Blog

what. connect: AI Adoption Without the Privacy Risk

Banning AI doesn’t protect your data. It just pushes your team into personal accounts you can’t see. We built one controlled AI layer that runs our whole agency, with per-person access and human sign-off. Here’s what months of running it on ourselves taught us, warts included.

what. connect illustration hero
Anand Patel
Written by
29.10.2026

Many teams hold back on AI because of one fair concern: where does our sensitive data end up? A client in education raised it with us recently, and we knew the feeling. Locking AI down would have left most of our team behind. Leaving it open would have put personnel files and client material at risk.

Rather than choose between protecting that data and getting everyone to use AI, we built one layer that does both. It’s called what. connect: a single, controlled route between our team’s AI assistants and the tools they work in. We’ve run our agency on it for months, and now we’re opening it to clients.

Why a separate AI chatbot in every app doesn’t work

A chatbot per app starts blind, because it only knows what happens inside its own tool. The one in your CRM can’t see the email that just arrived, and you end up as the human glue between assistants.

what. connect gives one AI assistant access to the tools we already work in, each person with their own login. Ask it to check whether a lead replied, update the deal and tell the team, and it handles the whole chain. For us, the chat itself was never the valuable part. The payoff came when the assistant could see the ticket, the thread and the email together.

The moment your AI can see your inbox, your CRM, and your project board at the same time, it stops being a chatbot and starts being an actual operations layer.
Luke Szkudlarek , Founder & Growth Strategy Lead at what.

It grew out of our internal “AI Brain” initiative. Before that, everyone wired up AI alone: a local Model Context Protocol (MCP) server here, a token in a config file there. MCP is the open standard that lets AI reach business tools, but it needs deliberate security work. Nobody knew who had access to what, and shared resources like client folders couldn’t be used safely at all.

Giving every employee a DIY AI setup isn’t innovation. It’s shadow IT with a friendlier name.

The biggest privacy risk isn’t AI. It’s AI with no approved way to use it

An earlier top-down rollout at what. had failed because it didn’t solve real problems. Meanwhile, we’re convinced that people were pasting confidential data into personal AI accounts. We just couldn’t see it.

That’s the real compliance gap. If the official tool is good, nobody has a reason to go around it. With one shared layer, everyone gets the same strong tool, HR data stays with HR, and each person sees only what their job needs.

How what. connect works: apps, team apps and shared skills

The assistant reaches Slack, Gmail, Google Calendar, Drive, Linear, GitHub, WordPress and Google Search Console, plus our time tracker what. works and Optiverse, a third-party meeting recorder. That’s just the list we use ourselves, not a limit: if a tool has an API, we can connect it, and we add more apps whenever a team needs them. It all comes down to three building blocks:

Building blockWhat it isExamples
AppsIndividual tool connections, permitted centrally, used with your own loginSlack, Gmail, Linear
Team appsShared accounts and folders for the whole teamPipedrive, Front, client folders
Shared skills and workflowsSaved step-by-step recipes that anyone can run by asking, and anyone can build“run worklog”
what. connect: One secure layer between each person’s AI assistant and the tools they work in.

Read also: Why we let AI log our own billable hours

Access control: who sees what, even on shared team accounts

Everyone gets what they need and nothing more, and an admin can prove it. For us, this is the most important part of the whole system.

The core rule: the assistant can only do what you could already do yourself in a tool, using your own account. Each person connects their own accounts once, via OAuth or their own API token depending on the tool, so every call runs with that person’s permissions in the source system. Some apps start switched off. Our DigitalOcean connection only has read tools and stays off until an admin enables it for a specific person, because not everyone needs a server dashboard.

By default, anything that writes, sends or deletes asks for confirmation first: writes ask once per session, deletes ask every time, and this can be tightened per app. Client and billing data is where that rule earns its keep: hours turn into invoice lines and emails leave under our name, so the AI prepares the draft and a human signs it off. Any Drive folder can also be put off-limits to the AI for the whole company, so no model reaches it through our tools.

