Want AI That Actually Pays Off? Start With the Boring Stuff.

A few months ago an AI agent deleted two weeks of my work.

I’d been running an agentic setup, the kind everyone’s selling right now, the AI that works on its own while you get on with your day. It corrupted a database of more than 30,000 processed documents, then deleted it. Gone. Thousands of dollars and two weeks, wiped out by the thing that was supposed to be saving me time.

I know what some of you will say. My fault, for not controlling the agent well enough. Fair enough. But let’s be honest about what that admission actually means. If someone like me can’t keep it on the rails at small scale, what hope does a business have running this across a large, complex environment with real money and real risk on the line?

I’ve spent a lot of this year building with AI. Properly building, with tools like Claude Code and agent bots, not posting about it. And that experience taught me something the hype is desperate for you not to hear.

The shiny version doesn’t hold up.

The 24-hour build. The demo that writes a whole app while you watch. The agent that runs your business while you sleep. It all looks incredible. Then you try to put it into production.

A weekend of vibe coding gives you something that looks finished and isn’t. It’s not built to be reliable, secure, or still standing in six months. The gap between a demo and a system you can run a business on is enormous. Most of the noise online skips straight past it.

So I stopped chasing the shiny thing and asked a duller question. Where does AI actually save a business money right now?

It isn’t agentic AI. Not yet. Agents are coming and they’re genuinely interesting, but they’re not where the savings sit today, and they’re nowhere near as safe to hand the keys to as people make out.

The savings sit somewhere far less sexy.

The four-hour purchase order

I had a conversation recently with a manager who’d just started at a client business. They’d been through hours of onboarding. And it still hadn’t answered the question that actually mattered.

These weren’t gaps in their ability. They had the subject matter expertise and the experience. What they didn’t have was the local knowledge. How this employer does things. The specific processes, approvals and ways of working that are unique to their new business. The transactional reality of the job, not the craft of it.

Their example stuck with me. They spent four hours chasing down whether they were meant to approve a purchase request from one of their own team. Not because they didn’t understand approvals. Because they didn’t yet know how this business handled them. The answer didn’t come from the induction pack. It didn’t even come from their manager, who didn’t know either. It came from a colleague who happened to point them in the right direction.

That’s not an onboarding problem. That’s a business that knows the answer but can’t get it to the person who needs it.

Where the hours really go

Your people already know how your business works. The problem is the knowledge is scattered. It’s buried in shared drives, old emails, policies nobody can find, and the heads of your most experienced staff.

So think about where the hours really go.

Hunting for the right form, the current policy, the template that was definitely somewhere.

The phone call to a senior manager asking the same question that’s been answered fifty times before.

The work redone because someone couldn’t find the existing version and started again.

Five people answering the same question five different ways, because there’s no single source of truth they can reach.

Sound familiar?

Add it up across an organisation. It isn’t minutes. It’s hours per person per week, every week, forever.

That’s the lever. Not a robot doing the work. Taking what your business already knows and making it findable, consistent, and instant.

What RAG does, and what most people get wrong

This is what RAG does well. You take your real material, your policies, procedures, contracts, hard-won knowledge, and you put a system in front of it that anyone can ask in plain language. They get a grounded answer, drawn from your documents. Not made up. Not a generic guess off the internet. And it doesn’t delete your database while you’re at lunch.

But here’s the part that matters, and the part most people get wrong.

RAG done badly is just an LLM with extra steps. You hand it your documents, it sounds confident, and it still makes things up. In safety, in compliance, in approvals, a confident wrong answer is worse than no answer. It manufactures false trust.

So we built ours differently. We’ve spent two years on it. The accuracy doesn’t come from trusting the AI. It comes from engineering. In our system, code does around 80% of the heavy lifting. The retrieval, the structure, the logic, the parts that have to be right every single time, are handled by software that behaves the same way on every run. The LLM only comes in at the end, to turn the right answer into clear, readable language.

We keep the part that can hallucinate well away from the part that has to be correct.

That’s the difference between a demo and a system you can actually put in front of a supervisor about to cut a trench. One sounds clever. The other is right, consistently, and you can see where the answer came from.

It doesn’t replace judgement. It accelerates it.

When the missing answer is a real risk

Now picture where this goes, because the missing answer isn’t always a wasted hour. Sometimes it’s a real risk.

A supervisor starts on a new site. They can’t recall the standard for cable trenching in this business, so they assume it’s the same as their last employer. Two problems in one. A possible compliance breach, and standards quietly drifting because nobody corrected the assumption.

A worker sees something unsafe but isn’t sure who to report it to, or how. So the report doesn’t happen. The signal is lost.

An employee is having a problem with a colleague and can’t work out if it’s something they should genuinely report, or who to raise it with. So they sit on it, and it grows.

In every one of those, the business already has the answer. The standard exists. The reporting process exists. The policy exists. The person just can’t reach it at the moment they need it.

The future state worth building towards

Here’s the future state I’m actually building towards.

Onboarding that never really ends, and never overwhelms. Instead of cramming a week of induction nobody remembers, you give each person a system built for their specific role. It guides them through how things are done here, at the exact moment they need to do them.

The new manager goes to approve that purchase request and gets walked through how this business handles it, right there in the moment. The supervisor about to start trenching gets this site’s standard before the first cut. The employee unsure about a colleague gets pointed to the right policy and the right person to talk to.

They learn by doing. The guidance arrives just in time, tied to the real task, instead of a forgotten slide from week one.

That’s what every business actually wants but struggles to deliver: learning and development, productivity, alignment and consistency, all at once. Not an agent making decisions for people. Their own knowledge, built into the work, reaching every person the instant they need it. People get better at their jobs by doing them, with the right answer always within reach.

This is the engine behind Frank

This is the infrastructure behind Frank. We built it first for HR and IR, the most complex, high-stakes rules a business has to get right, and Frank reads that real material and turns it into clear answers instead of another spreadsheet and another phone call.

And we haven’t stood still. Over the last six months we’ve rebuilt the engine underneath him. Frank 2.0 is dramatically more powerful, more accurate and more flexible than the version that came before. The same two years of hard lessons, now running on far better foundations.

That’s what lets us take it further. We’re applying the same engine well beyond HR and IR. The approach doesn’t care what the subject matter is. Safety, operations, finance, procurement, any business with hard-won knowledge trapped in documents and people’s heads has the same problem, and the same solution.

So here’s my question for you

How many hours is your business losing every week to questions it already knows the answer to? How much knowledge walks out the door every time an experienced person leaves? And what is it costing you when the right answer doesn’t reach the right person in time?

If those questions are worth answering, let’s talk. I’ll show you what we’ve built and how it would work for your business, not a generic demo, your actual problem.

Message me, or find us at lvlup.au. Stop guessing. Start asking the right questions of the knowledge you already have.