The cheaper artificial-intelligence system is not necessarily the one with the lowest computing bill. In a fictional software trial by Merrowstack, a software co-operative, a compact language model processes most routine invoices without human intervention but sends enough additional documents for checking to erase its operating-cost advantage. The invented evaluation illustrates a purchasing problem for businesses considering AI: a system that is inexpensive to run can become expensive to supervise, particularly when staff must examine exceptions individually before the output can enter a financial workflow.

The compact model handles 82% of the narrowly defined invoice-classification workload without escalation, against 89% for a larger system. That leaves 180 referrals per 1,000 documents for the smaller model and 110 for its counterpart. At an assumed $2 per review, the resulting checking bills are $360 and $220 respectively. Add illustrative computing charges of $40 per 1,000 documents for the compact system and $110 for the larger one, and the total becomes $400 against $330. On those assumptions, the larger model costs 17.5% less overall, despite a computing bill 2.75 times as high.

The scope of the demonstration matters. Neither the 82% figure nor the 89% figure measures general intelligence or establishes reliability outside the test. All documents use a fixed template, the task is confined to classification, and every uncertain classification is routed to a human reviewer. There is no basis for assuming that the same performance would hold across unfamiliar suppliers, altered layouts or other languages. Nor does avoiding escalation establish that an automated classification is correct. The trial describes how work is divided between software and people, not an independently verified accuracy rate.

Read more: The next computing bottleneck may sit around the processor

To make the purchasing implications concrete, an illustrative planning case scales the workload to 10,000 documents a month. At unchanged referral rates, the compact model would generate 1,800 reviews, against 1,100 for the larger system. Monthly computing expenditure would be $400 and $1,100, respectively, while checking would cost $3,600 and $2,200. The combined monthly bills would therefore be $4,000 and $3,300. Annualised at a constant workload and unchanged prices, the larger system would save $8,400. Those are scenario calculations rather than reported spending by Merrowstack, and they exclude integration, procurement and ongoing quality-assurance costs.

The composition of those bills shows why a model-price comparison can mislead. Human review accounts for 90% of the compact system’s combined computing-and-checking cost, compared with roughly 67% for the larger model. The smaller system saves $70 in computing for every 1,000 documents but requires 70 additional reviews. Its computing advantage therefore disappears when each review costs $1. Below that threshold it is cheaper on these two cost categories; above it, the larger system is cheaper. The relevant negotiating question is not merely how far a supplier will cut its price, but how much checking the buyer must retain.

Review time can make that threshold more tangible. At an illustrative fully loaded labour cost of $30 an hour, a four-minute check costs $2, matching the central assumption. The additional 700 monthly referrals in the 10,000-document scenario would then consume about 47 staff hours. If a redesigned interface reduced checking time to one minute, the cost per review would fall to 50 cents. Combined monthly expenditure would drop to $1,300 for the compact model and $1,650 for the larger one, making the compact option about 21% cheaper. Workflow design, rather than model capability alone, would reverse the purchasing outcome.

Read more: A model compact puts public contracts at the centre of democratic renewal

That leaves room for competing interpretations. The case for the larger system rests on reducing exception queues and preserving staff capacity: paying more for computation can be economical when review remains slow. The case for the compact model rests on improving the process around it, especially where checks can be completed quickly. Both arguments have limits. Fewer referrals do not necessarily mean fewer errors, while faster review is not automatically adequate review. A buyer would need to establish that any saving in checking time does not simply move mistakes further downstream, where correction may be more disruptive.

Several costs remain outside the calculation. Connecting either model to an invoice system, maintaining classification rules and auditing automatically accepted documents could change the comparison. So could different mistake rates among documents that never reach a reviewer. A sound procurement exercise would price those activities separately rather than treat unreviewed output as cost-free. It would also distinguish cash savings from released capacity: avoiding 47 hours of monthly checking does not necessarily reduce payroll, although it may allow existing staff to absorb more work without additional hiring.

The next useful step would be a broader, controlled evaluation, rather than an immediate decision based on either automation rate. Merrowstack’s fictional purchasing case would benefit from timed reviews, checks on accepted classifications and tests using changed templates. Buyers could then compare total costs at their own volumes and staffing rates. Until those measurements are available, the demonstration supports a narrower conclusion: benchmark prestige and low computing prices are incomplete guides to value. For routine business AI, the decisive expense may be the work required to establish that the answer can be trusted.