AI agents that do real work every day

An AI agent is a program that gets a task, reads what it needs, decides what to do and does it. In other words, besides answering it also acts: it sorts an email, updates a record, alerts someone or drafts a reply.

Unlike a chatbot, it does more than chat, because it works inside your processes. And unlike classic automation, it understands text, nuance and cases that weren't planned for.

It does have its limits, though. It won't handle everything on its own, and you don't install it on Friday to have it running on Monday. Nor will it decide for you how your company should work.

In sales, it reads each new contact, judges whether it fits and assigns it to its owner. So there's no more handing them out at the end of the day.

In customer service, it answers the questions that keep coming back, by phone, by voice on your website or by text. What it doesn't know, it hands to a person with the context already summarised. And bear in mind that if it talks to your customers, it has to be accessible to anyone.

In marketing, it watches ad performance, flags when something goes wrong and adapts content to several languages without losing the tone. In operations, it checks every day that what should work does work. And in management, it answers plain-language questions about your own data, without waiting for someone to pull a report.

I am Head of AI GTM Engineering at Sesame HR, with a mandate over the internal AI projects of the whole company. There I have more than ten tools in production across five markets. Between them, they follow competitors, answer questions about business data, prepare content in several languages, watch ad performance and check the website end to end.

Before that I was Global Lead, Automation & Tech at Dentsu International (Dentsu X and iProspect), in London. That is where I built GeoAdsFinder, a tool that ended up being used across the network. Along the way I have worked in agencies and large companies, in marketing, revenue and product.

Outside Sesame HR I build AIsWise: voice and text AI assistants with accessibility built in to WCAG 2.2, the international web accessibility standard. And back in 2021, when ChatGPT did not exist yet, I co-founded Chat Ergo Bot, a startup working on conversational AI, which I led. So I have spent years seeing what really works once the demo is over.

First, you need a process that already exists. Could you explain to someone new, step by step, how that task gets done today? If not, an agent won't guess it either, and all it will do is automate the mess.

Next, data the agent can read. If the information lives in one person's head or is spread across ten spreadsheets, that is the first job.

Finally, you need a person accountable for the result, who decides what is right and what needs fixing. And also a record of what the agent does, with what it read, what it decided and why, so anyone on your team can check its work.

Many pilots stay there because a demo is built to go well, with clean data, an ideal case and no awkward questions about what happens when it gets something wrong. In a real company, by contrast, the data is incomplete and the odd cases are half the job. And someone has to answer when something goes wrong.

So what is usually missing is the system around it: where the data comes from, who supervises and how you measure whether the agent helps or gets in the way.

I start by understanding your company from the inside: who does what, with which tools and where work gets stuck. Sometimes the answer is an agent and sometimes it is something simpler, like sorting out where information lives. And I pick what works in your case.

When it's time to build, I start with a single task, check that it helps and leave it running before moving on to the next one. Your team then stops copying data from one place to another and can spend that time on customers. And every decision the agent makes is on record.