Which workflow deserves AI investment?
Follow the work through to approval

Consider recurring preparation of supplier information for a purchase decision: gather authorized documents, organize terms, identify missing information and assemble a traceable pack. A reviewer checks the sources and unresolved issues. An authorized person makes the commercial decision. AI assistance could support preparation; it does not establish that the information is correct or that a purchase should proceed. A genuine need starts with observed recurring work and a credible business or service outcome. Establish qualified demand, backlog or avoidable spending, the actual constraint and usable inputs authorized for this purpose. Name the business owner and staff who will use the process. Set acceptance criteria, bound the technical task and establish willingness to change the workflow. NIST's AI Risk Management Framework supports considering trustworthiness in AI use and evaluation. It is voluntary guidance, not proof that this workflow is safe.

Compare assistance with the approved tools and processes already available. A common intake form, better document ownership, simpler requirements or a changed approval schedule may solve the problem more cheaply. A bottleneck is not a buying instruction: if reliable AI cannot address it, choose another intervention. Keep four baseline records separate: staff effort per accepted pack, elapsed time from request to decision, accepted packs per week and actual cash spending. A preparation share of labor does not automatically describe its share of elapsed lead time. Waiting, parallel work and staffing can change the relationship.

Under explicitly assumed serial, no-overlap, unchanged-work conditions, let eligible work occupy 20% of total duration. Halving that task's duration gives 80% + 20% × 0.5 = 90% overall duration: 10% time saved and a rate gain of 1/0.90 − 1 = 11.11%. “50% less task time” is not “50% faster task processing.” A 50% higher task rate gives 80% + 20%/1.5 = 93.33% duration: 6.67% time saved and 7.14% rate gain, rounded. These are our planning calculations, not measured AI performance. Separately, assume a pack contains 10 total staff-hours, including 2 preparation hours. Halve preparation to 1 hour and add 0.25 incremental checking hours: 10 − 2 + 1 + 0.25 = 9.25 staff-hours, releasing 0.75 hours. Deduct only checks or rework not already included. This does not automatically shorten elapsed time, reduce payroll or establish cash ROI. Now use a separate capacity assumption, not a measurement derived from those 10 hours. Comparable accepted-step capacities per week, with unchanged case mix, staffing and quality, are preparation/review/approval 10/14/12; qualified demand is 18. If preparation becomes 20, capacities become 20/14/12. Maximum final-flow capacity rises from 10 to 12 accepted packs/week, +20%, not 2x. Approval becomes limiting. Downstream capacities already allow for checks and rework. If preparation was not limiting, or qualified demand is lower, faster preparation may add zero accepted output.

Cost follows the actual contract. For a metered service, token charges can be written as (I × rI + C × rC + O × rO)/1,000,000. I, C and O are distinct billed counts for uncached input, cached input and output; the r values are their applicable per-million rates, not quoted prices. Include other paid token categories, such as thinking where billed, without counting them again if already in output. Add metered tools, storage and actually billed retries where applicable; already counted retry usage is not another charge. The provider's API billing documentation distinguishes billing categories and differing tier/time windows. Use the categories, rates and effective period in your actual contract, then reconcile usage with the invoice. A fee schedule does not establish business benefit. For subscription, fixed-fee or local offers, use the actual subscription, fixed and operating costs instead of inventing a token bill. Local operation is not automatically cost-free. In one decision horizon, include one-time setup, equipment, integration and data preparation, plus recurring usage, operation, support, security, governance, training and incremental checks/rework. Do not count both equipment purchase and its depreciation in the same cash calculation, baseline review again or review already deducted from net savings. More completed purchase-decision packs are not themselves extra sales or contribution. For procurement, count verified avoided expediting, outsourced preparation or other actual outlay where applicable, including overtime, fees or hiring spending actually avoided. Do not monetize purchase volume or invent a margin per pack. Additional genuinely billable accepted work is a separate pathway: count its contribution margin, not gross revenue, only with qualified demand and attainable downstream approval. Identify who can realize each benefit. Released hours under unchanged salaries remain capacity, not cash savings. Public-sector service, backlog and uptime improvements matter without unsupported monetization.

Use a short decision worksheet in prose. Open with the recurring purchase pack, its observed volume and the outcome sought. Record authorized source availability, the business owner, users and approval authority. Describe the current constraint and the cheaper approved alternative. State what bounded preparation assistance must do, what it must flag and what evidence reviewers need to accept a pack. Missing terms should remain unresolved, not become confident wording. Next record the four baselines and proposed changes in their own units. Assign released hours to named work; test whether downstream capacity and qualified demand permit more accepted packs. Have finance identify verified avoidable procurement outlay, or evidenced contribution from genuinely billable additional work. Set the same horizon for benefit and complete incremental cost, including adoption and review effort. Distinguish assumptions from verified observations. For genuinely incremental accepted billable jobs with positive evidenced contribution margin m per job, break-even volume = remaining incremental cost/m, after other verified benefit. If no cost remains, no additional jobs are needed to cover it. Otherwise round whole jobs up. Compare required volume with both qualified incremental demand and attainable downstream capacity over the same period. Do not assign m to procurement packs that earn no contribution, or count costs already deducted in m again. ROI = (verified incremental benefit − total incremental cost)/total incremental cost, only for positive known cost and the same period and valuation basis. For a cash ROI, both sides must be cash-based. Purely assumed normalized benefit of 140 units and cost of 100 units gives (140 − 100)/100 = 40% ROI. This is unit arithmetic, not a currency quote, customer result or product price. Payback in months = upfront cash/(verified monthly benefit − recurring monthly cash cost), only when that monthly net is positive. With zero or negative net, this formula provides no payback period. A favorable assumed calculation is not evidence that the benefit will occur. Microsoft's cloud-adoption strategy guidance connects investment with measurable business goals and shared business, IT, finance and security ownership. It is not an AI ROI study. Apply that ownership discipline: name the person who changes scheduling or spending, set a review date and decide what evidence would justify proceeding, redesigning or stopping. The same method applies to legal matter preparation, banking documentation, pharmaceutical document assembly and energy/manufacturing records. Measure preparation, required review and authorized decisions separately. It does not delegate legal, credit, clinical or scientific judgment to an agent.