To create a cost-benefit analysis with AI, define the decision and the no-action baseline first, then build an evidence register and perform every financial calculation in a reviewed spreadsheet or model. Use AI to challenge missing categories, standardize descriptions, and draft an explanation from verified figures. Do not let it invent values, choose the preferred option, or hide uncertainty behind polished prose.

This is an appraisal workflow, not a substitute for financial review. Start with the broader guide to using AI responsibly, use the spreadsheet-analysis workflow for deterministic calculations, and keep AI privacy risks in view before sharing internal figures.

Do not put names, health or grievance information, salary records, protected characteristics, identifiable performance records, or investigation evidence into a consumer AI service. Use only an organization-approved environment and the minimum de-identified fields needed for the analysis.

Key Takeaways

Compare every option with a dated no-action baseline.

Keep source facts, calculations, assumptions, and judgments in separate fields.

Include costs, benefits, risks, timing, non-monetary effects, and affected groups.

Test the assumptions that could reverse the recommendation.

Treat AI output as a draft and question generator, never as approval.

What decision and no-action baseline should you define first?

A useful analysis begins with one decision sentence: who must decide what, by when, for which population or business unit, and within which constraints. “Should we replace the support platform this quarter?” is bounded. “Is the new platform worth it?” leaves the model free to change the scope.

Write the business-as-usual or no-action option before listing alternatives. It describes what happens if current arrangements continue, including maintenance cost, expected deterioration, contractual obligations, and foregone improvements. The UK Green Book treats business as usual as a benchmark that remains in the shortlist so other options have a consistent comparison point.［1］

Record objectives and critical success factors separately. An option that fails a mandatory objective should not win merely because its estimated ratio looks attractive. Also state the appraisal period, currency basis, tax treatment, discounting method, inflation convention, and decision threshold. Those choices belong to the accountable finance or decision owner, not the model.

Build options and an evidence register

Include genuinely different options rather than cosmetic variations. A compact longlist might contain no action, a minimum change, the proposed change, and a materially different delivery approach. Use project-planning guidance later for delivery sequencing; the analysis here only compares plausible choices.

Create one evidence row for each cost, benefit, risk, or impact:

Field	What to record
Option	Stable option ID and short description
Type	Cost, benefit, risk cost, non-monetary impact, or transfer
Amount and unit	Verified value, currency or operational unit, and price basis
Timing	Start, duration, recurrence, and appraisal horizon
Source	Source ID, owner, extraction date, and location
Assumption	What must be true for the estimate to hold
Confidence	High, medium, or low with a reason
Range	Low, central, and high value where uncertainty is material
Owner and reviewer	Person accountable for the input and its review

Keep estimates at a useful grain. Combine trivial rows, but do not merge items with different timing, evidence quality, owners, or uncertainty. Preserve one-time implementation cost, recurring operating cost, transition overlap, training, support, decommissioning, and exit obligations as distinct categories.

Benefits need the same discipline. Name the affected process, measurement method, baseline, expected change, ramp-up period, and evidence. “Save time” is not an estimate. “Reduce manual review by 20 hours per month after a three-month ramp-up, based on the timed pilot” is testable.

Where should you calculate costs and benefits?

Use a controlled spreadsheet, reviewed script, or approved financial model for arithmetic. Store formulas, units, source IDs, and version information with the calculation. AI may explain a result, but the approved number should be reproducible without asking the model again.

Check for common structural errors:

counting the same benefit under revenue, productivity, and avoided cost;
combining nominal and real amounts;
applying inconsistent time horizons across options;
treating a transfer between parties as a net benefit without analysis;
omitting transition, termination, monitoring, or risk costs;
presenting an average while hiding who bears the cost;
using a precise figure when the source supports only a range.

The Green Book separates monetizable effects from effects that cannot credibly be expressed in money and requires risk, uncertainty, public financial impact, and distributional effects to remain visible.［1］ A single benefit-cost ratio cannot carry all those judgments.

If you need charts, build them only after the table is checked, following the data-visualization workflow. A polished chart cannot repair an incorrect denominator.

Create an assumption register for a cost-benefit analysis with AI

An assumption register makes the reasoning inspectable. Give AI only approved, minimized evidence and ask it to find implied assumptions, missing owners, inconsistent units, unsupported causal links, and ambiguous timing. It should return questions, not fill the gaps.

Use this prompt contract:

Review the option and evidence rows below. List every explicit or implied assumption. For each, return an assumption ID, the affected option and row, supporting evidence, confidence, plausible low and high values, owner, validation method, and review date. Do not invent a value, calculate a result, rank options, or recommend approval. Mark missing evidence as “unverified.”

Then review every generated item. Remove assumptions that are not grounded in the actual proposal, merge duplicates, and add any domain-specific constraints the model missed. A useful register includes demand, adoption, delivery duration, staff availability, unit cost, useful life, failure rate, benefit ramp-up, dependency, and residual-value assumptions.

