# Discovery: Connecting AI to Your Tools Is Easy. Letting It Act Safely Is the Real Work.

> We started by connecting a Thai accounting tool to Claude. Today our connector server has more than 30 tool groups. The biggest lesson: access is the easy part. Approvals, dry runs and permissions are what make AI safe to use on real business data.

Source: https://pulsarinteractive.ai/blog/mcp-connectors-discovery
Published: 2026-10-15 by Pulsar Interactive, Pulsar Interactive
Updated: 2026-10-15

Every business we talk to wants the same thing: *"Can I just ask the AI, and have it check our accounts, create the purchase order, or find that client?"*

The technology to do this now exists. It's called MCP (Model Context Protocol), an open standard that lets AI assistants like Claude and ChatGPT use other software directly. We have spent this year building MCP connectors for the tools businesses actually run on, and what we learned surprised us.

> **[ MEDIA SLOT 1 · VIDEO or IMAGE ]** Asking Claude a question about real business data through a connector (made-up company data). Replace this block with the media.

## What we found

Our first connector was a proof of concept for an industrial supplier in Thailand. They kept their accounts in a regional accounting tool that no AI assistant could talk to. We connected it, and the team could ask Claude about expenses and documents in plain language.

That was the exciting part. Then came the questions that matter:

- *"Can it create a document, or only read?"*
- *"Who approves before it posts something to our accounts?"*
- *"What happens if it gets the amount wrong?"*

Nobody asked how the connection worked. Everybody asked how to stay in control.

## What we built

Our connector server has grown from that single proof of concept to **more than 30 tool groups**: accounting, payments and business accounts, expenses, purchase orders and their documents, contacts and companies, memberships and vouchers, articles, reconciliation, activity logs and more.

The parts we are proudest of are not the connections themselves:

- **Switched on per business.** Each client only gets the tool groups they need. Nothing else is visible to the AI.
- **Role-based access.** Reading is one permission; creating, changing and approving are others. Anything sensitive needs a logged-in user with the right role.
- **Dry runs before writes.** The AI can prepare an expense or a purchase order and show exactly what it would submit, before anything is posted.
- **Approval gates.** For money and documents, a person approves. The AI prepares; people decide.
- **A control panel.** Clients on a monthly plan get a private dashboard to see documents, approvals and activity in one place.

> **[ MEDIA SLOT 2 · IMAGE ]** A dry run or approval step: the AI's prepared document waiting for a person to approve. Replace this block, or delete it.

## What we learned

**1. Access is becoming a commodity.** Software vendors are starting to ship their own official AI connectors. That is good news: plain access will soon be free. The value moves to what sits on top: the workflow, the approvals, and how the tools work together.

**2. Start read-only.** The fastest way to build trust is to let the AI answer questions first. Teams that start by asking questions adopt it; teams that start by letting it write things get nervous.

**3. Every write needs a dry run and an owner.** If the AI can post to your accounts, someone must see what it will post and approve it. That one rule removes most of the fear.

**4. Agree what "done" means before building.** Our clearest lesson from early projects: every milestone needs a written example of the input, the expected output and how it will be accepted. Without that, an AI project can work and still disappoint.

## Who this is for

- **Owner-run businesses** in South East Asia and Latin America with no in-house engineers.
- **Teams on regional or niche tools** (accounting, CRM, booking) that rarely get first-class AI support.
- **Anyone who wants AI to act on business data**, safely, with a person approving what matters.

## We can set this up for you and your team

We connect the software you already use to the AI assistant you already use, and keep it running:

- **Start with one workflow.** We review how your team works today and pick the workflow where AI saves the most time.
- **Built with your guardrails.** Read-only first, then dry runs and approvals for anything that changes data.
- **Kept running for you.** Monthly plans include new workflows, catalog connectors and your private control panel.

If your team keeps copying data between your tools and an AI chat window, **[see how we work](https://pulsarinteractive.ai/how-it-works)** or **[talk to us about your tools](https://pulsarinteractive.ai/mcp-connector-development)**.