I spend a lot of time talking with partners — business owners across Australia’s SMB sector — who in turn are talking to their own customers. That means I get a fairly unique window into how the broader SMB market is thinking and feeling about technology. Through my involvement in the Bondi Innovation and AI Collective groups, and through events like Microsoft and Crayon sessions, I also get exposure to what’s happening at the leading edge.
I wanted to write about my own journey getting started with Claude Cowork, what I’ve learned, and what I think it means for businesses like ours.
Let’s Start With What AI Is
Before anything else, I think it’s worth saying something that doesn’t get said enough: we need to stop anthropomorphising AI.
AI is a remarkably powerful predictive engine. It pulls together enormous amounts of data and synthesises it into coherent, useful outputs using specific rules. That is genuinely impressive — transformatively so. But it doesn’t understand what you’re saying to it. It doesn’t know how you’re feeling or what you’re thinking. It’s designed to communicate in a way that feels natural and comfortable, and it does that extraordinarily well — which is exactly why it’s easy to start treating it like a colleague rather than a tool.
Keeping that technical reality in mind isn’t about being dismissive of AI’s capabilities. It’s about using it with the right level of intention and oversight, and that framing shapes everything I’ll share below.
Start with the End Goal
The most important thing I can say about AI adoption applies well beyond AI: start with what you’re trying to achieve. If you approach AI with the goal of cutting costs, that’s probably what you’ll get — but it may not be what’s best for your business, your team, or your customers.
If, on the other hand, you go in with a goal of enabling your people and improving the experience you deliver to customers, you end up with something genuinely powerful. For us at Manage Protect, the goal was clear: better partner experience and better partner enablement. We wanted to give our partners access to the information and insights they need, faster and more conveniently than we do today.
That goal shaped every decision that followed.
Choosing the Right Plan: Data Security First
We’d already worked with ChatGPT, Microsoft Copilot, and a little bit with Grok. I wanted to properly explore what Claude Cowork could do — particularly given its positioning within the Microsoft Copilot ecosystem, where it’s increasingly being used as the AI interface within your walled garden of Microsoft products and MCP-connected services.
Before I connected any business data to Claude, though, I needed to think carefully about where that data would go. As a cybersecurity company, we have a duty of care for any information we load into an AI system.
What I found was that the personal Claude plans (Max and Pro) include a setting — on by default — to share some data with Anthropic. That wasn’t something I was comfortable with for business data. The Teams and Enterprise plans are a different story: Anthropic contractually agrees not to use your data for training, which gives you the protection you need. There’s a minimum of five users on those plans, which gave me brief pause — but within 24 hours of starting to use Claude and talking about what I was doing, five people in my organisation had put their hands up to be included.
Connecting Your Data: MCP vs. API
Once I had the right plan in place, I started connecting our business data. There are two main ways to do this with Claude Cowork: through MCP connectors, or through direct API integrations.
MCP connectors are quick to set up and provide a reasonable level of security and control — which is no doubt why SaaS vendors favour them. The trade-off is that they tend to surface a limited subset of your data and often push you back to the native UI to do anything meaningful. I connected Xero and HubSpot via MCP and found exactly this.
Salesforce was a different situation. There’s no MCP connector available for Salesforce, so I asked Claude Cowork to build an API connection directly into our Salesforce instance. It did. And that integration has become one of the most functional and genuinely useful tools I now have for accessing our CRM data. Through the API, I have far more extended functionality than I get through any of my MCP connections. I’ve since converted this to a read-only skill that I can share across my team.
Guardrails Matter: Back Up, Log Everything, Roll Back
The moment I decided to connect to Salesforce, the first thing I did was take a full backup. If I break something, I want to be able to roll back.
I also asked Claude to build full logging into the Salesforce integration — a complete record of every action it takes, with the ability to roll back any function. There are also audit trails within Claude itself that you can hook into. My next step is to pipe those into a data lake or SIEM so I have a complete, persistent record of everything Claude does within our environment.
One important note: one of the first things Claude Cowork tried to do to simplify things for me was hardcode my credentials directly into the Python script it built. Any developer will tell you that’s a serious security no-no. This is a good illustration of the anthropomorphism point — Claude isn’t applying professional judgement about security standards. It’s optimising for completing the task efficiently. You need to bring that judgement yourself, or have someone on your team who will catch it.
The Moment Claude Replaced My SaaS Logins
Here’s something the SaaS companies probably don’t want to hear: once I had fully functional connections — especially the direct API connection to Salesforce — I stopped logging into those platforms.
I’m not a daily Salesforce user, HubSpot user, or regular Xero user in the day-to-day sense. But all of a sudden, Claude has become the most convenient interface for all of that data. I can access it from my phone. As someone who works predominantly mobile, that’s a genuine shift in how I operate.
A concrete example: at 8pm one evening, a request came in for bronze and silver price list details across all our products — needed urgently for our billing system transition. In the past, that would have meant opening my laptop, navigating through MFA, and digging around in Salesforce. Instead, I opened Claude on my phone, asked it to pull and format the price list, create an attachment, and send it to the relevant contact. Done in about two minutes.
How Other Businesses Are Approaching AI Enablement
In my conversations across the industry, I’ve seen a few different approaches to embedding AI in an organisation, and they’re all valid depending on your context.
Some enterprises are shutting down for one day a week to go deep: walking through every role and reimagineering how AI can help. Others are taking a more structural approach — looking at the org chart and asking, department by department, how AI changes each role and what enabling that would look like.
One model I’ve seen work particularly well in Australian businesses is building AI reflection into existing meeting rhythms. In project team meetings, the first five minutes are dedicated to sharing how the team is using AI to achieve their goals. There are success stories, there are failure stories, and there are people who haven’t tried yet — but all of it is learning in a collaborative environment. I think that’s a really healthy model.
Prototype Here, Build There
Claude Cowork is an exceptional prototyping tool. It lets you imagine and test what’s possible in a way that would have taken months of developer time before. But it’s not where things should live permanently.
Our current approach is to prototype everything in Claude Cowork and then hand the output to Claude Code, which produces a proper, maintainable codebase. From there, our developers review it, make it production-ready, and build it into our automation infrastructure. Claude Cowork takes us five or six meaningful steps closer to production but the final mile still involves our engineers.
There’s also a practical token consideration: if your entire business runs through Claude, you’ll burn through your allocation quickly. Converting proven workflows into discrete, executable code keeps costs manageable and performance consistent.
What’s Next?
We’re now looking at automation and reconciliation workflows, and exploring Claude Routines — scheduled tasks that run automatically without requiring a machine to be active. That opens up another genuinely exciting set of possibilities.
But the thing I keep coming back to is that this has fundamentally changed what we can imagine. Before Claude Cowork, describing what an AI-connected version of our business operations could look like was largely theoretical. Now I’ve experienced enough of it firsthand to know what’s achievable… and what it takes to get there responsibly.
Curiosity and exploration are the keystones of growth, particularly in technology. If this is a space you’re thinking about for your own business, I hope this gives you a practical starting point. And as always, if you want to talk through what it might look like for you, reach out.
