How to build an AI agent for a real business workflow

Pick a job someone already does at least a few times a week, with a clear done state, and with examples you can show a partner. Inbox triage with a draft, invoice field extraction, internal Q and A over a known wiki. Avoid jobs where the expert cannot explain why they chose A over B.

Write the failure cost next to the job. A wrong internal summary is cheap. A wrong customer email is not. That number decides human review, not a taste for autonomy. If the job is really search, ship search. Chatbots vs agents is the fork.

The AI agent development pillar covers architecture. This page is the sequence. Job, data, tools, review, slice. Skip ahead and you will tune prompts for a workflow nobody owns.

List sources: tickets, docs, CRM, product DB, transcripts. Then list what is allowed in a prompt. Customer PII, other tenants, and unpublished strategy docs are common no-go items. Permission is a product requirement. Retrieval that ignores tenancy is a breach waiting for a curious query.

You need examples more than you need a lakehouse. Ten labeled threads beat a vague promise of plenty of data. If the examples disagree with each other, fix the process before you automate it. Models amplify mess. They do not referee it.

If facts live in documents, plan RAG in the slice. If facts live in APIs, plan tools. Most business agents need both. Do not wait for a perfect corpus. Wait for a corpus you are allowed to use and a way to refresh it.

Each agent tool should perform one clear business action with a tight schema, user-scoped authentication, timeouts, and safe retry behavior. Retrying create_ticket must not open several tickets. Put human approval before costly or customer-facing writes, and show reviewers the proposed action plus the evidence behind it.

Human-in-the-loop sits on writes that leave the building or move money. The UI must show the proposed action and the evidence. Slack dumps of JSON train people to rubber-stamp. RelayHub keeps a person on the hard send. Copy that instinct even if your UI looks nothing like an inbox.

MCP can wait until a second host needs the same tools. For a first business agent, native function calling inside your app is enough. Standardize later, when the verb list is stable.

A useful first release covers one job with a small tool set, retrieval when needed, a review queue, logging, and a kill switch. Test it in staging against anonymized cases and a repeatable evaluation set. Save broad platforms, custom models, and multi-agent designs for evidence-backed needs.

Measure tool errors, human edit rate, latency, and spend. If edit rate stays high, the job or the tools are wrong. If spend spikes, the loop is chatting with itself. Cost ranges exist so you can tell a slice from a program before you start.

Algo Vortex ships this as product work under AI development. Bring the job and the examples to contact. We will push back if the first slice is still a department-wide platform in disguise.