Questions
Should AI touch this process, or should a rule?
Use AI in a process when the task repeats often, its output can be checked, mistakes are cheap to catch or reverse, you have measured accuracy on your own past cases, and the full cost, human review included, beats both the current process and a simpler rule. Where an error is costly or hard to see, keep a person reviewing every result.
Argumont is decision support. It does not provide legal or regulatory advice about automating a process.
Four tests before automating
The first is the cost of an error: what happens when the output is wrong, who notices, and how quickly. A miscoded invoice caught at month end is a nuisance, while a wrong answer in a medical, legal, or lending decision can harm a person. The second is measurement, meaning whether you can score the tool on a few hundred past cases where the right answer is known.
The third is volume, because a task done twenty times a month rarely repays the setup and oversight. The fourth is the simpler alternative. Many processes improve more from a better form, a fixed rule, or a template than from a model, and those options are cheaper to run and easier to explain.
A worked example
A finance team codes 3,000 supplier invoices a month. A clerk takes four minutes on each, 200 hours a month at $35 an hour, so the current process costs $7,000. A tool that suggests codes was right on 92 percent of 500 past invoices, and with a person checking every suggestion for a minute, plus $1,200 a month for the tool, the cost falls to $2,950.
A plainer option deserves the same scrutiny. Seventy percent of invoices come from recurring suppliers whose codes rarely change, and simple matching rules code them correctly almost every time. Coding the other 900 by hand costs $2,100 a month, and with an allowance for keeping the rules current the total is about $2,400, cheaper than the AI plan and with errors that are easier to trace.
The strongest option combines them: rules for recurring suppliers, and AI suggestions with review for the remaining 900, at about $1,225 a month all told. The tempting shortcut of skipping review on invoices under $500, which are 60 percent of the volume, would leave about 144 miscoded invoices a month for the close to find.
Keeping a person in the loop
Review is how you learn the tool's real error rate in your own conditions, so treat it as measurement rather than distrust. Begin by checking every result, record the errors, and reduce the share you check only when the measured rate falls below your people's own, which you should measure too.
In hiring, lending, health, and legal work, rules may require a person to review or explain an outcome. Find out what applies to your process before deciding how much of it a tool may finish on its own.
How the room weighs automation
The current process, a rule-based change, AI with review, and a combination each get their own Champion. The Challenger presses on the accuracy figure, asking whose cases, how many, and how recent, and Inspector holds a vendor's accuracy claim as unverified until it is reproduced on your own records.
Red Teamer runs the month a new invoice format arrives and accuracy drops. Aslan weighs who bears the cost of an error, from suppliers paid late to the staff who correct the books. The Verdict names the error rate that would change it, and the Human Decision stays with the owner of the process.
Questions
- How accurate is accurate enough?
- More accurate than the process it replaces, after review. Measure your current error rate first, because many teams discover they never knew it.
- How many past cases do we need to test?
- A few hundred real cases with known answers usually give a usable first estimate. Include the awkward ones, because those are where tools tend to stumble.
- Can we remove the human review later?
- For low-risk cases, often yes, once months of measured results support it. Keep a regular sample under review so any drift is caught early.
Bring Argumont a Decision worth challenging.
Start with a paid 60 day pilot on your own real Decisions.