When businesses calculate the cost of a manual process, they usually start with the obvious number: the time it takes, multiplied by the hourly rate of the person doing it. That number is real. It's also incomplete by at least half.
The full cost of a manual process has four components that rarely appear on the same spreadsheet — and the gap between what organisations think they're spending and what they're actually spending is usually where the strongest case for agentic AI is hiding.
The four components of true process cost
- Direct labour: The fully-loaded cost of the people doing the work — salary, super, leave, overhead, management time. Not the salary number: the fully-loaded number, which in Australia typically runs 1.3–1.5× base.
- Error and rework: Every manual process has an error rate. Every error has a cost — the time to find it, fix it, and communicate about it, plus any downstream consequences (a customer complaint, a delayed payment, a compliance flag). This cost is real but rarely measured.
- Consistency penalty: A manual process run by different people at different times produces different outcomes. The variance might be small per instance — but at volume, inconsistency is a cost. It shows up in customer experience, in audit exposure, and in the management time spent investigating why similar situations were handled differently.
- Opportunity cost: The most expensive component and the hardest to quantify. Every hour a capable person spends on a repeatable manual task is an hour they're not spending on work that genuinely requires human judgement. What would your team do with that capacity back?
"The labour line is what you can see. The opportunity cost is what you're actually paying."
How to build a number the board will act on
The goal isn't precision — it's a number that's credible and directionally correct. The approach I use with clients:
- Pick one process. Map it step by step, timing each stage and identifying who touches it.
- Calculate the fully-loaded labour cost per instance.
- Estimate the error rate and the average cost of a rework cycle.
- Multiply by annual volume.
- Add a conservative estimate for management oversight and inconsistency-related exceptions.
In most cases, the number that emerges is 2–4× the figure the business had in mind when they said "it's not that expensive to do manually." That's not a dramatic finding — it's just the full picture.
What the comparison needs to show
The business case for agentic AI isn't just "what does the manual version cost?" It's "what does the agentic version cost, and what does the difference buy?" That means modelling the agentic option with the same rigour — including the governance overhead, the implementation cost, and the ongoing operating cost of the CAgO function that makes the whole thing defensible.
The businesses that get this right don't present a dramatic ROI slide. They present a credible process map with real numbers on both sides — and let the gap speak for itself.