An AI system can make a task faster without improving the company's cash position. It can also improve service quality without reducing anyone's workload. Start by naming the benefit and the mechanism that turns it into a business result.
This chapter gives the guide's common vocabulary for that work. Every workflow chapter applies it to a specific case, and the economics worksheet turns it into a procedure. The numbers below are illustrative assumptions; none is a measurement from any company. Accounting references are scoped to what their sources say, and no accounting opinion is given.
Three separate ledgers
Operational effect: minutes per completed task, correction rate, turnaround time, throughput, or another observable change.
Economic effect: additional capacity, avoided contractor spending, avoided future hiring, additional contribution from sales, or another measurable benefit. Name the decision needed to realize it.
Accounting presentation: where costs and benefits appear in this company's reporting. The work's relationship to the product being sold matters. A draft proposal prepared by an internal sales team and a client deliverable prepared by an agency may use similar tools but have different economic purposes. A controller must determine the applicable accounting treatment under the relevant reporting standards and transaction facts; tagging the software itself as COGS or OPEX does not settle the question.
For every initiative, record the expected benefit location separately from the cost location. An operating-cost increase can accompany a delivery-cost reduction. Revenue growth can require additional delivery expenditure. Working-capital improvements are not interchangeable with profit improvements.
Three ledgers, connected only by named decisions
Operational effect
- Minutes per completed task
- Correction rate
- Turnaround time
- Throughput
Only through a named decision that realizes it, such as cancelling contractor hours or not filling a planned role
Economic effect
- Additional capacity
- Avoided contractor spending
- Avoided future hiring
- Additional contribution from sales
Determined by a controller applying the reporting framework to the transaction facts
Accounting presentation
- Where costs and benefits appear in this company’s reporting
- Depends on the work’s relationship to the product sold
Not a link: tagging the AI tool as COGS or OPEX. Working-capital improvements are also not interchangeable with profit improvements.
The three ledgers fail in different ways. An operational claim fails when the measurement is wrong: the baseline excluded follow-ups, or the pilot handled easier work. An economic claim fails when no decision converts the operational change into money or value. A presentation claim fails when someone treats a label as an analysis. Keep a separate line in the business case for each, so a reviewer can challenge one without dismissing the others.
Where benefits land
The guide uses five benefit locations. Each has its own mechanisms, its own evidence, and its own way of being overstated.
| Location | Mechanisms | Example in this guide | How it is overstated |
|---|---|---|---|
| Cost of delivery | Less effort per delivered unit; less rework; avoided unpriced scope | Proposals: a $24,000 fee leaves $8,000 contribution at 160 delivery hours, and nothing if 80 unpriced hours slip into scope | Counting faster drafting without checking what was committed |
| Operating expense | Added capacity in support, sales, finance, or operations; avoided contractor, overtime, or hiring spend | Support replies: 120 hours a month released, worth $5,400 at an assumed $45/hour | Treating released hours as cash |
| Revenue | Higher conversion, retention, or expansion; more paid work; a new product | Professional services: released hours on a time-and-materials engagement reduce billed revenue unless redeployed or repriced | Valuing every AI-touched sale at full price instead of incremental contribution |
| Working capital | Less inventory or faster collection for the same service level | Inventory forecasting: a $50,000 inventory reduction is released once; its carrying value at an assumed 20% is $10,000 a year | Adding the one-time release to annual savings |
| Risk exposure | Fewer severe errors, earlier detection, a control that did not exist | Catalog enrichment: an uncertified safety claim is never published | Manufacturing a precise expected loss from a rare event |
A single initiative can touch several rows. The support assistant releases operating capacity, may improve retention, and adds a control against wrong-version instructions. Record each mechanism separately, attach its own evidence, and decide which ones the business case actually rests on. A case that only works when every row is counted at its most optimistic value is a hypothesis, not a plan.
A capacity calculation
The support-reply chapter works one example in full, and the economics worksheet derives it step by step. A team handles 5,000 requests a month; in an illustrative pilot, 60% are eligible and assistance is used on 80% of those, so 2,400 requests are assisted. Time per request falls from eight minutes to five, including review and correction, which releases 120 hours a month. At an assumed loaded labor value of $45/hour, that represents $5,400 of monthly capacity. If the recurring budget for software, knowledge upkeep, review, and monitoring is $2,000, the capacity value after those costs is $3,400. Setup costs $12,000. This example assumes the whole $2,000 is new cash spending, for instance contracted upkeep and review, rather than work absorbed by existing staff.
Capacity value is not cash. If payroll, overtime, and contractor invoices remain unchanged and the released time is not put to productive use, there may be no offsetting cash benefit at all. Record the $2,000 cash outflow and the 120 hours separately.
