Generative AI has moved from experiment to budget line, and finance has caught up. In PwC’s 2026 Global CEO Survey, reported by ITPro in January 2026, 56% of CEOs said AI had not yet delivered significant financial benefit, and only about one in eight reported both revenue gains and cost reductions. Gartner predicted in mid-2025 that more than 40% of agentic AI projects will be canceled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. CFOs have read these numbers. The proposal that worked in 2023, an impressive demo plus a claim that staff will be 30% more productive, no longer clears an investment committee.
That is good news for well-run AI programs. A business case that finance can test is also one that forces the team to choose the right problem, design for adoption and know when to stop. This guide walks through how to build one for a generative or agentic AI initiative in 2026: what CFOs now look for, how to model value and cost honestly, and how to structure funding so that approval is a sequence of small, evidence-based decisions rather than one big bet. (Tracking value after a pilot goes live is a related but separate discipline; this article is about building and approving the case.)
Why Generative AI Business Cases Get Rejected
Most rejected proposals fail for predictable reasons, and few of them are about the technology.
- Value expressed as “hours saved.” Time saved is not cash unless it changes headcount plans, overtime, outsourcing spend, cycle time that customers pay for, or capacity that is actually redeployed. Finance discounts unconverted hours heavily, often to zero.
- No baseline. If nobody knows today’s cost per ticket, error rate or cycle time, no improvement can be proven later.
- Costs limited to licenses. Integration, data preparation, evaluation, security review, change management and ongoing operations often exceed the model or seat cost.
- Single-point forecasts. One optimistic number invites skepticism. Ranges with explicit assumptions invite discussion.
- Solution looking for a problem. MIT NANDA’s widely cited 2025 report on enterprise generative AI observed that roughly half of AI budgets flowed to sales and marketing, while some of the clearest returns came from back-office work such as reducing outsourcing and agency spend. Whatever one thinks of the report’s headline “95%” figure, the pattern of funding visible use cases over measurable ones is familiar.
- No owner on the business side. A case sponsored only by IT signals that nobody will change the process to capture the value.
What Has Changed for AI Business Cases in 2026
From seats to consumption
Early generative AI spending was dominated by per-user copilot licenses. Agentic systems are different: they consume model tokens, tool calls and compute on every task, and an agent that reasons through ten steps costs far more than a single chat reply. Costs become variable and usage-driven, which finance understands well, but only if the case models cost per unit of work, not cost per month.
Falling unit prices, rising volume
Epoch AI’s analysis of LLM inference prices found that the cost to reach a fixed level of performance has fallen by between 9x and 900x per year depending on the task, while cautioning that the steepest declines may not persist. That is a strong argument for not over-engineering cost today. It is not an argument for ignoring cost: cheaper tokens tend to be consumed by longer contexts, more agent steps and broader rollout. A sound case shows sensitivity to both price and volume.
Higher scrutiny on agentic claims
Gartner has warned about “agent washing,” estimating that only about 130 of the thousands of vendors claiming agentic capabilities actually deliver them. CFOs increasingly ask whether a proposed “agent” is genuinely making decisions inside a workflow or is a chatbot with a new label. The case should state plainly what the system will do autonomously, what humans will still review, and why that split is safe.
Governance is now part of the cost and the value
Regulation such as the EU AI Act and internal AI policies mean that risk assessment, testing, logging and human oversight are not optional extras. They belong in the cost model. PwC’s survey also found that CEOs whose organizations had strong AI foundations, including responsible AI frameworks, were about three times more likely to report meaningful financial returns, so governance is a value enabler, not only a tax.
Step 1: Start From a Business Problem With a Measurable Baseline
Pick a workflow, not a technology. Good candidates share four traits: high volume, measurable outcomes, a clear owner, and work that is partly language-heavy (reading, writing, classifying, summarizing, deciding against rules). Examples include invoice exception handling, customer service resolution, contract review against a playbook, claims intake, IT service desk triage and sales proposal drafting.
