Twelve months of building with AI: what I’ve actually learned

For the last year, I’ve been in the trenches building an AI-powered finance operations platform. Some of it has been exhilarating. Some of it has been hard yakka. All of it has been instructive.

This is my honest take on where AI is genuinely transformative, where it still has a long way to go, and what that means for finance professionals trying to figure out where to lean in.

1. Where AI is genuinely brilliant today

There are three areas where the productivity uplift has been undeniable.

Forming ideas. AI is an extraordinary thinking partner. Pressure-testing hypotheses, surfacing angles you hadn’t considered, moving from a vague notion to a structured strategy in a fraction of the time it used to take. That alone has changed how I work.

Generating proof-of-concepts. What once took weeks of scoping and prototyping can now be stood up in days. POCs that would have required real investment are now cheap to spin up and equally cheap to throw away. That shift changes the economics of innovation entirely.

Sparking insight. Feed AI a dataset, a transcript, a process map, and ask it to look for patterns. The signal-to-noise ratio is genuinely useful. Not always right, but almost always a starting point worth having.

2. Where AI still has work to do

This is where the realism kicks in.

Established workflows are hard. Getting AI to perform a known task seamlessly, every single time, at the level of accuracy a finance team requires, is significantly harder than the demos suggest. You get there. But the path involves iteration, testing, regression, and a fair amount of frustration along the way. It is not turn-key.

“A POC that works 7 times out of 10 is fascinating. A production system that works 7 times out of 10 is unusable.”

The gap between those two states is where most of the real engineering effort lives. For finance use cases in particular, the bar is high. We are dealing with money, compliance, audit trails, and trust. There is no margin for the kind of cheerful unreliability that is tolerable in other domains.

3. Why “simple” workflows aren’t simple

The single biggest underestimation I see, in myself and in others, is how complex a “simple” finance process actually is once you try to automate it.

Take accounts payable. On the face of it, AP looks straightforward. An invoice arrives, you check it, you code it, you pay it. How hard could it be?

Then you start building an AI solution and the variables surface. A skilled AP person is silently working through a checklist that nobody ever wrote down:

• ✓ Is the supplier registered for GST, and does that match what is on the invoice?
• ✓ Is this a valid tax invoice?
• ✓ Do the line items actually sum to the total?
• ✓ Should this amount be recorded ex GST or inc GST?
• ✓ Is this supplier known to us, or is this the first time we have seen them?
• ✓ Do the services billed match what was actually provided?
• ✓ Have the services been provided?
• ✓ Are there supporting attachments, and do they line up with the invoice?
• ✓ Does the GL coding match how we have treated this supplier in the past?
• ✓ Is the amount within the expected range, or is something off?

Every one of those questions is a variable the AI needs to learn, validate, and handle. And that is just the happy path. Now layer in duplicate invoices, supplier name variations, credits, partial deliveries, and situations where two of the checks contradict each other.

What looked like a straight line from inbox to ledger is actually a decision tree with dozens of branches. The AP person has been making those judgement calls for years, often without realising they are doing it.

This is why building production-grade AI for finance takes the time it does. Not because the technology is weak, but because the work itself is more layered than it appears. The skill of an experienced AP person is hiding in plain sight. Replicating it means surfacing every one of those variables and teaching the system to handle each one with the same instinct.

4. The reality of building robust AI solutions

There is still a meaningful need for development skill to take an idea through to a robust production system. The marketing around AI might suggest otherwise. The reality is that bridging the last 20%, the validation, exception handling, audit trail, error recovery, is where dedicated engineering effort earns its keep.

The good news is that this is changing quickly. Development cycles are shrinking. Tooling is improving. The cost of building a robust AI system today is materially lower than it was a year ago, and it will be lower again in another year.

But today, in this moment, building something genuinely production-grade still requires:

• Real development capability
• Disciplined testing and iteration
• Patience through regression cycles
• A clear-eyed view of where AI ends and human oversight begins

The teams that will win are the ones that respect this reality rather than wish it away.

5. The question every finance professional should ask

Here is the takeaway that matters most, and it has nothing to do with technology.

Time and bandwidth are finite resources. As finance professionals, the most important question we can ask ourselves is this: where is our time best spent?

Not on every shiny tool. Not on chasing every demo. And not on grinding through the same repetitive admin work that AI is increasingly capable of handling.

Our time is best spent on the work that actually requires us. The judgement calls. The strategic decisions. The relationships with the business. The oversight and critical thinking that no AI can replicate.

That is the real opportunity. AI does not replace the finance function. It frees the finance function to do the work it has always wanted to do but never had the bandwidth for.

Where I’ve landed after twelve months

1. AI is a brilliant partner for ideas, prototypes, and insights. Use it freely for these.

2. AI is not yet turn-key for production-grade workflows in finance. Building robust solutions still takes real effort, real iteration, and real engineering discipline.

3. The opportunity for finance professionals is not to fear AI or to over-invest in it. It is to use it deliberately, in the places where it earns its keep, so we can spend our time on the work only we can do.

That, more than anything, is what I have taken away from the journey so far. The technology will keep moving. The cost of building will keep falling. But the question for finance leaders stays the same: are you spending your time on the work that only you can do?

If the answer is no, that is the problem worth solving.