I Saved $4M a Year Before AI Was a Thing

There is one number on the front page of brianarfi.com that people ask about more than anything else: $4M+ in yearly costs saved across the companies I have worked at.

And almost everyone guesses the same thing: "That was AI, right?"

No. Those savings happened before generative AI took off. No chatbots, no LLMs, no agents. Just automation, process work, and someone willing to sit down and trace where the money was leaking.

That is exactly what makes the story worth telling now.

Where the $4 million a year comes from

The number is not from one magic project. It is accumulated across several companies, and every piece of it is public on my track record page.

The first part comes from Flip, one of Indonesia's leading money-transfer companies. We cut money-transfer costs by about 32% in 6 months, worth roughly USD 2.12 million per year. That was 450% of the original target.

Nothing about it was glamorous. We re-routed transfer traffic, eliminated redundant API calls, and stripped out wasteful steps. Cleanup work, not clever work.

The second part comes from Tokopedia, years earlier. My team cut OTP costs by around USD 2 million per year while improving the success rate. Again, no exotic technology. Just someone willing to follow every rupiah of that bill to its source.

Those two items alone add up past $4 million a year. Other efficiency work, like deflecting support tickets away from expensive human queues, pushed it further.

So when people say efficiency is a new trend that AI started, I see it the other way around. Efficiency is old work. AI just changed who gets to do it.

Why it used to require engineers and expensive systems

This is the part nobody talks about.

Every one of those savings had a long line of people behind it: engineers writing and maintaining the code, analysts validating the numbers, infrastructure running the systems, and months of roadmap negotiations with other teams.

My role as a product person was clear: find the leak, quantify it, convince people, and drive the execution. But I could not do any of it alone. Even a simple automation idea had to wait in a backlog behind dozens of competing priorities.

Which means leverage at that scale was only available to companies with strong engineering teams. The regular employee staring at a wasteful process every single day? Their best move was filing a ticket and waiting.

That was the world before AI. Ideas were cheap. Execution was expensive.

What changed

I now have living proof of the other side.

I am not an engineer. But over the past few months, working with AI, I built an 8-channel content distribution engine for my own personal brand: one long video gets automatically cut into clips and spread across YouTube, Shorts, Reels, Stories, WhatsApp, LinkedIn, and Threads. It schedules itself and reports to me every day.

Work that used to require a content team now runs daily, and I operate it alone. I review, I approve, it ships.

Alongside that, I built an Applied-AI certification system that runs end to end, from assessment to badge, and I run AI Circle, a community for people who want to learn this way of working.

The difference from my Flip and Tokopedia days is not the principle. The principle is identical. The difference is the tooling: leverage that used to require engineers and expensive systems can now be learned and used by anyone on your team.

What failed, to keep this honest

So this does not read like a highlight reel, here are some of my own failures from the AI era.

My auto-reply system once failed at a 99% rate for almost two full days. A platform login session got invalidated, but the system kept running: it kept generating replies with an LLM, kept paying for every single one, then failed to post them. Thousands of failed replies in a single run, barely a dozen successes. Not one alert fired. I only found out because I happened to check.

I also had a post publish to the wrong brand account entirely, because an account-matching rule compared credentials in the wrong direction. The system worked perfectly. It just did the wrong thing very efficiently.

Those failures taught me a principle I now refuse to compromise on:

Automation without monitoring is not savings, it is cost waiting in line.

And this is not a new lesson either. In the pre-AI days, unmonitored systems were always time bombs. AI just made the bombs cheaper to assemble, so more people can build one by accident.

The selection principle never changed

If you ask me which process to automate first, my answer today is exactly what it was ten years ago.

The most profitable automation is not the smartest one. It is the one doing the most boring work at the highest volume.

When I attacked OTP costs at Tokopedia, I did not start with technology. I started with a question: which transactions have the highest volume and the most repetitive pattern? When I attacked transfer costs at Flip, same thing: find the most-traveled route and fix it there.

Today, when I decide which content work to hand to AI, it is still the same question.

One extra filter people skip: price the cost of a mistake first. If an error is cheap and easy to correct, automate fully. If an error is expensive, keep a human on the final decision and let the system prepare the other 90%.

That principle held in a payment system moving millions of transactions, and it holds in a one-person content pipeline. The only thing that changed drastically is the price of admission.

If you want to start at your company

Here is what I would do if I were starting from zero again:

1. Find the work your team complains about every week. High volume, clear rules, and nobody loves doing it. That is your starting point.
2. Measure in money, not in "cool". Automation you cannot translate into dollars will not survive the next budget cut.
3. Do not chase 100%. The complicated remainder is often better left with humans.
4. Ship monitoring with the system, not after it. The most expensive failures are the silent ones.
5. Do not wait until you have an engineering team. That is the difference from my early days. The tools are already in your hands.

One last thing. I reached that $4 million figure without AI, which proves the principle was right, not that the tools were magic. The tools have since been democratized. What has not been democratized is the willingness to sit down, break a problem into small pieces, and start with the most boring work.

What is the most boring, highest-volume work on your team? Start there.