Suppose instead 70% of the capacity translates into a demonstrably avoidable expense at the assumed rate. Realized monthly benefit would be $3,780, net recurring benefit $1,780, and simple setup payback approximately 6.7 months. That scenario requires evidence for the 70% conversion, not merely managerial optimism.
The same 120 hours, three different answers
- Capacity value
- Deduction
- Monthly result
Show the data
| View | Capacity value | Not realizable | Recurring cost | Monthly result | Setup recovered |
|---|---|---|---|---|---|
| Capacity view | $5,400 | — | −$2,000 | $3,400 | 3.5 months (capacity, not cash) |
| Cash, 70% realizable | $5,400 | −$1,620 | −$2,000 | $1,780 | 6.7 months |
| Cash, nothing realizable | $5,400 | −$5,400 | −$2,000 | −$2,000 | Never |
How capacity becomes money
Released time turns into cash only through a decision someone can name. Four mechanisms recur in the guide, and each needs different evidence.
Avoided contractor or overtime spending. The most direct route, because the expense already exists and stops. In the finance leaders guide, 100 released hours worth $5,000 at an assumed $50/hour produce $2,000 of avoided expense, because only 40 of those hours replace paid overtime. Evidence: the invoices or timesheets that would otherwise have been paid, and someone with authority to stop them.
Avoided hiring. Released time replaces a planned hire. Evidence: an approved hiring plan that changes because of the pilot, and enough released time to matter. Small amounts spread across many people may never add up to a role; the onboarding chapter notes that its modest volume may not reach a whole avoided hire.
Redeployment to paid work. Released hours move to work someone pays for. In the professional-services guide, 20 hours redeployed at a $250 rate add $5,000 of revenue. Evidence: the paid work actually exists and would otherwise have been declined or delayed.
Demand absorbed without growth in headcount. Volume rises and the team handles it without adding people. This is real value, but it is a counterfactual: the evidence is a credible forecast of the hiring that would otherwise have happened, set before the pilot.
Each mechanism needs an owner and a date. A business case that says "capacity will be redeployed" without saying to what, by whom, and when has a capacity value and no cash value. The onboarding and knowledge retrieval chapters show how realization changes the answer: onboarding is negative in cash at 50% realization and needs about 60% to break even, and knowledge retrieval needs about 52%.
The downside case belongs beside the base case
If assisted time is seven minutes rather than five, released capacity falls to 40 hours, worth $1,800 at the same assumed rate. Recurring costs now exceed that capacity value by $200. A seemingly small change in review effort can reverse the result.
In cash terms the same downside is worse. At the 70% realization assumption, the worksheet example's net monthly cash falls from $1,780 to −$740. Capacity breaks even at about 1.11 minutes released per assisted request; cash at the same realization needs about 1.59. The gap between those two thresholds is the room realization takes up. The economics worksheet shows which input moves the result most; in this example it is the two effort assumptions, because they act on a small difference between them.
Measure the whole workflow. Saving drafting time while adding approval, integration repair, or downstream correction work can move work between teams without eliminating it. Include the cost of quality incidents and monitoring where material. Keep one-time setup separate from recurring expense, and do not count review labor twice if it is already inside assisted handling time.
What it costs
In this guide's worked examples, the costs that matter most are continuing work rather than the license.
| Category | One-time or recurring | What it includes |
|---|---|---|
| Software and usage | Recurring | Licenses, model usage, hosting |
| Integration | Both | Connectors, permission mapping, write paths and their monitoring |
| Data and knowledge preparation | Both | Labeling, cleaning, retiring stale material, keeping sources owned |
| Evaluation | Both | Building the case set, scoring, re-running after changes |
| Review | Recurring | Human checking, already inside assisted minutes if measured that way |
| Maintenance | Recurring | Prompt and configuration updates, provider changes, incident handling |
| Training | Both | Initial rollout and refresh for new staff |
The support chapter splits its illustrative $2,000 monthly budget into $800 for software and usage, $600 for knowledge upkeep, $400 for quality review, and $200 for integration monitoring. The split is invented, but its shape is the point: half of the continuing budget in that example pays for people maintaining sources and reviewing quality, not for the software. The calculation above treats all of it as new cash spending. If existing staff absorbed the $1,000 of upkeep and review instead, that $1,000 would be consumed capacity, and net monthly cash at the 70% realization assumption would be $2,780 rather than $1,780; the capacity value after recurring cost would not change.
When existing staff absorb continuing work, record it as consumed capacity, not as cash. Counting the same hours as both a cash expense and a labor deduction overstates the cost; counting them as neither understates it.
Double counting
The same benefit can enter a business case more than once. The traps the guide's chapters flag:
- Review counted twice. If assisted minutes include review, a separate review line repeats it.
- Capacity and cash from the same hours. The $5,400 capacity value and the $3,780 realized at 70% describe the same 120 hours; they are not additive.
