What makes enterprise AI worth the cost?

The cheap answer is only part of the cost

The transition can lose money even when the finished AI workflow looks profitable. I would budget the period in which the old process must remain available, and identify what evidence allows its cost to end.

I tend to ask who will do the work an AI system leaves behind. A wrong field, an ambiguous document or an unresolved exception still needs somebody’s time. A small inference bill can sit beside a large review queue.

Brynjolfsson, Li and Raymond’s 2023 NBER study examined 5,179 customer-support agents and found uneven productivity gains across workers. It is a useful warning against carrying one uplift across an organisation. The process and the people using the tool matter.

Synthetic model, not observed ROI; your inputs stay in this browser.

Electric motors left the old factory in place

Early factory electrification offers a useful picture of this problem. In their 1999 economic-history study, Paul David and Gavin Wright describe factories that added electric motors while retaining the shafts and belts that had transmitted steam or water power. The power source changed, while much of the old production arrangement remained. Later, separate motors for individual machines allowed different layouts, smoother material flow and less dependence on a shared mechanical drive. David and Wright, pp. 6–8

Picture a new motor turning an old shaft above the same rows of machines. Buying the motor was one decision; changing how materials and people moved through the factory was another. I use that distinction when looking at AI costs. A faster extraction step can leave the old checking and escalation process intact. Before counting the saving, I would trace one case through to completion and identify which work actually disappears. The history suggests a question to investigate; it gives no timetable or guaranteed return for an AI rollout.

The same model can make or lose money

Consider 1,000 cases a month, 20 manual minutes per case and labour at $60 an hour. The manual baseline is $20,000. With 60% adoption, 25% of adopted cases needing 30 minutes of review, $1 AI usage per adopted case and $2,000 fixed monthly cost, the model gives $4,900 net benefit and 125 hours of released capacity.

Hold the volume and costs steady. At 85% adoption, 10% review and 20 review minutes, the benefit rises to $12,450. At 75% adoption, 60% review and 50 review minutes, it becomes a $10,250 loss and adds 125 hours of work. The difference comes from the work around the answer.

Fewer exceptions can still cost more

At $60 an hour, a 25% chance of a 30-minute review costs $7.50 per adopted case. In the starting scenario, that is more than the $1 AI usage cost. The useful question is how much attention each exception consumes.

Five percent needing two hours costs $6 per adopted case; twenty percent needing fifteen minutes costs $3. Automating the easy work can leave a smaller queue of harder cases. I would measure the spread of resolution times as well as the average.

More adoption scales the economics already present. If the value per adopted case is positive, adoption helps spread fixed costs. If it is negative, adoption grows the loss. A review bottleneck can also lengthen turnaround even when the average cost looks attractive.

Budget the period when both processes must stay available

The starting model shows $4,900 of monthly benefit. Now add a separate, fictional transition cost: $5,500 a month to keep a fallback team available while the new process is proved. Assume this cost is additional to the model’s fixed costs and remaining manual handling. The transition then loses $600 a month, although the eventual operating case remains positive. Counting the fallback twice would be equally misleading; each cost needs one place in the budget.

I would attach an exit condition to that extra cost. Which decisions can move without the fallback, which failures must still reach a person, and what happens if the new process has to stop? A calendar date alone cannot answer those questions. Equally, keeping the old arrangement indefinitely can erase the saving the project was intended to produce.

This suggests two business cases: the cost of learning whether to switch, and the value after the switch. Compare a narrower trial, a staged transfer and retaining the manual process. The useful decision is when enough work and responsibility can actually move to release capacity; a cheaper model call cannot make that transfer by itself.

Decide how saved hours become business value

The 125 hours in the starting case are an operating choice. They could support more throughput, shorter turnaround or less overtime. Turning them into cash savings needs a staffing or demand decision.

For a stable, repetitive process with little review, a simple cost-per-case comparison may be enough. Before carrying that result elsewhere, I would check eligible-work adoption, intervention frequency and the time needed to resolve an exception. Those three observations tell me more than the price of a successful demonstration.

How the calculation works

Manual hours = monthly cases × manual minutes ÷ 60. Review hours = adopted cases × review rate × review minutes ÷ 60. Net benefit subtracts remaining manual labour, review labour, AI usage and fixed cost from the manual baseline.

Break-even adoption is fixed cost divided by monthly cases and the value per adopted case. With zero fixed cost and negative per-case value, only zero adoption breaks even; with zero fixed cost and zero per-case value, every adoption level breaks even. “Not reachable” means the model finds no break-even level between zero and 100%.

The calculation assumes constant volume, average labour rates and linear costs. Adopted cases without exceptions need no further handling in this model. Routine human handling across all adopted cases needs a separate cost estimate; review minutes apply only to cases counted as exceptions. Implementation, transition costs, undetected quality failures and changing demand need a separate estimate.