What happened when the whole office got Claude

TWO months ago, Aspire rolled out Anthropic’s Claude to everyone in the organisation.

The company has more than 500 employees. About 100 are engineers, and most of them were already using artificial intelligence in some form in their work.

The question for us was really about the other 400: operations, finance, compliance, customer success, marketing, product. The people whose tools don’t usually get the first AI upgrade.

We – the AI squad – set a quarterly goal of 250 weekly active users outside of engineering. It felt ambitious at the time. We hit it in two weeks.

This took two things, mostly: a small amount of education (a couple of workshops run by people inside the company, not a vendor pitch), and a clear signal from leadership that this was a tool everyone was meant to use, not a privileged few.

That was it. No mandate, no scorecard, no gamification. People showed up because they were curious, and they kept showing up because the tool started doing real work for them.

The goal at this stage was never to have everyone using AI for everything. It was to give every team a fair chance to find the most boring thing on their desk and see if they could automate it. The only caveat was staying true to our stringent data-handling policy.

How people are using it

In the early days, a lot of usage was what you’d expect. Drafting e-mails, summarising threads, getting unstuck on a doc. Useful, but mostly surface-level stuff.

What I find more interesting is what showed up a few weeks in, when people started reaching past Claude as a chat window and started using it to do work from end to end.

Two stories stand out.

A product manager was working on straight-through processing (STP) for new customer onboarding. In fintech, STP means automating the background and compliance checks so that a new customer can be onboarded quickly, while ensuring that anything the system flags as uncertain is reviewed by a human.

It is high-stakes work. Get it wrong and you onboard the wrong person. Get it overly cautious and you make every legitimate customer wait. The hard part is figuring out which cohorts of customers can safely be automated first.

That is a data analysis problem: Which specific customer segments could responsibly be considered for lighter-touch review, using data already held within a user acceptance testing (UAT) environment?

The product manager, who is not a data engineer, got the analysis to a working answer in about a week.

Using Claude, they explored different modelling approaches, and arrived at a cohort definition. The first cohort now lives in shadow mode or UAT, which means we are running it alongside our existing process – but not yet using its decisions for real onboarding.

Honest caveat: UAT is not production. What lies ahead is testing at scale, backtesting for accuracy and hardening data sanity – which is, realistically, the ceiling on how fast we can grow this.

The point of this story is not that AI replaced a data engineer; it is that the analysis itself, which I assumed would be the bottleneck, was not.

In another telling instance, someone on our customer success team had been using a third-party tool to do sentiment analysis on inbound queries. This involved flagging which messages were likely to escalate, so the team could get to them first.

In one afternoon, this team member rebuilt this analytical core inside our own stack, using Claude (along with its governance and security layer), with help from the AI squad.

The final outcome was more tuned to our customers. It costs us less, and can be changed whenever the team’s needs shift.

I don’t think the takeaway is that AI will lead you to cancel all your software-as-a-service contracts. Sometimes the right answer is still to buy.

But what someone with no engineering background can put together in an afternoon is genuinely remarkable.

What we’re still figuring out

If I had to compress the last couple of months into one sentence, it would be that the most interesting AI work inside a company is not built by the AI team. It is built by the teams whose work it changes.

Our job, as the AI squad, is mostly to remove friction as well as to enable security, governance and connectivity. Beyond that, we’re constantly educating people on what’s possible, so nobody wastes a quarter on a workflow that was never going to hold up.

There remain a few questions with no clean answers yet, including how to keep the quality of internal builds high as more people build them. There is a real difference between a workflow that one person uses and one that a department depends on, and the gap between them is mostly testing.

We are also working through how to decide, at scale, when an automation should be deterministic (a defined workflow, every time) versus agentic (the model makes the call).

For a regulated business, this choice matters a lot, and it is not always obvious in advance.

Fortunately for me, while we operate in a highly regulated industry with the compliance constraints that come with it, we also have a culture that’s willing to say, “Let’s try it”.

Every experiment still goes through the same rigorous data and risk reviews before it reaches a real customer, but even within those guard rails, the possibilities have been fascinating.

The writer is head of AI at Aspire, a business-to-business fintech startup headquartered in Singapore.

The commentary is based on the writer’s observations and argument. AI tools were used for drafting and editing. The writer remains fully accountable for the commentary’s accuracy, originality and final form.