We start with a conversation about the one job worth automating, and we draw a tight boundary around it. A narrow job is one we can build in days, that you can actually trust, and that costs cents to run. We build it to sit on your own machine, use your own AI accounts, and connect to the data you already have.

Whatever it does, it starts by showing you its work instead of doing it. You watch it run on your real data, in the open, for as long as you need to judge it. You will know quickly whether it understands the job, because it is working on your business and not a demo.

For as long as you want, a person approves what it does before anything happens. Some jobs you will hand over quickly. Others you will want to keep a hand on for good, and that is a perfectly good place to stop. You are never forced past the point you are comfortable with.

When you trust it, you let it run on its own, and it carries the job without you. If the work changes, that is what the monthly management is for: we tune it and adjust it as your business moves. The agent stays on your machine whether you keep us on or not.

We built this to look at each file once, a photo, the first page of a document, a frame from a video, and record what it is: a short description, any text it contains, what kind of document it is, and how sensitive it is. It writes that alongside the file and builds a search record, so the file becomes findable by what is actually in it. It reads your files where they already live and copies nothing out.

It runs across your files and you spot-check what it wrote. Search for something you know is in there and see if it surfaces. Look at what it flagged as sensitive and see whether it is right. Because it is describing things you can see with your own eyes, checking it is quick.

What it writes is a catalogue for you: a way to find any file, and a list of the sensitive ones. Nothing here goes to a customer. Approving it means deciding to trust the catalogue enough to act on it, whether that is finding a photo in seconds or deleting the private documents you should not still hold.

Left on, it catalogues every new file as it arrives, so the search stays complete and the sensitive-document list stays current without anyone maintaining it. Sweeping an entire history is a one-time job. After that it just keeps up.

We built this to read the questions customers actually ask across our own businesses, and write a clear post answering each one. It writes from what your team already told customers, so the advice is yours, and then a second, different model fact-checks the draft against current rules before anyone sees it. Two models, one writing and one checking, because a wrong answer in public is worse than no answer.

Every post it writes waits for you to read first. You see the question it picked, the answer it wrote, and you decide whether it is good enough to carry your name. Since it is answering questions you get all the time, you are usually just confirming it said what you would have said.

You publish the ones you like and send back the ones that need work. Many businesses keep this review step permanently, because a blog post is public and it is reasonable to want eyes on it. The agent still does all the writing and checking. You only decide.

When you are ready, you let it publish on a schedule you set, a post a day or a post a week, on its own. It keeps answering the real questions and linking each post to the service it relates to, so your site steadily gains the pages that bring people in. Nobody had to sit and write them.

We built this to read the notes, messages, and bookings we already had, and pull out the durable facts about each customer: what they want, the language they prefer, how they like to be contacted, which services they use, where they are. It writes those as clean tags on the customer's profile. It never contacts anyone. It only organises what you already know.

It runs across your customers and you look at the profiles it builds. Because it is only tidying facts you already had, it is easy to check. Open a customer you know well and see whether it got them right. It usually does, and where it does not you can see why.

The tags it writes are there for your team and for the other agents to use. Nothing about this reaches a customer, so approving it really means deciding you trust the profiles enough to rely on them, in a reply, in a campaign, or as the memory another agent works from.

Left on, it keeps every profile current on its own. A new booking or message updates the customer's tags the same day, so the picture is never stale. This is the most useful agent to run first, because the tidy customer memory it builds is what makes the others sharper.

We built this against the real jobs in our own businesses. It looks at what work actually cost, the real numbers rather than made-up ranges, and writes a clear page answering how much a given service costs. It only ever publishes a figure it can trace back to a real job, so the prices are real rather than guesses.

At first every page it writes waits for you to read before it goes live. You check that the numbers match what you would quote and that it sounds like your business. Because it is working from your own jobs, the prices are usually right, and reading a draft takes a minute.

You approve the pages you are happy with and they publish. You send back the ones that need a change. A few businesses like to keep this review step for good, so a person always sees a price before the public does.

Once you trust it, you let it publish and keep the pages current on its own. When your prices move or a new service appears, it updates the page without being asked. Your pricing pages stay accurate for the first time, and nobody spent an afternoon on them.

We built this reading every message across our own businesses, in every language our customers write in. Its job is narrow: decide whether a conversation is a real opportunity waiting on a reply, and flag it if it is. It never writes to the customer. It just marks the thread in the system your team already uses.

While you are getting to know it, it flags threads and you check its judgement against your own. You will see the ones it caught that you would have missed, and the occasional one it raised that did not need raising. Within a week or two you know how good it is, because you are watching it work on your real conversations.

In practice the flag is the point, so approving here means your team works from the agent's shortlist instead of the full inbox. They open the flagged threads first, reply, and the flag clears. Nothing is sent on anyone's behalf. The agent decides what deserves attention, and a person still does the talking.

When you trust its eye, you let it run continuously, around the clock, in every language. New conversations get read the moment they land, and the ones that matter are waiting at the top of the list when your team logs in. Nobody has to keep the inbox open. The important messages come to you.

We built the first one for a cleaning company we run in Bangkok. The job is deliberately narrow: watch who is overdue for their next visit, and when someone has clearly drifted, write them a message worth reading. It uses the customer history the team already keeps, on the server they already use. Something this specific takes us a few days to build, because we are not trying to make it run the whole business.

At first it does not contact anyone. It writes its messages into a list only your team can see. Each one is addressed to a real customer, in the language that customer normally uses, and suggests the service that makes sense for them next. You read a couple of weeks of these before a single one goes out, so you know exactly what it would say in your name.

When someone likes a message, they approve it and it sends. When they want to fix a line, they fix it. Nothing reaches a customer until a person says so. A lot of owners are happy to leave it here, with the agent finding the people and a staff member giving each message a quick look before it goes.

Once you have watched it for a while and you trust it, you turn on automatic sending. Now it finds the customer, writes the message, and sends it on its own, in your voice, while your team is busy with something else. They never go back to working the list by hand. They just handle the customers who reply.