The hidden cost of automating friction
A process may be slow because of redundant approvals, incomplete information or decisions nobody owns. Automating each step without questioning its purpose can deliver faster answers while preserving the same bottleneck. Start with the outcome the customer expects and follow the work end to end.
A study of AI assistance for customer-support agents observed productivity improvements in one specific setting; its results are not a rate applicable to every business. The operational lesson is to measure actual work, quality and conditions of use rather than extrapolate a percentage.
Map the whole journey
Map the input, output, owner, handoffs, waiting time and exceptions. Separate active work from queue time. Then ask what can be eliminated, what requires judgment and what could be supported by rules or AI.
In a hypothetical quotation process, drafting might take minutes while waiting for commercial data and discount approval dominates the elapsed time. Accelerating only the draft may produce little improvement the customer can see.
Design controls before connecting systems
Define data permissions, action limits, traceability and when a person must intervene. For a generated response, specify permitted sources, tests with difficult cases and a correction mechanism. For an automated action, define what can execute, what needs approval and how mistakes can be reversed.
NIST recommends considering generative AI risk across its lifecycle. A small test should include errors, oversight and quality as well as speed and cost.
Measure the flow that matters
Compare total turnaround, rework, cost per case and satisfaction or quality before and after. Distinguish theoretical capacity from a reduction in spending that can actually be realized. If volume grows, check that quality holds.
Jevons can help analyze, design and coordinate automation and implementation with the appropriate team; technical responsibilities and boundaries are defined in the proposal.
Frequently asked questions
Which process should be evaluated first?
Look for a frequent journey with a visible problem, a willing owner and enough data to compare before and after. Avoid starting solely with what is easiest to automate.
When should a person intervene?
That depends on potential harm, reversibility and available reliability. Define review and escalation for sensitive decisions, exceptions and outputs that fail agreed criteria.
Sources and further reading
Let’s discuss your next decision.
Tell us what is at stake, what remains uncertain and where a clearer strategic direction would help.
Start a conversation