In this last week I burned through over 20 million tokens building things for clients and for my own business. That number sounds abstract until you translate it into what it actually bought: hours of work that used to sit on someone's desk every single week, now handled automatically. Here are the three builds that made the biggest difference, and what they can tell you about where AI automation for business actually pays off.
Why Token Count Isn't the Point, But It's a Useful Signal
Twenty million tokens isn't a brag stat, it's a workload indicator. Every one of those tokens was spent on a task that had a business reason behind it: pulling data, writing drafts, checking logic, generating output. When I look back at a week like that, the useful question isn't how much AI I used, it's what got built with it. So let's get into the three things that mattered most.
Cutting Hours Off a Recurring Reporting Task
The first project was a reporting job that used to eat a real chunk of someone's week, every week. It's the kind of task a lot of businesses have buried in their operations: pull numbers from a few different places, format them, write up a summary, send it out. Nobody loves doing it, and it rarely gets done faster no matter how many times you've done it before.
I built an automation that handles the pulling, the formatting, and a first-pass write-up, so the person who used to own that task now spends a few minutes reviewing and editing instead of building it from scratch. That's the pattern worth paying attention to: if a task is repetitive, rule-based, and produces a predictable kind of output, it's a strong candidate for this kind of build. It doesn't have to be glamorous to be worth doing. Reporting, status updates, weekly summaries, these are exactly the tasks that quietly cost the most over a year because they happen so often.
Building a Research and Drafting Tool That Actually Gets Used
The second build was a tool for research and first-draft writing, the kind of work that normally means someone opening a dozen browser tabs, reading through them, and then staring at a blank page trying to turn it into something coherent.
I set it up so it does the gathering and the first pass at writing, and a person takes it from there. The key detail here isn't that it writes something, plenty of tools do that. It's that it gets used. A lot of AI tools get built, get shown off once, and then quietly die because they don't fit into how someone actually works day to day. This one stuck because it slots into an existing process instead of asking someone to change how they work to accommodate it. If you're building or buying automation, that's the test that matters more than any feature list: will this still be open on someone's screen in three weeks?
Automating a Decision-Support Task That Used to Require a Person's Judgment
The third one is the most interesting because it goes a step past pulling data or drafting text. It's a task that required someone to look at a set of inputs and make a judgment call, and I built a system that handles most of that judgment automatically, flagging only the edge cases for a human to look at.
This is the category people are usually most nervous about, and for good reason. Judgment calls feel like they should stay with a person. But the reality in most businesses is that a big share of those judgment calls are actually pretty routine. There's a smaller set of genuinely tricky cases sitting inside a much larger pile of straightforward ones, and most of the time isn't spent on the hard cases, it's spent grinding through the easy ones to get to them. Automating the routine calls and routing the hard ones to a person doesn't remove judgment from the process. It puts a person's judgment where it's actually needed instead of spreading it thin across everything.
What Ties These Three Builds Together
None of these three projects were flashy. No one of them is a story you'd tell at a dinner party. But between them they took a recurring reporting job, a research and writing bottleneck, and a judgment-heavy decision task, and turned all three into something that runs mostly on its own, with a person checking the output instead of producing it from scratch.
That's what a heavy week of building actually looks like in practice. It's not one big splashy AI project, it's several smaller, unglamorous ones stacked on top of each other, each one clawing back a few hours here and a few hours there. Multiply that across a year and it adds up to a lot more than 20 million tokens' worth of time saved.
If you're looking at your own business and wondering where something like this could apply, start by listing the tasks that happen on a schedule, produce a predictable kind of output, or involve a judgment call that's usually pretty easy to make. Those are your candidates.
Talk to Us About What This Could Look Like for You
If any of this sounds like a task sitting somewhere in your business right now, reporting nobody wants to do, research that eats a day, decisions that follow a pattern, we'd like to hear about it. Reach out to Level Up AI and let's talk through what a build like this could look like for your team.