- Working capital as annual savings. The forecasting chapter's $50,000 inventory release happens once. At an assumed 20% carrying rate it is worth $10,000 a year, and its $6,000 annual operating cost leaves $4,000; the break-even reduction is $30,000. Adding the $50,000 to the annual figure is a different kind of number added to the wrong period.
- One missed order counted twice. A stockout avoided is either expedite spending avoided or margin preserved, not both.
- Revenue at full value. An extra sale is worth its contribution after delivery cost, not its price.
- The counterfactual twice. If headcount would have grown anyway, "demand absorbed" and "avoided hire" are the same benefit.
Billing models change the sign
Where work is billed by the hour, faster work can reduce revenue. In the professional-services guide, an illustrative 200-hour engagement at $250 an hour releases 20 hours. Under time and materials, if the hours sit idle the engagement's contribution falls by $5,600; if they are redeployed to other paid work the change is −$600. Under a fixed fee with redeployment, contribution rises by $4,400. If the client reprices the fixed fee to the reduced hours, the change is −$600, the same as time and materials with redeployment.
Neither model is better in general. The point is that the same operational improvement can be positive, negative, or neutral depending on how the work is sold, and the business case must say which applies.
Classification is a separate question
For your finance lead
Where costs and benefits appear in the financial statements depends on the applicable reporting framework, the facts of the arrangement, and the company's policy. The guide gives examples of what sources say; it does not give an accounting opinion. Those questions belong with a qualified reviewer who knows the company's framework and the facts of the arrangement.
For an IFRS example, IAS 2 includes purchase, conversion, and other qualifying costs of bringing inventory to its present location and condition; it recognizes the carrying amount as expense when the inventory is sold. This does not classify an AI invoice by itself. IFRS Foundation, IAS 2 overview
Software arrangements raise a different question. The IFRS Interpretations Committee's 2021 SaaS configuration/customisation decision considers whether the customer controls a qualifying separate asset and, if not, when services are received. That fact pattern illustrates why setup spending is not automatically capitalized. It is not a universal rule for AI projects or US GAAP. IFRS agenda decision, questions I–II
US GAAP needs a separate assessment. FASB's ASU 2018-15 addresses qualifying implementation costs for cloud service arrangements. ASU 2025-06 changes internal-use software guidance for annual periods beginning after December 15, 2027, with early adoption permitted. Record the entity's adoption status before applying a recognition checklist. Neither update supplies a universal COGS/OPEX label for AI.
Treat company policy as an application of the relevant reporting framework, not a free choice of COGS, OPEX, or capitalization. The economic examples in this guide do not determine financial-statement treatment, and the finance leaders guide describes how to keep the two conversations apart.
A reusable decision worksheet
| Input | What to record |
|---|---|
| Volume | Completed units per period, including seasonal variation |
| Eligible share | Fraction of units within the pilot's approved scope, from its written eligibility rules and exclusions |
| Adoption share | Fraction of eligible units actually processed with assistance, including declined suggestions |
| Baseline time | End-to-end human effort for a comparable case mix |
| Assisted time | Human effort after assistance, including review and rework |
| Quality | Errors by severity and downstream consequences |
| Unit value | Labor value or contribution basis, with its limitations |
| Realization | The specific expense avoided or productive use of capacity |
| Recurring cost | Licenses, usage, maintenance, evaluation, training refresh |
| Setup | Integration, preparation, rollout, and initial training |
| Sensitivity | What happens when volume falls or review doubles |
Revenue applications need a different numerator: incremental contribution after delivery costs, not the full value of every AI-touched sale. Establish a comparison that can distinguish the system's effect from seasonality, pricing changes, or a stronger sales team. Until there is evidence, higher conversion remains a hypothesis.
For risk applications, describe the loss mechanism and uncertainty. Do not manufacture precise expected-loss reductions from rare incidents. A system can have a worthwhile control benefit before a defensible monetary estimate exists.
The decision rule
Proceed when the observed workflow improvement, a credible realization mechanism, and acceptable quality justify the full cost. If the business case depends on unmeasured accuracy, perfect adoption, or removing all review, the next investment should be in evidence rather than expansion.
In practice that means asking four questions in order. Is the operational effect measured on comparable work, including review and rework? Does a named decision convert it into money or a stated value, with an owner and a date? Is every continuing cost funded, including the people who maintain sources and run evaluations? Does the case survive its downside, and does anyone know which assumption it is most sensitive to? A "no" to any of them does not mean stop; it names the next thing to measure.
Some projects are worth doing on non-financial grounds: a control that prevents a rare but severe error, a better experience for customers, or learning that informs a larger decision. Say so plainly and fund them as such, rather than forcing them into a savings case they cannot support.
Related chapters: economics worksheet, finance leaders, software support, services proposals, order intake, inventory forecasting.