AI, automation and development

What we build with language models

We use language models in day-to-day client work, not as a demonstration.

• Agentic workflows. An agent takes a multi-step job and works through it with the tools it needs: ad platform APIs, analytics, spreadsheets, a database. A typical job pulls search terms, classifies them, checks them against the account and drafts negatives. It reports what it did and what it wants a person to approve.
• LLM pipelines. Query classification at scale, search term and n-gram analysis, ad copy variants within brand and regulatory limits, and feed enrichment: better titles, missing attributes, cleaner categories.
• Custom internal tools and dashboards. Built around how one team actually works, and maintained after launch.

Tested before it is trusted

A model that is right most of the time is not good enough to change a live account. So we build evaluation harnesses. A harness is a fixed set of real examples with known correct answers. Every change to a prompt, a model or a workflow runs against it, and the scores are compared. If a new version classifies search terms worse, we find out before it ships, not after the spend.

People stay on the decision

Language models are good at first drafts, classification and spotting patterns across thousands of rows. They do not know what your business can promise or what a regulator will accept. People stay responsible for that. Every automated change is logged and has an off switch.

Automation that does not need AI

Much of the most useful automation is plain code. Google Ads scripts for bids, budgets, pacing and anomaly checks. Reporting pipelines that pull, clean and check data, with no copying of cells. Anything done every week is a candidate, and automating it removes errors as well as hours.

We have done this for a long time. In 2019 we wrote Google Ads scripts for Confused.com’s in-house team: hygiene checks, alerts, keyword review and n-gram analysis.

Development that removes the bottleneck

A lot of PPC and GA4 friction is a developer queue. A button that fires nothing, a data layer that was never built, a booking system in an iframe that the ads cannot see into. We fix those ourselves, or alongside your developers, and we take on a landing page or a full site build when it serves the account. The development we do is the development the measurement needs, done quickly and documented.

Recognition

Our work for Slizza won Best Use of AI in PPC at the 2025 European Search Awards, a category open to agencies of every size. The Slizza work won three more categories that year and took a silver.

Where we usually start

1. An automation audit, with the five biggest time sinks scripted or handed to an agent
2. A reporting pipeline from ad platforms and analytics into one checked dataset
3. A small internal tool or agent, from prototype to something the team uses daily