ROI and measurement
No AI transformation without a scoreboard.
Most automation business cases are one multiplication: hours saved times hourly rate. That number is always large and almost always wrong. Here is the arithmetic we actually use, including the two terms that usually get left out.
- Baseline measured, not estimated
- KPI agreed before we build
Try it
What does the process cost you today?
Four numbers gets you the order of magnitude. Then we validate it against the real workflow and replace the estimates with measurements.
Four numbers. Everything is an estimate until we measure the real workflow — this is the arithmetic, not a promise.
Everyone who touches it, including part of someone’s week.
Across the whole team, not per person.
Time a real sample, including the awkward ones. Estimates run low.
Salary plus taxes, benefits and overhead — roughly 1.3× base pay.
We plan against 60% on a first workflow. Anything above 80% assumes very stable inputs, so treat it as the optimistic end.
What this process costs today
$7,404/month
195 hours of work a month, or $88,852 a year.
Estimated capacity value if 60% is removed
$53,311/year
About 117 hours a month — roughly 0.7 full-time equivalents of capacity handed back to the team.
This is capacity, not cash. It becomes money when it absorbs growth you would otherwise hire for, clears a backlog that is costing you revenue, or moves experienced people onto work only they can do. We will not call it guaranteed savings, because it isn’t.
We validate this against the actual workflow and send back the real numbers.
The arithmetic
Every term, and the mistake usually made with it
If a vendor’s business case does not contain all eight of these, ask which ones they left out and why.
| Term | How we calculate it | The usual mistake |
|---|---|---|
| Current workflow cost | Volume × minutes per item × loaded hourly cost, plus rework and the cost of errors where we can observe them. | Using base salary instead of loaded cost, and timing only the straightforward items. |
| Residual manual cost | The share that will still need a person after automation, at a longer handling time, because the leftovers are the awkward ones. | Assuming this is zero. No automation of unstructured input reaches 100%, and pretending otherwise turns a success into a failure. |
| Build cost | The fixed implementation price for the workflow, quoted before we start. | Treating discovery as free. Mapping the process properly is most of why the build works. |
| Ongoing operating cost | Infrastructure and model usage, maintenance, and the time somebody spends working the exception queue. | Forgetting maintenance. Vendors change APIs and models change behaviour whether or not you budgeted for it. |
| Capacity recovered | Hours removed from the process, expressed as capacity — and named as capacity, not as cash. | Calling it savings. It only becomes money through one of three specific mechanisms. |
| Error reduction | Fewer items requiring rework, and fewer downstream consequences: wrong shipments, credit notes, disputes. | Ignoring it, which usually understates the case more than the labour maths overstates it. |
| Revenue impact | Only where it is measurable: faster quotes, faster invoicing, shorter customer wait times. | Claiming a revenue number nobody can trace back to the workflow. |
| Payback period | Build cost ÷ (current − residual − run), in months. | Comparing build cost against gross hours saved, which is how automation projects get approved and then disappoint. |
For a worked example and the complete formula, read how to calculate whether an AI automation is worth building.
Being straight about it
Recovered hours are not saved money.
Removing 400 hours of manual work a month does not reduce payroll by 400 hours unless you actually reduce headcount, and most of our clients have no intention of doing that. What you get is capacity — which turns into money in exactly three ways.
01
It absorbs growth you would otherwise hire for
The most common and the most defensible: the next 20% of volume arrives without the next two hires. This is the one we usually build the case on.
02
It clears a backlog that costs you revenue
Quotes going out late, invoices going out slowly, customers waiting. Here the recovered capacity is directly traceable to money.
03
It moves experienced people onto work only they can do
Real, and the one most people feel most strongly. Also the hardest to put a number on, so we do not lead with it.
We say estimated capacity value, not guaranteed savings. If a vendor tells you the second thing, ask which of those three mechanisms turns their number into cash.
After launch
The number should keep moving.
Most of the gain in an automation is not on launch day. It is in the weeks afterwards, when the exception rate gets worked down and the thresholds get tuned against real data.
Illustrative example
Manual processing
32 minper request
Before DevnTech
Week 4
11 minper request
System launched
Week 8
6 minper request
30 days of optimization
What the management dashboard shows at day 60
0%
reduction in processing time
0 hrs
of manual work removed per month
$0
estimated monthly capacity value
Questions
About measurement
Does recovered employee time equal money saved?
Not automatically, and we will not pretend otherwise. Removing 400 hours of manual work a month gives you 400 hours of capacity. It becomes money when that capacity absorbs growth you would otherwise have hired for, or when it stops a backlog costing you revenue. We call it estimated capacity value, not guaranteed savings.
How do you calculate the baseline?
Volume times minutes per item times the loaded hourly cost of the people doing it, plus the cost of rework and errors where we can observe it. We measure a real sample rather than accepting an estimate, because estimates of one’s own process are almost always wrong in the same direction.
What ongoing cost should we plan for?
Three things: infrastructure and model usage, which is usually small and metered; maintenance, because vendors change APIs and models change behaviour; and improvement work, if you want the system to keep getting better. We quote these separately so you can see what is fixed and what is variable.
What payback period is realistic?
For a high-volume, rule-stable workflow, the implementation typically pays back within a few months of the capacity it frees. For low-volume or judgement-heavy work it can be much longer, or never — which is exactly why we measure before building instead of after.
Want us to validate this against the actual workflow?
Send us the process. We measure a real sample, replace the estimates with numbers, and tell you honestly whether it is worth building.
- For businesses with $5M to $30M in revenue
- No long-term commitment
- Start with one workflow
- Keep ownership of everything we build