How We Built an AI Automation System That Saved 40 Hours a Week

The Client and the Problem

Our client is an 18-person professional services firm based in Manchester providing consulting services to mid-size businesses across the UK. They had grown quickly — from 6 to 18 staff in two years — but their operational processes had not kept pace with that growth.

The founding team was spending an increasing proportion of their time on administrative work rather than client delivery. A time audit revealed the scale of the problem:

53 hours per week across the team — the equivalent of 1.4 full-time employees — spent on tasks that were repetitive, structured, and well-suited to automation.

The Solution Architecture

We designed a five-workflow automation system built on a shared infrastructure layer:

• n8n (self-hosted on a £15/month VPS) as the orchestration layer for all workflows
• GPT-4o as the primary reasoning model for all language tasks
• Whisper API for meeting audio transcription
• HubSpot API for CRM operations
• Xero API for accounting / invoice processing
• Google Drive API for document storage and retrieval
• Pinecone vector database for proposal knowledge base
• Slack for notifications and human review requests

Workflow 1 — Enquiry Triage and Response

Before: 3 staff rotated inbox monitoring, manually reading and forwarding enquiries, drafting replies, and updating HubSpot.

After: New emails trigger an n8n workflow that reads the email via Gmail API, sends it to GPT-4o for classification and reply drafting (with context from HubSpot and the company knowledge base), creates a HubSpot record, and posts the draft reply to a Slack channel for 1-click approval.

Result: Triage became zero-touch. Approval and send takes under 30 seconds per email. 14 hours/week reduced to 2.5 hours/week (review and approval only).

Hours saved: 11.5 hours/week

Workflow 2 — Proposal First Drafts

Before: Senior consultants spent 3–6 hours writing each proposal from scratch, referencing past proposals manually.

After: A consultant fills in a brief 10-field intake form (client name, problem, scope, budget range, timeline). This triggers a workflow that retrieves the 5 most similar past proposals from Pinecone, sends them alongside the intake to GPT-4o, and generates a 1,500-word first draft in the firm's house style, saved to Google Drive. The consultant reviews, refines, and sends — not writes from blank.

Result: First draft generation: 4 minutes. Total proposal time: down from an average of 4.5 hours to 1.2 hours. 12 hours/week reduced to 3 hours/week.

Hours saved: 9 hours/week

Workflow 3 — Invoice Processing

Before: Accounts admin manually opened each invoice PDF, extracted supplier name, amount, date, cost code, and entered into Xero.

After: Invoices emailed to a dedicated address trigger a workflow that extracts the PDF, sends it to GPT-4o Vision for field extraction, matches the supplier against the Xero contacts API, and creates a draft bill in Xero — flagging any that cannot be matched for human review.

Result: 92% straight-through processing rate. 8 hours/week reduced to 0.7 hours/week (reviewing flagged items only).

Hours saved: 7.3 hours/week

Workflow 4 — Weekly Client Reports

Before: Each project manager spent 2–3 hours writing a status update, pulling data from multiple sources, and formatting a Word document to send to the client.

After: Every Friday at 3pm, n8n triggers a workflow per active project. It pulls the week's completed tasks from the project management tool, retrieves any relevant notes from CRM, and generates a formatted status report in the client's preferred template via GPT-4o. Reports go to a review queue; the PM approves or edits and sends with one click.

Result: Report preparation reduced from 2.5 hours average to 25 minutes (review and occasional edits). 10 hours/week reduced to 1.7 hours/week.

Hours saved: 8.3 hours/week

Workflow 5 — Meeting Notes and CRM Updates

Before: After every client call, consultants manually wrote meeting notes, extracted action items, and updated HubSpot — averaging 20–35 minutes per meeting.

After: Calls are recorded via Google Meet. The recording is automatically sent to Whisper API for transcription, then to GPT-4o which generates a structured summary: key decisions, action items with owners, next steps, and client sentiment. Summary is posted to Slack for confirmation, then pushed to HubSpot and saved to Drive.

Result: Post-meeting admin reduced from 25 minutes to 3 minutes (reviewing the auto-summary). 9 hours/week reduced to 1.1 hours/week.

Hours saved: 7.9 hours/week

Results Summary — 3 Months After Launch

What We Learned

• The proposal workflow had the highest perceived value. The actual hours saved were not the largest, but the quality uplift and reduced senior consultant burden made it the most celebrated change internally.
• Human review steps were not a compromise — they were a feature. Building in 1-click review for all client-facing outputs meant the team adopted the system immediately. Nobody felt their judgment was being bypassed.
• The knowledge base was the foundation. The proposal and enquiry workflows only worked because we spent the first week building a high-quality vector store from 3 years of past work. Garbage in, garbage out.
• n8n self-hosted was the right call. At 5 concurrent workflows and moderate volume, the £15/month VPS cost versus a cloud n8n plan saved approximately £200/month and gave full control over logging and error handling.

Related Services

Related Articles