AI & automation consulting, built for real teams.

If the people delivering the service are fighting the tools all day, chasing the same errors, and typing the same reply for the hundredth time, they are not going to be in a position to look after your customers properly. That is where I start. Every project below took repeated, grinding work off a team first. Better customer experience came out of that, not the other way round.

Their contract charged them for every registered user, not every active user. So every dormant account was costing money every single month.

Once a month the team got handed a list. Everyone who had registered over thirty days ago, never bought anything, and gone quiet since. Somewhere between three and eight thousand accounts. Five people then worked through that list in a spreadsheet by hand. Open the account, check it, delete it, move to the next one. Full days, every month.

I built a tool in Python that used Playwright to drive the browser and do exactly what a person would do. The rules were strict. No ICCID, no purchase history, no activity in the last fourteen days.

Deletion cannot be undone, so it re-checked every account at the moment of deletion rather than trusting the list it had been handed. That caught the case that actually mattered. Some users bought an eSIM in the gap between the list being generated and their account coming up for processing. Those got skipped. Every run wrote a CSV showing what was deleted, what was not, and why.

It took around 68 hours a week of manual work off the team, and turned a recurring monthly cost into something managed rather than accepted.

Two people ran the support desk on HubSpot, and the tool gave them nothing to work with. No ticket assignment. No routing. No agreed response times. Every reply was typed from scratch, including “hi, how can I help you today”. Web chat first response was sitting over ten minutes. Email was over an hour. That was not the agents being slow. Nothing in the system was helping them.

So the first thing I did was move the desk off HubSpot onto Zendesk and rebuild how it ran. Tickets got assigned and routed, so every request had an owner. We agreed response targets, so there was a standard to work to. And I built macros for the repetitive openers and the common troubleshooting paths, things like walking someone through a Windows or connection problem.

I deliberately left the body of email replies alone. Those still needed an agent to read the problem and write the answer. A macro would only have made that worse.

Alongside that I reworked the help centre and added deflection into web chat, so people got pointed at an article that answered them faster than a queue could.

Web chat first response went from over ten minutes to under fifteen seconds. Email came down from over an hour to roughly ten.

All of their global shipping ran through Zapier. Label creation, warehouse routing, the whole path from someone placing an order to a device arriving at their door. The problem was it had no tolerance for failure. If an API was rate limited or went down for thirty seconds, the run failed. It posted an error into a Slack channel, added a row to a Google Sheet, and then somebody in the logistics team had to go in, work out what had gone wrong, and create the label by hand before the warehouse could ship anything. Failures that would have sorted themselves out were creating real manual work.

I rebuilt it in n8n. Failed calls now retry, five seconds apart, three attempts, thirty second timeout. Most of those failures resolve on their own and nobody hears about them.

The rebuild also made proper fallbacks possible. If someone in the UK asks for next day delivery and FedEx is not available, it books DPD automatically. Before, that was another Slack message and another manual booking.

All of that across two product lines, one for the US market and one for the EU, with warehouses in America, Mexico, Australia and Poland, and different customs rules in every one.

The logistics team stopped being the retry mechanism.