Example engagements

Every morning two analysts match payment-provider settlement reports against bank statements and the internal ledger in spreadsheets. About 4% of lines do not match and each takes 10-20 minutes to trace. Month-end close slips by two days.

n8n self-hosted in the client's AWS account (EU region) pulls settlement files from two payment providers by SFTP and API, and bank statements in camt.053 format. A Python matching service applies deterministic rules first (amount, reference, date window). Unmatched lines go to Claude Sonnet (Anthropic API under the client's own key, zero-retention settings requested) with the line and candidate matches; the model proposes a match and a reason, never posts. An analyst approves or rejects in a small review screen; every decision is logged with the model's reasoning. Daily summary to Slack.

n8n self-hosted on AWS (EU region), Python matching service, Claude Sonnet under the client's own API key, Slack

Systems Project, 7 weeks, after a 2-week audit

85-95% of lines matched automatically by rules, 60-80% of the remainder resolved by one-click approval, analyst time on reconciliation down from about 30 hours to 8-12 hours per week, close back on schedule.

“We still sign off every exception. We just stopped hunting for them.”

Invented for this scenario, attributed to a head of operations.

Coordinators open each email, read bills of lading, commercial invoices and packing lists, and re-key 20-30 fields per shipment into the transport management system. Around 600 documents a week, with typos that surface at customs.

Make scenario watching the Microsoft 365 shared mailbox, attachments sent to Azure AI Document Intelligence for layout and table extraction, then GPT-4.1 via Azure OpenAI in the client's tenant maps fields to the TMS schema and flags low-confidence values. Shipments with all fields above the confidence threshold are created through the TMS API as drafts; the rest go to a review queue in the TMS with the source page highlighted. Weekly accuracy report from a 200-document evaluation set.

Make, Microsoft 365, Azure AI Document Intelligence, GPT-4.1 on Azure OpenAI, the client's TMS API

Build Sprint, 4 weeks, plus one extra integration

70-85% of documents created as drafts without manual typing, re-keying time down 50-65%, field error rate below the manual baseline measured in the audit.

“The team checks shipments now instead of typing them.”

Invented for this scenario, attributed to an operations director.

Tenant requests come by portal, email and text. A coordinator reads each one, decides urgency, picks a vendor and replies. After-hours emergencies (leaks, no heat) wait until morning unless a tenant calls the on-call line.

n8n Cloud flow collecting requests from the property management system's API, a Twilio SMS number and a shared inbox. Claude Haiku classifies category and urgency against the client's written policy and drafts a reply; emergency keywords and any low-confidence case page the on-call manager through Twilio. Routine jobs are created as work orders with a suggested vendor; a coordinator approves in one click. Every tenant-facing message states that it was drafted by an automated assistant.

n8n Cloud, the property management system API, Twilio SMS, Claude Haiku

Build Sprint, 4 weeks; Run Standard afterwards

First response to tenants under 10 minutes for 80-90% of requests, emergency escalation under 5 minutes around the clock, coordinator triage time down 40-60%.

“Nights stopped being the time things go wrong.”

Invented for this scenario, attributed to a director of property operations.

Demo requests and contact forms land in HubSpot unqualified. Reps answer within 4-6 hours on average and spend time on students, competitors and tiny accounts. Monday pipeline reporting is assembled by hand.

Custom Python service on the client's Google Cloud project, triggered by HubSpot webhooks. Enrichment from the company domain via a data provider the client already licenses, then GPT-4.1 mini scores fit against a written ICP and drafts a first reply in the rep's voice for approval. Qualified leads are assigned round-robin with a Slack alert; poor-fit leads get a polite automated reply with self-serve resources. A Looker Studio report replaces the Monday spreadsheet.

Python on Google Cloud, HubSpot webhooks, GPT-4.1 mini, Slack, Looker Studio

Build Sprint, 3 weeks, then Run Basic

Median first response under 15 minutes in working hours, 25-40% of rep time on unqualified leads returned, weekly report fully automated.

“Reps open their day with leads that are worth a call.”

Invented for this scenario, attributed to a VP of sales.