Then capture the baseline before anyone writes code:
- Volume: units of work per month (tickets, invoices, contracts, claims).
- Unit cost: fully loaded labor and outsourcing cost per unit, including rework.
- Cycle time: elapsed time from request to resolution, and what that delay costs (late-payment penalties, lost discounts, churn, working capital).
- Quality: error, escalation, reopen or complaint rates.
- Constraints: backlogs, seasonal peaks, hiring difficulty or service-level breaches that limit growth.
If the data does not exist, a two- to four-week time-and-motion sample on a few hundred cases is usually enough. It is far cheaper than defending an unmeasurable claim later.
Step 2: Map Value to Specific Financial Levers
Translate expected improvements into lines a CFO recognizes, and state the mechanism by which each becomes cash.
- Cost avoidance: absorbing volume growth without proportional hiring, or reducing outsourcing and contractor spend. This is often the most credible lever because it does not require redundancies.
- Direct cost reduction: lower external spend (agencies, translation, BPO), fewer penalties, less rework.
- Revenue: faster quote turnaround, higher conversion, reduced churn, new chargeable services. Treat revenue claims with the most caution and tie them to a measurable leading indicator.
- Working capital: faster invoicing, collections or claims settlement that change cash timing.
- Risk reduction: fewer compliance errors or missed obligations. Express it as expected loss avoided, with conservative probabilities, or keep it qualitative.
For each lever, name the business owner who will make the change happen, for example a service director who agrees not to backfill attrition, or a procurement lead who will reduce an outsourcing contract at renewal. Without that commitment, the benefit is a hope.
A generative AI business case is approved on mechanisms, not on percentages: who will change what, when, and how the money shows up.
Step 3: Build an Honest Total Cost of Ownership
Model costs for at least three years, separating one-time build from recurring run costs.
One-time costs
- Discovery, process redesign and use-case validation.
- Data preparation: cleaning, permissions, document conversion and indexing.
- Integration with core systems (ERP, CRM, ticketing, document management) and identity.
- Building the evaluation set and acceptance criteria.
- Security, privacy and AI risk assessments; legal review of vendor terms.
- Change management and training for users and supervisors.
Recurring costs
- Inference: cost per task, modeled as tokens per step × steps per task × tasks per month, with a range for retries and long cases.
- Platform and licenses: AI gateway, vector store, observability, orchestration, seats.
- Human review of outputs that the design keeps in the loop.
- Evaluation and monitoring: re-running evals when models, prompts or data change.
- Model and vendor changes: migrations when models are deprecated or prices change.
- Support and continuous improvement by a product team.
Present the result as cost per unit of work alongside the baseline unit cost. A finance team can reason about “AI-assisted cost per invoice of X versus a current Y” far more easily than about a monthly platform bill.
Step 4: Model Scenarios, Adoption and Ramp
Benefits never arrive on day one. Build three scenarios with explicit assumptions:
- Conservative: low automation rate, slow adoption, higher human review, inference prices flat.
- Expected: the rates observed in the pilot, normal adoption curve.
- Optimistic: higher automation as the system improves, faster rollout, falling prices.
The most sensitive assumptions are usually the share of cases the system handles end to end, the residual human review time on the rest, and adoption, meaning the share of eligible work that actually flows through the new process. Show a sensitivity chart for these. If the investment only pays back in the optimistic scenario, say so; it may still be worth a small, gated experiment, but not a full program.
Be explicit about the ramp: months to production, months to full adoption, and when each benefit lever starts. A realistic payback for a well-chosen workflow is often within 12 to 24 months, but your case should derive this from your numbers rather than borrow a benchmark.
Step 5: Fund in Stages With Kill Criteria
The single most effective way to win CFO approval is to ask for less money with clearer decision points. Structure funding as gates:
- Gate 0 – Validate (4–6 weeks): baseline measured, eval set built, prototype tested on historical cases. Pass criterion: offline quality above an agreed threshold and unit-cost model within target range.
