoms) ✓
- Existing data: 18 months of triage decisions to fine-tune on ✓
- Latency budget: under 5 seconds — small model fits ✓
- Owner: Chief Medical Officer, with weekly review meeting ✓
Result: 60% faster triage, zero misclassified critical cases in 6 months of production.
The 80/20 stack we recommend in 2026
For SMBs that want AI without becoming an ML company:
- Claude Sonnet 4.6 or GPT-4o for question-answering
- Claude Haiku for cheap high-volume classification
- LangChain or LlamaIndex for retrieval — but only if you genuinely need it
- Postgres + pgvector for the vector store (one less thing to run)
- Trace everything with Langfuse or LangSmith from day one — you cannot debug AI without traces
What this is not
This playbook will not turn your business into an "AI-first company." It will turn a handful of specific workflows from manual to assisted. That's the realistic win in 2026, and it's plenty.
If you want a written assessment of which workflow at your company is the right first AI bet, book a 30-min call — we'll come back in 24 hours.