An AI system can look highly reliable step by step and still make costly mistakes across a long business workflow. In a Forbes Council essay, Certinia technology chief Raju Malhotra argues that businesses need to measure the reliability of the finished task, not just the model’s accuracy on individual steps.
Watch Desk analysis
What happened
Malhotra illustrates the problem with invoice processing. At 99% accuracy, an AI system handling 10,000 invoices a month would leave around 100 errors for people to catch or explain. In a 20-step workflow, even 99.9% accuracy at each step works out to roughly 98% success across the whole task: about 200 invoices a month with an error, under his example’s assumptions.
His advice is to ask vendors for multistep performance figures and to check what reliable information, such as contract terms and approved rate cards, the system uses to make decisions. Malhotra is a technology executive making an argument from his professional experience, not presenting a universal benchmark.
Why it matters
A small error rate can turn into a steady stream of errors when software repeats a task at scale. In finance and operations, those mistakes may take time to uncover and could affect billing, audits or client relationships. A polished demo can tell you what a system gets right; it may say much less about the work left behind.
Our read
Malhotra’s most useful point is that accuracy needs a unit of measurement. A model’s score on one step is not the same as the chance that a whole invoice, approval or payment workflow finishes correctly. Ask for results on the complete task, and ask what information the system relied on. If those answers are missing, “99.9%” is doing more reassuring than informing.
What to watch
- Whether vendors report error rates for complete workflows, not only individual model steps.
- How results change when contracts, rate cards or other relevant business context are incomplete.
- How often people must review and correct the system’s output.
Discussion spark: Should companies demand end-to-end error rates before deploying AI in finance, or can human review make step-level accuracy good enough?
Sources and evidence
- The Balance Sheet Liability Of 'Mostly Right' AI – Forbes (30 October 2026, 11:45 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.