Automations that read documents, make a decision, and act — from a single n8n workflow to an autonomous agent running a whole process.

Most workflow automation stops where the messy part begins: a PDF lands in an inbox and a human still has to open it, read it, and decide what happens next. Putting a model inside the workflow closes that gap. Most AI agent projects fail the opposite way: they work in a demo and quietly make expensive mistakes at volume. Everything here is built for the day after launch — deterministic first, model second; a real escalation queue; an audit log a non-engineer can read.

Automation Audit — $1,200, five days. For when you know you want "AI" but can't yet say where. I map how work actually arrives, who touches it and where it stalls, then sort every workflow into three buckets: safe to hand to an agent today, automatable with plain code and no model at all, and leave alone for now. You get a build plan with a cost per workflow, whether or not you hire me to build it. If you do, the fee comes off the build.

n8n Workflow with AI built in — from $2,500, two weeks. Document intake from email, forms or uploads; extraction into structured data your systems can use; routing and approvals with a human step wherever the decision warrants one; replies that read like a person wrote them. I built one that watches a Gmail inbox, reads PDF attachments with Claude's vision (scans and phone photos work the same as typed files — no OCR stage to maintain), and replies in-thread with a summary, line items and action items. Typically two to four workflows of moderate complexity.

Autonomous Agent for one business workflow — $7,500, five weeks. Not a chatbot and not a demo. The agent receives real work, does the job, and only escalates what it genuinely can't settle. Week 1 scopes one workflow precisely and builds an eval set from your history; weeks 2–4 build integrations, tool loop, guardrails and monitoring; week 5 runs it shadowed against live work before handover. In production I've built an order-intake agent that ingests emailed purchase orders, matches every line against the catalogue, allocates stock or raises supplier orders, and emails back an ETA — no human on routine orders.

Deterministic first, model second. Rules and exact-match lookups resolve ahead of the LLM, so common cases never depend on a generation being correct. Also the single biggest lever on cost.

A real escalation path. Anything ambiguous lands in a queue for a person instead of being guessed at. Every decision is logged, including the ones a human overrode.

Evals and cost controls. A test set drawn from your real historical cases, plus per-run cost tracking and ceilings.

Stack. n8n by default (self-hostable, no per-task pricing) — Make or Zapier if you're committed. Claude or OpenAI. Node.js or Python, PostgreSQL. Integrates over REST, webhooks and OAuth with CRMs, ERPs, e-commerce, email and Microsoft 365.

Who this is for. Operating businesses with real volume and a manual process someone is tired of. Not a fit if you're pre-revenue or want a chatbot on a marketing site.