Some tools only make sense as one shared account: our team Pipedrive, Front, Sentry, shared Drive folders, a client’s Shopify store. An admin connects the account once. The team uses it through their own AI, and nobody ever sees the password or token, which stays on our server. The rules:

  • One Access page. An admin opens an app to everyone, holds it back from one person, or gives it to just a few. Every grant shows who made it and when.
  • Read-only by default. The AI can look things up, but changes are made with someone’s personal account, in their own name.
  • Writing is a deliberate grant. Where needed, an admin or appointed app manager gives Read or Write per person. Anyone without access can’t even see the account.
  • Drive “collections”. Instead of opening a whole drive, an admin publishes a curated set of folders with its own search. Nobody sees it until they’re added.

With AI and company data, the question is never only “can it do this”. It’s “who can see what, and can we prove it”. One place that answers that for every person and every tool is what made us comfortable letting AI near shared data at all. In Switzerland that’s no small point: only 34% of SMEs have clear rules on what staff may enter into AI tools, and the revised data protection act (FADP) has applied since September 2023.

Related: Data privacy and security for AI systems and agents

Shared skills and workflows: build once, the whole team uses it

Shared skills and workflows are what got our team using it. A skill is a saved, step-by-step way of doing a task, so we treat the two as one: anyone can run one by name, and since September anyone can also write and publish one. Our current favorites:

  • worklog: logs your hours from one sentence, finds the right offer and ticket, writes a comment a client can read, and holds the entry until you approve it.
  • followup: turns a meeting transcript into a summary and email drafts.
  • client-pulse: a seven-day status for one client: shipped, in progress, blocked.
  • drift-check: compares Linear, Slack and meeting notes and flags dates at risk, stale tickets and open questions.
  • content-relaunch-audit: checks Search Console and GA4 data before a relaunch, so we don’t delete pages that quietly bring in traffic.

Also relevant: AI now handles the meeting notes we used to write

Lessons from months of internal use: adoption and fixes

Three things changed once everyone had the same layer:

  • Adoption. Anyone can ask any question in plain language, and what. connect picks the right tool and combines answers from several sources. Engineers no longer have the head start.
  • Compliance. The approved route became the easiest route, and compliance went up along with adoption.
  • Shared knowledge. When someone builds a skill, the whole team gets it, so nobody builds the same thing twice.

Adoption is still wildly uneven, and that’s fine. A few people use it all day, others only for one workflow like worklog, some barely at all. What changed things wasn’t adding more apps. It was workflows: people who’d never write a long prompt will happily type “run worklog”.

The rollout was more about habits than a programme. Everyone got training and access, and every other Tuesday the team swaps what worked in a short session we call AI Tuesdays. The rest came from small wins: a colleague mentions a slow chore, such as setting up a new project, an engineer turns it into a shared skill, and from then on anyone can run it with a single sentence.

Real use also finds gaps fast. Most improvements started with a colleague telling us in Slack that something didn’t work, so failed tool calls are now collected every night and the common ones get fixed. If we started again, we’d ship workflows earlier and connectors later.

Is what. connect right for your business?

Probably, if your team stitches tools together by hand or you’re nervous about AI near client data. what. connect is the AI-agent layer on top of our Tools Integration work, now available to Swiss SMEs. We implement it across the tools your team already runs: mainstream ones like Gmail and Slack within days, niche or in-house ones very likely too.

We didn’t build this to sell it. We built it because we needed it. We add tools and workflows almost every week, so treat it as a living layer, not a finished product. Talk to our AI automation team about connecting AI to your business tools.

What our clients ask us

FAQs

Do I need to understand MCP to use what. connect?

No. The layer handles the protocol behind the scenes, so your team just talks to the assistant in plain language. The hard part isn’t the technology, it’s deciding who may use which tool, and that’s a decision for management, not just IT.

Can I switch a tool off if something feels wrong?

Does the whole team have to use it?

Which AI assistants does it work with?

Does our data end up with US providers?

Where does it run, and what is logged?

Who is responsible for the data?

Related blog posts