LLM Adoption
Built in-house by Uppercut Digital
AI in every role.
A person on
every change.
Everyone at Uppercut Digital uses large language models every day, in every department. We built our own tools on top of them, with guardrails the team actually trusts.
The models we use, and the systems our MCP server connects them to
7KQ2Stylised illustration with a demo client and sample figures
Training
No person left behind.
That's what we call it, because AI has a use in every job here, not just in PPC. Most of what we know, we taught each other.
Taught one to one
Most of what we know, we taught each other, one person at a time.
Training and lunch and learns
Team training sessions and lunch and learns, plus dedicated resources we set aside for AI education.
The part people like most
Anyone can show something they've built, and that's how a good idea gets past the person who had it.
Good workflows become shared skills
A workflow that proves itself becomes a shared skill, so it runs the same way for everyone on the team.
What we use it for
I've spent years at other agencies, and AI was always something one or two people tinkered with on the side. Uppercut is the first place I've worked where everyone is expected to use it, and taught how. The show and tells are my favourite part. You see what a teammate has built and you're using it yourself soon after.
How it works
One engine. Connected to everything we run on.
Our own Model Context Protocol (MCP) server connects Claude to the systems we run on, from HubSpot to Google Ads. The same engine answers live data questions, builds campaigns and drafts strategy decks.
What it reads
HubSpot, Google Ads, GA4, Search Console, Business Profile and Meta, for every client.
Our MCP server + Claude
Reads by default, with every call logged. Before any change to a live system, the person says which account and what change.
What it does
Answers questions from the real numbers, drafts builds and decks, and proposes changes for a person to approve.
What we built
Tools we built on top of the models.
Built by our own team, for the agency and for our clients. Two of them have their own pages.
Campaign Build Console
Turns a signed deal into a Google Ads build and strategy deck. Keyword research runs through the Google Ads API, so volumes and costs are real. A new client's research, deck and build now take under 72 hours instead of 2 to 3 weeks.
See how it works DataOko, our own assistant
Checks every client every few hours. It puts Google Ads, Meta, GA4, Search Console, Business Profile and HubSpot on the same 7 day comparison and tells the team which clients need them, and why. Anyone can ask "What needs me?" and get the answer from the live numbers.
Meet OkoOur MCP server
Connects Claude to HubSpot, Google Ads, GA4, Search Console, Business Profile and Meta, and logs every call.
Daily account audit
Budget pacing and search term reviews used to wait until someone had time. Now an audit runs every day and proposes the actions for a person to approve.
AI reviewer with a scorecard
Every Google Ads and Meta build goes past it before a person approves. Fixed checks first, then a judgement call against our standards.
Quality control
A reviewer that earns trust one rule at a time.
A rule isn't trusted until it has agreed with the team 10 times in a row, and every human correction is stored as a learning for that client.
Agreed with the team 10 times in a row. Now trusted.
What we got wrong: an early version marked a dental client's free consults as "not on the site", because it had only read the top of the page. It now reads the parts of a page that carry offers.
Demo client, for illustration
Governance
The rules we wrote down early.
Because so much of this is new, we wrote hard rules down before anyone needed them.
4F7KAnything that changes an account (negatives, budgets, bid targets, tag publishes) hands back a preview and a confirm code. Nothing applies until a person approves that exact preview.
Demo client, for illustration
Challenges
What we got wrong, and how we fixed it.
Wary account managers
Our account managers were wary of a model touching client money, and they were right to be. So we fixed it in the design. The model reads freely, changes nothing without approval, and the scorecard makes each rule earn its place from the team's own verdicts.
Wrong answers that look right
The reviewer once missed an offer because it had only read the top of the page. We changed the data tools as well, so when a source can't be read they say so plainly instead of reporting zero.
Keeping up
The tools change faster than anyone's habits, which is why the rules are written down and the training never stops.
For our clients
The time we save goes back to clients.
Faster launches
Research, strategy deck and campaigns in under 72 hours, so a new client is live sooner.
Answers from real numbers
"Why did our dental client's calls drop last week?" gets answered from the client's own data, in chat or in Oko.
Accounts checked every day
Budget pacing and search terms are reviewed daily, not when someone has time, and every build is checked against our standards first.
Key numbers
Every role. Every day.
Figures from Uppercut Digital's own records, October 2026.
See the tools