Time saved is not yet a realised financial benefit
This analysis examines 2025 AI investment decisions from a 2026 perspective. The case and amounts are teaching examples and do not claim to describe a CYTIZEN engagement. An assistant that shortens a task can release capacity, improve quality or reduce expenditure. These effects have different owners and financial treatments. Adding them together without explaining how they materialise turns a useful calculation into a fragile promise.
A business case starts with a specific decision: which use to fund, within what scope and with which stopping condition? It should not justify buying technology by adding every hour that might theoretically be affected. Describe the current journey, measure actual work and build an alternative without AI to understand the investment's incremental contribution.
Illustrative scenario: contract summaries lose their advertised advantage
Élodie, the legal director, reviews a contract-summarisation assistant. The provider advertises fifteen minutes saved per document. Marc, the financial controller, multiplies that figure by 12,000 annual documents and €60 per hour, producing €180,000 of gross value. The project appears profitable. Élodie asks whether every contract is comparable and whether checking the summary has been included.
A stratified sample shows that 8,000 documents are eligible for the proposed workflow. Summarisation reduces first-reading time by twelve minutes but adds four minutes of checking and an average minute of correction. Net savings are seven minutes. At €60 per hour, theoretical capacity is worth €56,000 annually: 8,000 × 7/60 × 60. An effective adoption assumption of 75% reduces that valuation to €42,000.
Marc corrects the model. No headcount reduction is planned, and lawyers do not invoice for each internal document. The main gain is therefore not €42,000 in cash savings. It is usable capacity, conditional on reducing a backlog or completing more work. Actual savings may exist if identifiable external services are avoided, but those savings need their own evidence.
A completed calculation with visible limitations
| Annual assumption | Illustrative value | Check before commitment |
|---|---|---|
| Eligible documents | 8,000 | Exclude degraded scans, untested languages and complex annexes |
| Gross reading saving | 12 minutes/document | Compare equivalent tasks |
| Checking and correction | 5 minutes/document | Include unusable outputs and silent errors |
| Effective adoption | 75% | Active use on eligible documents, not accounts created |
| First-year cost | €68,000 | Integration €25,000; preparation €12,000; subscription €18,000; operation and evaluation €13,000 |
| Annual recurring cost | €31,000 | Subscription plus operation; confirm volume and terms |
The case presents €42,000 of valued capacity, no committed payroll saving and a reduction in response time that remains to be demonstrated. It does not count the same hour as capacity, avoided cost and additional revenue. First-year cost exceeds the capacity valuation. That does not automatically rule out investment, but the sponsor must explain the strategic reason and non-financial benefits.
A €68,000 expenditure does not become acceptable by presenting a return based on the original €180,000 estimate. It may be justified if it resolves a specific bottleneck, if an alternative costs more or if a service requirement warrants the expenditure. In each case, the sponsor explicitly accepts that decision rather than hiding its rationale inside an inflated monetary benefit.
Measure the right counterfactual
Compare the assistant with an improved process, rather than only a defective existing one. A clause library, a better request form or a triage rule may reduce processing time without a generative model. For the example, test three versions: current workflow, simplified workflow without AI and simplified workflow with AI. The difference between the latter two estimates the assistant's own contribution.
Allocate documents to users with comparable experience and retain categories of complexity. If simple cases go to the AI group and difficult cases to the control group, the result cannot support the decision. Do not let the project select only examples for which it already knows the answer. Prepare a control set that has not been used to configure the tool.
Processing time starts when the user takes the document and ends when the output is acceptable to its recipient. Count launch, context retrieval, source checking, corrections and rework. A measure that ends when the summary appears measures generation speed rather than legal work. Record interruptions consistently and avoid attributing unrelated delays to whichever group happens to experience them.
Assess risk without inventing numerical precision
An omitted clause can cost more than many hours saved. Multiplying an arbitrary probability by a spectacular loss does not make the risk measurable. Separate observed events, scenario assumptions and risks that cannot yet be quantified. The legal owner defines unacceptable errors; the sponsor decides whether the pilot can continue with adequate controls.