- Gate 1 – Pilot (8–12 weeks): limited production with real users and human review. Pass criteria: measured quality, user adoption, observed cost per task and no unresolved risk findings.
- Gate 2 – Scale: rollout across teams or regions, with benefit owners executing their committed changes (hiring plans, contract renewals).
- Gate 3 – Optimize: increase automation share, move suitable work to cheaper models, retire legacy steps.
Each gate has explicit kill or pivot criteria, such as quality below threshold after two iterations, cost per task above the baseline, or adoption below an agreed floor. Pre-agreeing them turns a potentially awkward cancellation into a planned, cheap outcome, and it is precisely the discipline that would have prevented many of the cancellations Gartner predicts.
Worked Example: Invoice Exception Handling
Consider a generic shared-services center processing 40,000 supplier invoices a month, of which 18% raise exceptions (price or quantity mismatches, missing purchase orders, unclear tax). Each exception currently takes an average of 14 minutes of analyst time, and late payments trigger some penalties and lost early-payment discounts. Volume is growing about 10% a year, and the team plans to hire four analysts next year to keep up.
The proposal is an agentic assistant that reads the invoice, purchase order and receipt, explains the mismatch, drafts a resolution or supplier query, and routes it for one-click analyst approval. The business case is built as follows:
- Baseline: 7,200 exceptions a month × 14 minutes, a fully loaded hourly cost from finance, plus last year’s penalties and missed discounts.
- Mechanism: the shared-services head commits to absorbing growth without the four planned hires (cost avoidance), and the treasury lead tracks discount capture (working capital and direct savings).
- Assumptions: in the expected scenario, the assistant reduces average handling to 6 minutes on 70% of exceptions; the remaining 30% are complex and unchanged.
- Costs: build (integration with the ERP, eval set of 300 historical exceptions, security review, training), run (inference per exception from pilot measurements with a buffer, platform fees, a part-time product owner and monitoring).
- Gates: Gate 0 tests the prototype against the 300 historical cases; Gate 1 runs with one team for ten weeks; kill criteria include accuracy of proposed resolutions below the agreed threshold or average handling time not falling below 9 minutes.
Notice what the case does not claim: no vague productivity percentage, no revenue upside, no headcount reduction. It asks for a modest first tranche, shows the cost per exception before and after, and ties the main benefit to a hiring decision that finance can see in next year’s plan.
The One-Page Summary Finance Wants
However detailed the model, lead with a single page:
- The business problem and its current cost, with baseline data.
- What will change in the process, and who owns that change.
- Benefit levers with ranges and the mechanism that turns each into cash.
- Three-year cost of ownership and cost per unit of work, before and after.
- Scenarios, payback range and the three most sensitive assumptions.
- Funding requested for the first gate only, with pass and kill criteria.
- Key risks (quality, adoption, regulatory, vendor) and their mitigations.
Common Pitfalls
- Counting the same hour twice across overlapping use cases.
- Assuming 100% adoption because the tool is available.
- Ignoring the cost of keeping humans in the loop, which the design may require for years.
- Treating vendor list prices as fixed, in either direction.
- Leaving governance and evaluation costs out, then cutting them under budget pressure.
- Approving a portfolio of 30 small pilots instead of funding three cases through to scale.
Conclusion
In 2026, the generative AI business case is a test of management discipline as much as of technology. Start from a costly, measurable workflow; capture the baseline; map benefits to financial levers with named owners; model full cost per unit of work; show ranges instead of promises; and ask for funding one gate at a time with agreed kill criteria. A case built this way is easier to approve, easier to stop if it is not working, and much more likely to show up in the P&L when it does.
How Hanoi Technology Can Help
Hanoi Technology helps leaders validate generative AI use cases, measure baselines and build stage-gated business cases, then delivers the solution through projects, dedicated teams or managed services.