Most companies have “adopted AI” by giving every employee a chat window to paste things into. An agent is something else: a program that reads mail on its own, opens files on the company drive, checks the calendar and carries out multi-step tasks — but only once someone configures it for a specific person's work and decides what it must not do.
That is why we do not run classroom sessions. We sit at one person's desk, on their computer, with their mailbox and their tasks for the week. After four hours the agent works, the person knows how to use it, and the company has one set of data rules for everyone. We work this way ourselves: Invisio has run dozens of projects on Claude Code since June 2026 and publishes its own AI policy.
Who it is for For people who live in mail, offers, documents and calendars — not only developers
The people who gain most are those whose day is made of mail, documents, offers, spreadsheets and deadlines: office staff, sales, customer service, marketing, accounting, executive assistants. For them the agent genuinely gives time back: it reads the mailbox, drafts replies for approval, tidies files on the drive, assembles offers from your templates and keeps an eye on the calendar — and they do not have to learn a terminal or “prompting”.
We implement for developers too, but honestly: most of them manage on their own, and our value lies where a team wants shared rules, configuration files for its stack and order in access rights, not in learning the tool from scratch.
- office, sales, customer service, marketing, accounting, executive assistants
- owners of small companies who answer everything themselves
- developers: shared rules and stack-specific configuration, not a beginners' course
How it works Four hours at one person's desk
We start with a conversation: what this person's week looks like, what takes most of their time, which of those things the agent can take over entirely and which it should only prepare for approval. Then we install the Claude desktop app, enable Claude Code and remote sessions, sign in to the company account and connect the tools this person really works in. The scope of integrations follows their role, not a ready-made list: for one person it is mail and the calendar, for another a sales system, spreadsheets or an industry application.
The second half of the session is work on real tasks from this week: the agent does, the person corrects and learns what to delegate, how to check the result and when not to trust it. We finish with rules: what the agent must not pass on, which data is confidential, who approves sending. The person gets a one-page cheat sheet, the company a training confirmation.
- about 30 min: a conversation about the person's work and the tasks for the agent
- about 60 min: installation, account, Claude Code, remote sessions, integrations
- about 2 h: work on this person's real tasks, cheat sheet and data rules
What you get The person gets a working agent, the company gets rules and proof of training
Every person leaves the session with an agent configured for their role: connected tools, saved instructions on which tasks to do and how, and a one-page cheat sheet. After two weeks we connect remotely for 45 minutes: what works, what gets in the way, what to add.
At the first visit the company receives a short AI policy (what is allowed, what is not, who is responsible), data rules for the roles being onboarded, a list of required licences and a confirmation of the training with a list of topics — it can be added to your documentation of the obligation to ensure staff AI literacy under Article 4 of the EU AI Act.
- person: an agent for the role, integrations, instructions, cheat sheet, check-in after 2 weeks
- company: AI policy, data rules, licence list, training confirmation
- optional: agent care on a monthly plan, so questions have their own channel
Data and security The agent gets access only to what it is meant to do
An agent with access to mail and files is a real risk if nobody sets the rules. That is why at every workstation we state explicitly: which mailboxes, folders and calendars the agent can access, what it must not send without approval, which data (customers, HR, finance) is out of reach and how results are checked before sending.
We work on Claude's business plans (Team), where company data is not used to train models under Anthropic's terms for business customers. We apply the same rules to ourselves and describe them in Invisio's public AI policy — on request we sign an NDA and a data processing agreement.
- the agent's scope of access written down for each role
- human approval before sending and before changing client files
- a Claude business plan, NDA and data processing agreement on request
Our AI usage policy →
Licences and hardware What you buy yourself and what we do
You buy licences directly from the vendor — we do not resell them and add no margin. The implementation needs a Claude Team plan for the company (accounts for the people being onboarded; in the vast majority of cases a Standard seat is enough), a computer running macOS or Windows, and access to the accounts and applications the agent is meant to handle at that workstation. Before the visit we send a short list of what should be ready, so the four hours go into work, not passwords.
On request we also implement OpenAI Codex in companies working on ChatGPT Business — the rules and the process are the same.
- Claude Team: accounts for the people being onboarded, bought by the company
- a macOS or Windows computer, access to the tools the agent is meant to handle
- a preparation checklist sent before the visit
What licences and the whole rollout cost — worked examples →
First implementations Four workstations in two days at Margot Travel — and agents for managers
We did the first client implementation at Margot Travel, a Krakow incentive travel agency: four workstations in two days, Claude Code in the Claude desktop app on Macs, remote sessions and integrations with mailboxes, the company network drive, Canva and Google calendars. On the first day, three people in six hours turned out to be too fast a pace — hence four hours per workstation today.
We have also prepared agents for managers at our clients: the managing director of the AKmed clinic, the project manager of E-wywiad and the IT director of the Goldruk printing house — each time for different tasks and different tools. Invisio has worked on Claude Code since June 2026: dozens of client projects, our own mail, analytics and documents. We implement a tool we use every day, not one we have read about.
- Margot Travel: 4 workstations in 2 days — mail, network drive, Canva, Google calendars
- agents for a managing director (AKmed), a project manager (E-wywiad) and an IT director (Goldruk)
- the same tool runs Invisio's daily work since June 2026
Margot Travel — see the whole collaboration →