Five jobs I used to do by hand in Karbon. Now I just ask.

Turning a client email into a full list of Karbon jobs usually means an afternoon of data entry. The other day I turned one email into nine work items and typed none of them. I asked Claude to read the email, build the work and set up the client, and it did.

That probably sounds like it needs a developer. It doesn't. I run Growthwise, a firm in Newcastle, and I am not a technical person. I spend a lot of time on this, but the week before the summit I vibe coded against the Karbon API and worked it out inside an hour.

At the summit I walked through five jobs we all still do by hand in Karbon, and how AI can take them on. You can watch the session here. The barrier to starting is far lower than it looks.

MCP just means driving your software in plain English

Most of what follows runs on something called MCP, the Model Context Protocol. Anthropic released it as an open standard, so it works whether you are in Claude, Copilot or anything else.

The idea is plain. An API is how software talks to other software, and the documentation is miserable to read. I gave Karbon's docs for one simple task to Claude a year ago and still came away confused. MCP puts a plain-English layer over that. You skip the jargon and just say give me my work items.

One word in the name is worth holding onto. Protocol. It is the handshake that sets what an agent is allowed to do. Some connectors, like Xero's, ship read-only on purpose. The protocol is what stops a plain-English request from doing something you never intended.

You don't need an MCP to start, just a spreadsheet export

Two of the five jobs need none of this. No server, no agent, no setup.

Karbon lets you export your data. Export your contacts, drop the spreadsheet into Claude and ask who has no mobile number. Across 157 contacts it ran the check and flagged the gaps in seconds. When you are setting a firm up, clean data is half the battle, and this is a fast way to check mobiles, emails and anything else.

The same trick builds a dashboard. I exported a timesheet and asked for a breakdown of time by task type. One tip saves you money here. Look at how the export is actually labelled first. I asked for work type, but the column was task type, and naming it correctly meant Claude spent fewer tokens guessing. A minute later I had an interactive dashboard of hours by user and task, the sort of thing you would normally build by hand.

With a connector, AI reassigns and creates the work for you

Karbon has been teasing its own MCP server. It is not out yet, so we built our own from the OpenAPI spec Karbon publishes on GitHub. You hand that spec to Claude and ask it to generate a server. It does exactly that. I used Python, and I am still not a developer.

With the connector switched on, the plain English does real work. I told Claude a staff member was on leave and to move her jobs to a colleague. It found her four work items and reassigned them in one request. It even hit a limit on the way. One method would not do the job, so it switched to another on its own.

Then the email. I pointed Claude at a client request sitting in Gmail and asked it to build the work. It read the email and saw the client did not exist in Karbon yet. It offered to create the record, then turned that one message into nine work items and assigned them to the right person.

It can reach outside Karbon too. With an SMS connector added, I asked it to text a client good luck for an event. It pulled the number straight from Karbon, sent the message and logged a note on the timeline so the team could see it was sent.

The rough edges are real, and paid does not mean safe

I am not going to pretend this is polished. MCP servers misread requests. On the live demo the text one went hunting for a Karbon event that did not exist, and I had to hand it the client's ID to get it back on track. You nudge it more than you would like.

Security matters more than the demos. Read the terms of every tool. Do not assume paid means safe. On Claude's personal plans, a thumbs up or thumbs down can be retained for up to five years. Team and business plans shrink that sharply and do not train on your data by default. For anything sensitive, ask providers about a zero data retention agreement, where your data is deleted once it has been processed.

Cost is smaller than people fear. Match the model to the task. Reading a short email off a file you already gave it does not need your heaviest model. Across the whole session my usage moved from three per cent to six per cent of my plan. Ask for three hundred work items and it would climb, but nothing here was expensive.

Karbon's own server is days away, so most firms should wait

I need to be straight about who should build this. Almost none of you. Karbon's official MCP server is days away, and the first release will be read-only. The spec exposes sixty-six capabilities, so the full picture is coming, just in stages.

Building your own carries real security risk, and you would be rebuilding something Karbon is about to hand you. Treat what I have shown as a preview of what you will soon click on, rather than a weekend project. The one habit worth starting now is leaning on your workflow tools for what they already automate, and saving AI for the gaps they leave.

Where this goes next: agents draft, humans approve

Every task so far has started with me asking. The next step removes the asking. Picture Vinyl transcribing a client meeting into Karbon, and Claude watching that transcript, seeing the follow-ups and drafting the work without being asked. That is agent to agent.

How far you let it run is your call. You might be happy for an agent to draft the email but want to press send yourself. That is human in the loop. You let it work up to a point, then a person approves the final step. That line is the whole game for a profession built on getting things right, and you get to decide where it sits.

The building was always going to get easy. Knowing where to stop is the real skill now.

Watch the full session here.