In the illustrative test of 200 contracts, twelve summaries require material correction and two omit an obligation considered critical. These counts cannot precisely estimate a rare risk across the whole portfolio. They are sufficient to prevent unchecked distribution within the tested scope and to require a cause analysis. The material-correction rate is 12/200, or 6%, using a definition of materiality written before testing.
Also include the cost of returning to manual work, provider dependency, model changes and confidentiality constraints. Not every risk should be converted to euros. Some become launch conditions: authorised data, a reversible output, a known version and the ability to stop without blocking the service. A risk marked as non-financial still needs an owner and an acceptance decision.
Test sensitivity before announcing a return
The cautious case assumes 6,000 eligible documents, five net minutes and 60% adoption. It produces €18,000 of valued capacity. The central case produces €42,000. The favourable case assumes 10,000 documents, nine minutes and 85% adoption, or €76,500. The recurring €31,000 cost is covered by capacity valuation only in the last two cases; none of these comparisons automatically represents a cash flow.
Ask which variables can be influenced. Training may improve adoption; document standardisation may increase eligibility; more demanding review may reduce net savings. If attractiveness depends entirely on the favourable case, fund a bounded pilot instead of a complete rollout. An illustrative continuation condition is at least five net minutes saved, no critical omission in the control set and support effort within the planned allowance.
The pilot decision specifies a maximum budget, admitted data, document count, users and a date for reviewing results. Observing no critical error is not a statistical guarantee of no future error. It is one condition among several, accompanied by review of each sensitive output. If the control set is too small to address a material uncertainty, the decision should state that limitation rather than present a definitive percentage.
Separate first-year cash requirements from recurring economics. A recurring benefit can eventually exceed recurring expenditure while the initial investment still takes time to recover. Conversely, an attractive first-year discount may conceal rising renewal costs. Include renewal terms and the workload needed to maintain test sets, permissions and approved document sources.
Name the benefit owners
Élodie owns legal capacity and its reassignment: reduce the older-contract backlog, handle urgent requests sooner or review more clauses. Marc approves the valuation method and distinguishes monetary flows. IT owns operating and integration costs; procurement owns pricing terms and contractual exit. One person maintains the model but does not approve on behalf of every benefit owner.
The first post-launch review reconciles active use, net time, errors, costs and the work actually completed using released capacity. If 500 hours are reported but the backlog does not fall and no additional activity is identified, the benefit has not yet been realised. Greater complexity may have absorbed it. Explain that with data rather than removing the discrepancy from the case.
The business case remains revisable. A supplier price increase or a change in checking requirements can alter its economic value. Before launch, set the threshold that triggers a new decision so that sunk expenditure does not become the only reason to continue. Keep the original baseline and successive revisions available: otherwise changing assumptions can make an underperforming investment appear permanently on target.
Benefits are not counted identically in every function
In procurement, less time spent reviewing suppliers becomes a saving only when it avoids expenditure or releases capacity that is actually used. Discounts secured through negotiation should not automatically be attributed to AI. Compare outcomes and explain other factors, including market conditions and the negotiator's own intervention.
In quality, a fast answer has value only if it identifies the approved procedure and reduces rework without weakening controls. Additional checking time may be justified by risk and belongs in the net calculation. An assistant that increases the number of drafts but also increases correction work may be useful for some users and uneconomic for others.
For customer service, track cost per correctly resolved request, including reopenings and escalations. A shorter first interaction accompanied by twice as many callbacks can worsen both service and economics. Satisfaction and response times are separate outcomes to measure, rather than values to convert arbitrarily into revenue. An improved experience can support a strategic choice without an invented sales uplift.
Sources and method
Primary sources consulted on 4 October 2026: HM Treasury Green Book; HM Treasury Magenta Book; NIST Generative AI Risk Management Profile. These references inform evaluation and risk discussion; they do not validate the scenario's amounts.
The method separates capacity, cash, quality and risk, compares alternatives and tests sensitive assumptions. Every numerical example is illustrative. Actual financial decisions must follow the organisation's rules, rates, tax treatment and contractual terms. This article supports a verifiable investment discussion rather than promising a return on investment. Its calculations should be reproduced using measured local data before any commitment is made.