One email to the courier saved £1,931.
AI operations for Omnia, a London Notting Hill Carnival mas band. Through the 2026 season, Claude worked as an operator inside the band's live systems — orders, courier, inbox, payments, books and the band's website. Not a chatbot on the side: an extra pair of hands with a strategist attached.
Nobody had noticed London parcels were billed as the remotest zone in the country.
Last year's courier invoice charged every London parcel at the most expensive rate. One email got the official rate card. That's the pattern of this engagement: the biggest wins weren't automation, they were catching what the systems got wrong before a customer or an invoice did.
Cost of one dispatch run
Same parcels, two rate cards
£1,931.13 saved on a single run. Merging parcels to shared addresses (135 down to 115) saved a further £131.86.
One season, in numbers.
Orders posted, reconciled to the label
295 items and 133 labels, every one accounted for.
Courier overcharge avoided
On a single dispatch run, from one email asking for the rate card.
Phantom courier bookings traced
Every one matched to a real parcel. Zero lost.
Emails turned into a service playbook
149 answers across 30 question types.
Customers not wrongly contacted
A false “missing size” alarm on 176 of 738 items, defused before anyone hit send.
Sponsor decision-makers contacted
A record-label partnership and 1,500 units in kind.
Seasons of galleries back online
2019 to 2026, on a branded site that's live all year instead of a blank page between carnivals, with a 2027 list collecting sign-ups.
One immovable deadline, and everything piles up in the four weeks before it.
738 costume items across seven sections, some sold by the band and some by independent section leaders taking payment into their own Stripe accounts, then packed, labelled, posted or handed out in person.
A small team, and systems that didn't talk to each other. Customer records sat in at least four places with no shared ID. One section's orders had gone unprinted from April to August because nothing checked every section was covered.
- Orders
Playmas, a carnival platform, after a 2025 season on Shopify.
- Delivery
Shiptheory and a courier account.
- Money
Stripe, bank exports, and bookkeeping that had fallen behind.
- Customer service
A shared info@ inbox on Zoho Mail.
- Everything else
A Dropbox folder of spreadsheets, PDFs and exports.
Thirteen tools across three areas, one operator in the middle.
Nothing was ripped out. The AI worked inside the tools the band already used, and added one, Claude Design, for the website.
Orders & delivery
Costumes from order to doorstep
- Playmas
- Shiptheory
- Courier account
- Python scripts
- Dropbox
Money
Takings, payouts, books
- Stripe
- Bank exports
- Bookkeeping
Customers & growth
Inbox, emails, sponsors, website
- Zoho Mail
- Brevo
- Shopify
- Claude Design
Claude, the operator
Connectors where they exist; otherwise a browser session the owner logged into, or scripts on the shared Dropbox folder.
- Owner logs in
- Rollback before bulk changes
738 items posted right
Packing pipeline, section-coverage guard, phantom bookings traced to zero lost.
A playbook from 8,140 emails
149 answers, each with a send policy: auto, draft for a person, or human only.
Costs caught, partners signed
£1,931 courier saving, two years of books rebuilt, a record-label partnership.
Where a tool had a connector, Claude used it directly. Where it didn't, Claude worked in a browser session the owner had logged into, or ran scripts on the shared Dropbox folder from his own machine.
Fulfilment, service, finance, sponsorship and the website, inside the band's own tools.
738 costume items, packed and posted right.
Seven costume sections, some sold by the band and some by independent section leaders, all packed, labelled, posted or handed out in the four weeks before carnival.
- A scripted packing pipeline. Order data fed scripts that produced packing slips, per-section checklists, collection slips, correction sheets and courier upload files. A late order or a reprint became a rerun, not a rebuild.
- A section-coverage guard. One section's 8 orders had gone unprinted for four months. Every print run now checks all seven sections are covered, and leaves out a partner band's section that shares the same data.
- The “null size” diagnosis. Packing slips showed no size for 176 of 738 items. The order platform had stopped writing one field while its screen read another, so every size was intact. Apology emails to 78 customers were withdrawn before sending, and a bug report went to the supplier.
- 159 orders updated in minutes. The admin screen had no bulk action, so posted orders were marked collected and tagged in one pass inside the owner's logged-in session, with every previous value saved and a rollback ready.
- 112 phantom courier bookings traced. August showed 111 delivery “exceptions”. All were duplicates from an upload run twice. Each one was matched to its printed twin by reference, recipient and date. No parcel was lost.
8,140 emails turned into a playbook.
The whole history of the band's info@ inbox became a knowledge base with a clear rule for what may be sent automatically and what always needs a person.
- The inbox, mined. 3,300 customer emails classified by question type, sentiment and escalation risk, and 340 facts pulled from 952 team replies.
- A knowledge base with a send policy. 149 answers across 30 question types, each marked auto-send, draft for a person, or human only. It ships as a review spreadsheet, a handbook for the service lead, and a file ready for automation.
- A voice brief. Replies sound like a seasoned masquerader who has seen every issue before and treats customers as carnival family.
- Inbox plumbing. Send-as for info@, forwarding to the team, and a filtered folder so customer mail stays out of the main inbox.
