The AI draft is cheap. The exceptions may not be.

“It saves ten minutes” leaves out most of the calculation

A demonstration turns an enquiry into a draft in seconds. The current process takes ten minutes, so someone multiplies the difference by monthly volume and announces a large saving. The missing question is what happens after the draft appears.

Somebody checks it. Some cases need repair. Somebody maintains the approved answers. The system costs money even when an output is unusable. If those activities are left outside the calculation, an assistant that produces more work can look profitable.

Here is a fictional small-business example in Thai baht. All volumes, time values and allowances are teaching assumptions, not vendor prices, salary guidance or DLVX performance claims. Replace them with your records. The point is to see which assumption changes the decision, not to copy a flattering number into a proposal.

Start with what the same work costs today

Suppose a team handles 200 eligible enquiries a month. Each takes ten minutes of human work today. The baseline is 2,000 minutes, or 33.33 hours. The business chooses THB 300 per hour as an illustrative value for that capacity, giving a baseline economic cost of THB 10,000.

This does not mean the company will receive THB 10,000 in cash if it automates the task. Salaries may remain unchanged. Capacity becomes useful if it can be used for another needed job, reduce overtime or avoid an actual additional expense. Record which benefit is realistically available.

Keep existing software out of the incremental cost if the business would pay for it anyway. Include a new licence only when this workflow creates that expense. Use the same scope for the manual and assisted versions: a draft is not comparable with a fully completed enquiry.

Now count the assisted process

The assisted process releases 680 minutes of capacity, about 11.33 hours. After the software allowance, the illustrative economic benefit is THB 1,900 per month. If setup costs THB 9,500, simple economic recovery takes five months while the assumptions remain true.

That is not cash payback unless the capacity value corresponds to a real avoided expense. Label it accurately. If training, support or taxes create additional costs for your situation, add them rather than hiding them inside contingency.

What if half the enquiries need extra work?

Keep everything else the same and raise the exception rate from 25% to 50%. Extra exception work becomes 1,200 minutes. Total human work becomes 1,920 minutes, or 32 hours. At the same capacity value, that is THB 9,600, plus THB 1,500 software: THB 11,100.

The workflow now costs THB 1,100 more economically than the baseline. The generation is still fast. The business outcome is worse. A salesperson showing only the attractive first draft would miss the assumption that matters most.

With these inputs, the break-even exception rate is approximately 40.8%. The formula is THB 10,000 minus THB 1,500 software, minus THB 3,600 for base review and maintenance, divided by THB 12,000 of exception cost at a 100% rate. That threshold is a planning boundary, not a prediction.

Also test lower monthly volume and longer review time. Fixed maintenance can dominate a small queue, even when each individual draft is good.

Avoid counting the same minute twice

In this example, every item receives three review minutes. The twelve exception minutes are additional repair work after that review. If your exception measurement includes the review already, subtract the overlap before calculating. Otherwise you may unfairly reject a useful system.

Record who did the work and what happened. A difficult case might need clarification from the customer, correction of a source, or a complete manual restart. Those are different reasons and suggest different fixes. Improving intake may be more valuable than buying a more capable model.

Count missing outputs and failed runs. An enquiry that disappears still requires recovery. If the team rescues it outside the tracking tool, bring that time back into the calculation. Measure first-pass acceptance separately from eventual completion so that repeated human intervention remains visible.

The money calculation cannot excuse a serious error

A wrong price, unauthorized booking or disclosure of another customer's information should not be averaged away as a few extra minutes. Define hard quality conditions before testing. A workflow that violates one may need to stop even when its overall time calculation is positive.

Keep the pilot narrow enough to inspect. Use the kinds of enquiries the team actually handles, including awkward and incomplete ones. Avoid presenting a handful of easy examples as a reliable production estimate. Run a supervised period, record uncertainty, and update the decision as evidence accumulates.

OpenAI's evaluation guidance supports task-specific testing. The THB budget, break-even rule and business decision in this guide are our illustrative calculations, not vendor benchmarks.

Bring this worksheet to the next proposal review

Eligible items per month:
Current human minutes per completed item:
Review minutes for every assisted item:
Exception rate, including failed or missing outputs:
Extra minutes per exception, excluding base review:
Monthly source upkeep and monitoring minutes:
New recurring cash cost:
Setup and training cost:
Chosen capacity value per hour and why:
Actual cash expense avoided, if any:
Hard errors that stop the pilot:
Base case / worse case / decision owner / review date:

Ask the implementer to fill this with measured pilot values and show the records behind the largest assumptions. Then change the exception rate together. If a modest change reverses the decision, continue a bounded trial or repair the input process before expanding the rollout.

Sources and dates

Sources checked on 20 October 2026. Worked examples are teaching examples unless explicitly attributed. Provider documentation describes product behaviour; the operating recommendations are DLVX editorial judgement.

• OpenAI: evaluation best practicesTask-specific evaluation, representative cases and comparison methods; our business worksheets are editorial recommendations. Checked 2026-10-20.