Behind the Build: How We’re Evolving the AI Colleague

From rigid scripts to proactive conversations. The story behind our latest CloudFO release.

Before CloudFO was a company, it was a side project.

But from the start, we knew we weren’t building a chatbot. We were building something that felt like a colleague, a conversational interface that could guide, respond, explain, and think with you.

Back then, generative AI wasn’t widely accessible. So we went the long way around.

We built decision trees. Mapped out rigid conversation flows. Wrote scripts to try to cover every question a user might ask.

It was complex, brittle, and a nightmare to maintain. The experience felt more like a form with a friendly face than a real conversation.

Then came generative AI, a new way forward and a new set of challenges

Large language models made conversations feel more natural. But they brought their own problems:

• ❌ Responses were unpredictable
• ❌ Financial calculations couldn’t be trusted
• ❌ Ad hoc reasoning slowed performance
• ❌ Taming outputs required prompt engineering, fine-tuning, and RAG

We didn’t trust most Gen-AI models to handle financial logic, especially with hallucinations so common. So we blended generative output with deterministic machine learning, the kind our team had built in finance and banking for decades, to ensure answers stayed grounded and more accurate.

We added code generation for dynamic calculations, but it created latency. As the space evolved, we layered in custom tools and a lightweight knowledge retrieval system (RAG) to give models better context and keep answers relevant.

That got us to our early versions,  with charts, summaries and a searchable knowledge base we kept updated manually.

But it still felt like a chatbot. Not a colleague.

What it feels like to talk to CloudFO now

Our latest release goes far beyond static Q&A.

When users interact with CloudFO today, they get:

• 🧠 Proactive suggestions: It suggests key things you should know about.
• 📊 More dynamic charts and graphs: Presented in ways that suit how you think
• 💬 A more conversational tone: More natural, less robotic
• 🕘 Persistent chat history: So you can pick up right where you left off
• 🧭 Contextual guidance: It doesn’t just show dashboards; it explains them
• 🌐 Awareness: It understands your company data, knows your role, references prior interactions, and can search the internet when needed

It’s the difference between looking at numbers and actually understanding what they mean in plain language, at speed.

Under the hood: what’s changed

Our tech stack has matured.

We now use knowledge graphs to help CloudFO understand relationships and patterns between business drivers, not just return values from a table.

That unlocks more exploratory, strategic questions:

• “Why are repeat purchases up?”
• “What’s driving churn?”
• “What are the most important things I should be paying attention to right now?”

CloudFO can even search the web when its own knowledge and memory don’t have the answer.

We also introduced a model-context protocol (MCP) server to coordinate tasks across a growing set of tools and AI agents:

🔍 Data Agents

Connect to finance, commerce, and banking systems to generate live diagnostics from cash flow to operational health.

📈 Finance Agents

Produce real-time metrics and plain-English weekly summaries. Think: Your finance meeting is already written.

🧭 Strategy Agents

Simulate scenarios (such as pricing changes, hiring, or market entry) and generate structured outputs, reports, and next steps.

It’s not just a generative layer. It’s a hybrid brain.

There’s a lot of noise around “agentic AI” right now and while the hype is real, so is the opportunity.

We’re finding the most value by combining generative agentic AI with the traditional machine learning we know works. Also incorporated elements of recommendation systems to make CloudFO more proactive, contextual and helpful.

Think of the conversation with CloudFO like working with a great teammate:

Sometimes you ask the question. Sometimes they bring it up before you even notice it.

Why we've built it this way

We learnt a lot watching early users.

If they didn’t know what to ask, they didn’t ask. And if they didn’t ask, they didn’t get value.

That’s when we stopped treating chat as just a support layer and started designing it to drive the discussion.

What’s next

A conversation with CloudFO, as described in this post, is like the ad hoc chats you’d have with a helpful teammate. Someone who can look over your shoulder, explain what your dashboard is telling you and flag things you might have missed.

But sometimes, you need to go deeper.

In upcoming posts, we’ll show how CloudFO:

• Runs your weekly business review meeting -  explaining what’s changed, what’s working, what’s not, and why. Then summarising everything into a clear, plain-English report delivered straight to your inbox
• Leads strategic deep dives - helping you work through key decisions like pricing changes, hiring plans, or new market opportunities, with full context and clarity

We will discuss how CloudFO supports structured meetings and planning sessions from forecasting to goal tracking to modelling decision scenarios, led by an AI colleague who understands your business and helps you move faster, with more confidence

Follow along for updates. We’re not just building dashboards and chatbots. We’re building better decisions through conversation.

 Backed by intelligence. Designed for action.