We build automation, and then we run it.

ColdSilver is a small studio in Delft. We build systems that do real work with real money in them, and we run them afterwards. When one breaks, you reach the person who built it. Moonlight, Sterling and Sophie came out of this work.

Write to us What we build

A colour-checker strip and a grey card lying on the studio floor, the tools for getting colour right before a shot

CALIBRATE BEFORE YOU SHOOT

What we build, and keep running

Customer-service agents

Where the studio started. An agent reads every inbound message, looks the answer up in your own systems (orders, bookings, your sent mail) and replies in your voice. The judgment calls go to a person. Sophie is the self-serve version of this.

E-commerce operations

The unglamorous middle of a store, where money leaks: refunds issued twice, orders cancelled on one platform and not the other, stock that disagrees with itself.

Finance back office

Invoicing, payment reminders and reconciliation, wired into the accounting system you already use. We have built this on Exact Online.

Marketing intelligence

Google and Meta ad accounts audited and run against what actually converts, bookings and orders rather than clicks. Every change is logged and reversible.

Integrations and internal tools

The systems that should talk to each other and do not, and the small internal tool nobody sells because it only makes sense for you.

Operating AI in production

Running what you or someone else already built: monitoring, guards around money, a person to escalate to, and an alarm when it goes quiet.

How we work on a live system

Four rules, learned running our own products on other people’s money.

Save the state before you change itBefore we touch a live system we save how it was, so we can put it back.

Test on the real thing firstA new agent runs in shadow on real mail before it sends a word. A change to a live account is tried on real cases, not the one we had in mind.

Money waits for a personA refund, a charge or a cancellation runs only after someone has said yes to it, in writing or in a switch they set themselves.

Nothing fails quietlyOur systems mail us when they stop. A failure reaches us first, not your customer’s inbox three days later.

You talk to the people doing the work. The person who built it is the one who answers when it needs to change.

One deployment we can talk about

Client work stays anonymous on principle. The specifics come on a call.

The whole support inbox of a UK and EU fashion brand

About 50 cases a day, 1,500 a month: roughly 1.25 full-time jobs. Thin margins, and return shipping ate a real part of each order. Orders ran through a storefront and a separate payment platform that did not always agree with each other.

We built the inbox: order lookups across both systems, replies in the brand’s voice, return handling, and refunds behind guards: the order has to belong to the sender, a refund is issued once, and the ceiling is the owner’s. Senders the owner flagged went to a person with the thread attached.

~80%: of return requests settled without a shipping label
~99%: uptime; failures reached our monitoring, not the complaints box
~1,500: cases a month, answered from the brand’s own inbox

Tell us what should not need a human

We say whether automation fits, and show you a real anonymised deployment. A 15-minute call; no slide deck unless you ask for one.

Write to us