- Delivery emails. Branded Brevo notifications with segmented send lists, and a separate version for customers whose parcels were merged into one box.
Costs checked, books rebuilt.
When the work started in July, the books had no reconciled bank data newer than late 2025.
- The courier rate card. Last year's invoice billed London parcels at the remotest zone's rate. One email got the official card, and the next dispatch run cost £752.85 instead of £2,683.98.
- Fewer, fuller parcels. Merging parcels going to shared addresses (135 down to 115) saved a further £131.86.
- Two years of books rebuilt from raw exports. Income and expense for 2024 and 2025, reconstructed and cross-checked against the business's own figures, with gaps named rather than papered over.
- A true-revenue view. Card takings with the platform's customer service fee stripped out, so gross takings are never mistaken for sales.
- Section-leader money mapped. The live data showed independent sections' payments already route through Stripe Connect, with each order tied to the leader paid.
From a cold start to signed partners.
The float had premium space to sell and no warm list. Between mid-May and mid-August the drive reached 37 decision-makers at 33 brands.
- Researched, not scraped. 45 named decision-makers at 41 brands across drinks, tech, fashion, finance and music, each with a verified title, a channel and a confidence rating, checked again before sending.
- LinkedIn first, warm where possible. 31 of the 37 were approached on LinkedIn. For the top leads: a message to a mutual connection first, then a personal InMail once the intro had landed.
- A repositioned offer. The float was pitched as a content and production asset, not a logo slot: access to Europe's biggest street festival and the content that comes out of it, in three tiers, with a deck and one-pager for each warm lead.
- A tracker that kept score. 50 dated follow-up packs drafted between late May and late August, with bounced emails flagged the same day.
- The result. A paid partnership with a major record label, a headline snack partner delivering 1,500 units in kind, and a long-standing drinks partner kept on. Not everything converted; one global drinks group asked to talk again for 2027.
A front door that never closes.
Between seasons, the band's own domain forwarded to a ticketing page that reads “No events found” until costumes go on sale. Every link, search result and Instagram bio was landing on a blank page for months at a time.
- A permanent branded home. The domain moved to the band's own site, live all year, with the ticketing platform linked only when sales open. The web records were switched over without touching a single email record.
- A 2027 interest list. A holding homepage built around the 2026 float video, a one-click sign-up with a single clear consent box, and a welcome email styled to match the band's despatch emails.
- Pages that stay in the team's hands. A design system in the band's colours, a looping video hero and reusable sections, built in Claude Design as templates where every word and photo stays editable.
- Galleries for every season, 2019 to 2026. A gallery hub and year pages with a lightbox and search-friendly structure. 100 photos picked from a 707-shot shoot and 97 from another, then resized, renamed for search, watermarked and alt-texted.
- A logo error caught and fixed. The white logo had a letter filled in that shouldn’t be. It was rebuilt from the master, the bad files removed, and all 100 photos re-watermarked in place without breaking a single page.
- An About page with a point of view. Lines like “Trinidad raised us. Notting Hill gets the band.” and “Strangers in July. Family by Monday.” Every line was pushed past the obvious first draft before it went live.
- Changes that can’t break the live site. Every edit is made on a copy of the live theme, checked in preview and published by the owner, with the previous version kept as a one-click rollback.
How the 149 answers are allowed to go out
Knowledge base v0.1, built from 8,140 emails
- 41Auto-send · Safe, proven answers sent without waiting
- 85Draft for a person · AI writes it, a person approves it
- 23Human only · Refunds, complaints and anything sensitive
Next: a scheduled task that drafts info@ replies for approval. Refunds and complaints are never sent automatically.
What to run 2027 on.
Every function the band used in its 2025 Shopify store was mapped against Playmas in 2026, then each alternative was tested against one hard rule: section leaders keep control of their own Stripe accounts.
Shopify still earned a role, just not at the till: it now runs the shop window — the branded site, the galleries and the 2027 sign-ups — while costume checkout stays where the section payments already work.
The customer-service build made the same kind of call: a keyword chatbot was parked in favour of AI-drafted replies inside the existing mailbox, approved by a person first.
The biggest wins weren't automation.
They were catching what the systems got wrong before a customer or an invoice did.
Trust the physical evidence, not the export.
“What actually shipped” was rebuilt from the printed labels, not from order reports. That's how 112 phantom bookings were proven harmless.
A status field is not a record.
If one edit can silently undo it, you need a durable marker and a check.
Check the premise before you contact the customer.
The 78 apology emails were written, ready, and wrong.
One email can be worth more than a month of tooling.
Asking the courier for its rate card was the single largest saving.
Draft first, automate later.
Every customer reply starts as a draft for a person. Only proven question types graduate to auto-send, and refunds and complaints never do.
People log in, AI does the bulk work.
Credentials stayed with the owner. The AI worked inside sessions he opened, and saved a rollback before every bulk change.
Carnival, events, or any business with one big deadline?
This all started the same way every engagement does: a week inside the business, in numbers. Book a free 20-minute call and we'll find your courier email.