What you can automate today
What works best are specific, repetitive tasks where a person can check the result:
- Reading documents: invoices, delivery notes, payslips or forms that the AI reads, pulls the data from and sorts. Your team just reviews.
- Answering from your documents: an assistant that answers your team or your customers with the company's information and says where it got it.
- Putting together and sending reports: gathering data from several tools, summarising it and sending it on its own.
- Replying to repetitive emails: drafting the reply, which a person approves before it's sent.
- Taking steps inside your tools: what we call an agent, an AI that doesn't just answer but reads, decides and acts in your email, your CRM or your ERP, within the limits you set.
Two examples of our own:
- An assistant that knows a manufacturer's catalogue. We're piloting, with a client, an assistant inside their app that knows their whole product catalogue. Users ask it which product suits them, which materials it's used with, what other products they need and in what quantity, and the assistant answers with the catalogue's information. It's a pilot: it isn't live yet.
- An app that files documents. Lyncat Tria, our Mac app, reads the documents you give it, names them and files each one in its folder following rules written in your own words. This one you can try.
What not to automate (yet)
- Decisions you can't review: if nobody looks at the result, a mistake can go unnoticed for weeks.
- Tasks you do once a year: automating them costs more than doing them.
- Processes your own team never does the same way twice: first agree on how they're done.
Start with a single task
The most common mistake is wanting to “bring AI into the company” all at once. It works better the other way round:
- Pick the task that takes the most hours and that can be measured: how many invoices, how many emails, how much time.
- Build a prototype with your real data, not demo examples. That way you know whether it's accurate enough before investing more.
- Connect it to your tools (email, CRM, ERP, Drive) only once the prototype works.
- Keep it supervised: a log of every action, a spending cap and a person who approves what matters.
It's the process we follow in AI and automation.
What it costs
There are two different costs:
- Building the automation: the work of setting it up, connecting it and testing it. It depends on the task and your tools, and it's quoted after the prototype, once you know what works.
- Running it: what the AI provider charges for each task, which is usually very little. Our guide Automate tasks with AI agents, safely has real examples: between $0.15 and $0.35 per run at API prices. Every automation has a spending cap, so a mistake doesn't turn into a bill.
Your data
It's the question we're asked most, and the one to ask anyone who offers you AI:
- Do they train on my data? We don't train any model on our clients' data. When we use models on AWS Bedrock, they run in the European Union.
- What does the AI see? Only what it needs for the task, with the minimum permissions: if it has to read invoices, it doesn't need your email.
- Is there a log? Every action is recorded, so you know what it did and when.
- Who decides? Important decisions are always approved by a person.
What about my team?
The idea is to take the mechanical work off their plate so they can focus on what adds value: looking after customers, reviewing what the AI isn't sure about, making decisions. Your team still has the final say.
If you'd rather build it yourself
If someone in your company writes code, Automate tasks with AI agents, safely explains the five rules we follow to leave an agent working on its own and the open-source engine that enforces them.