Working out the ROI of automation before you buy anything
Vendors quote hours saved. Hours are only money if the time gets used for something else. Here is the arithmetic that survives contact with a finance director.
Founder·28 August 2026·6 min read
Automation proposals almost always lead with hours saved, because it is the easiest number to make large. Multiply a task by a frequency by a headcount and you can justify anything.
The version that holds up later is narrower and more convincing.
Step one: measure the current process, do not estimate it
Ask the people doing the work to record the real numbers for two weeks. How many documents, enquiries or reconciliations, how long each takes, how often something goes wrong and what fixing it costs.
Almost every business is surprised by this data. Volumes are usually higher than management thinks and the time per item is usually lower, because people have found shortcuts nobody documented.
Step two: count only capacity you will actually use
- Time that lets you handle more volume without hiring is a real saving.
- Time that lets you stop paying overtime is a real saving.
- Time that lets a specialist stop doing clerical work is real, if they use it for something that produces value.
- Ten minutes returned to each of eight people is not a saving. It is a nicer day.
Step three: price the errors
This is usually the biggest and least counted number. Duplicate payments, mis-keyed quantities, an invoice missed until the supplier calls, a lead never contacted. Each has a direct cost and a relationship cost.
You do not need precision. An order of magnitude is enough to change the decision, and it is often larger than the labour saving.
Step four: subtract the honest costs
- Build or licence cost, including the integration work into your existing systems.
- Running cost: hosting, usage of any AI model, monitoring.
- Human time on the exception queue, which never goes to zero.
- Maintenance when a layout, an API or a rule changes.
- The change management cost of getting people to work differently, which is real even though it never appears in a proposal.
Step five: run it in parallel before you commit
The most useful thing you can do is run the automation alongside the manual process for two weeks and compare outputs item by item. It costs a little time and tells you the true auto-processing rate rather than the demo one.
It also converts a debate about whether it will work into evidence, which is a much better basis for extending scope.
Pick projects with these properties
High volume, low judgement, clear inputs, one system to write to, and an owner who wants it to work. Those succeed. Low volume tasks with heavy judgement and three stakeholders who disagree about the process are where automation budgets go to die.
SupraAuto is built for the first category: document processing, CRM and API integration, WhatsApp and email workflows, with everything logged so you can audit the savings rather than assert them.
If the arithmetic only works when you count minutes returned to everybody, it does not work. If it works on volume handled, errors avoided and hires deferred, it will still work next year.
Frequently asked questions
Is hours saved a real saving?
Only when the hours turn into something else you value: more volume handled, faster response, work that was being skipped, or headcount you did not need to add. Twenty minutes returned to five people is usually absorbed invisibly. Twenty hours returned to one person is capacity you can point at.
What ongoing costs get forgotten?
Model or API usage if the workflow calls a language model on volume, hosting, monitoring, and the human time spent handling exceptions and reviewing the queue. Also budget for maintenance: when a supplier changes their invoice layout or a system changes its API, something has to be adjusted.
What is a reasonable payback period?
For a single well scoped workflow, most of the ones worth doing pay back inside six to twelve months. If the honest arithmetic says three years, the process is probably too small to automate, and the right answer is to simplify or drop it instead.