The GAO Cost Estimating and Assessment Guide emphasizes documenting assumptions, data, methods, sensitivity, and risk so an estimate can be traced and updated.［3］ That traceability is more valuable than a longer narrative.

Test sensitivity, scenarios, and optimism bias

Sensitivity analysis asks which uncertain inputs matter enough to change the decision. Choose variables because they are uncertain and material, not because they are easy to edit. Test one variable at a time to understand leverage, then combine coherent variables into scenarios.

At minimum, calculate:

a central case using the reviewed working assumptions;
a downside case with plausible adverse values;
an upside case that remains evidence-based;
switching values for assumptions that could make the preferred option cease to meet the threshold.

The Green Book calls for sensitivity analysis and switching values, and it requires explicit adjustment for optimism bias—the tendency to underestimate costs or duration and overestimate benefits.［1］ Its supplementary guidance provides structured approaches for those adjustments.［2］ Your organization may use different methods, but it should not silently assume the central case is certain.

Ask AI to summarize the completed scenario table with a strict rule: state which inputs changed, show the direction of impact, preserve the approved figures, and identify which assumption crosses the decision threshold. Verify that summary line by line against the model output.

Why keep non-monetary and distributional effects visible?

Some effects matter even when a defensible monetary value is unavailable. Record service quality, safety, accessibility, privacy, staff workload, resilience, environmental impact, customer trust, and reversibility where relevant. Describe the evidence and decision criterion instead of inventing a dollar value to force comparability.

Also identify who receives each benefit and who bears each cost. An option may have a positive total while concentrating transition burden on one team or excluding a user group. Put material distributional effects next to the financial results, not in a footnote.

For high-impact decisions, ask the relevant specialists to review privacy, accessibility, employment, procurement, security, tax, and legal implications. NIST’s Generative AI Profile highlights confabulation, data privacy, information integrity, and human oversight as risks that need lifecycle controls.［4］ Redaction, source binding, independent calculation, and named approval are practical controls here.

Draft the recommendation and approval record

Once the model and scenario table are complete, provide AI with a locked fact table and request a decision brief. Require it to separate evidence, interpretation, recommendation, and unresolved question. Every numeric statement must cite a row or scenario ID.

Use this final structure:

Section	Required content
Decision	Decision owner, deadline, and scope
Options	No-action baseline and viable alternatives
Results	Checked financial metrics and non-monetary effects
Assumptions	Material assumptions, confidence, and switching values
Risks	Risk costs, mitigations, residual uncertainty, and affected groups
Recommendation	Human judgment and why it remains robust or conditional
Approval	Owner, reviewer, version, date, and recheck trigger

Run fact-checking checks on external claims. The finance owner should verify formulas and price bases; data owners should verify inputs; specialists should review regulated effects; and the accountable decision maker should approve or reject the option. Record dissent and conditions rather than asking AI to produce artificial consensus.

Set recheck triggers such as a unit-cost change, delivery delay, pilot result, contract revision, demand shift, or assumption crossing its switching value. Link the approved recommendation, evidence register, calculation model, and assumption register by version. If one changes, create a new decision version instead of silently editing the old record.

Summary

Start with a precise decision and a no-action baseline.
Bind every figure to a source, owner, date, unit, and assumption.
Keep calculations deterministic and outside the language model.
Test downside, upside, optimism bias, and switching values.
Preserve non-monetary and distributional effects.
Require finance, specialist, and decision-owner approval.

FAQ

Can AI calculate the benefit-cost ratio for me?

Do not use a language model as the system of record for business-critical arithmetic. Calculate the ratio in a reviewed spreadsheet, script, or financial model, then let AI explain the verified result.

What is the no-action baseline?

It is the expected outcome if current arrangements continue. Include ongoing cost, deterioration, obligations, and foregone benefits so alternatives are compared with a realistic benchmark.

How many options should I compare?

Use enough materially different options to avoid a false yes-or-no choice. A practical set often includes no action, a minimum change, the proposed approach, and a different viable approach.

Should every benefit have a monetary value?

No. Monetize only when the method and evidence are defensible. Keep material non-monetary effects visible with their evidence, affected groups, and decision criteria.

How do I check an AI-generated assumption?

Trace it to a proposal row or source, assign an owner, define a validation method, and test a plausible range. If there is no supporting evidence, label it unverified rather than accepting a plausible sentence.

What is a switching value?

It is the value an assumption would need to reach for the option to fail the decision threshold or lose its preferred status. It tells reviewers which assumptions deserve the closest monitoring.

Can AI choose the best option?

No. AI can organize checked evidence and expose gaps, but accountable people must weigh objectives, financial results, non-monetary effects, risk, uncertainty, and distributional impact.

Can I paste confidential financial data into a public AI tool?

No. Follow your organization’s tool approval, confidentiality, access, retention, and data-minimization rules. Use de-identified examples or an authorized environment for restricted information.

Disclaimer: This article provides general information, not financial, accounting, procurement, legal, tax, or investment advice. Use your organization’s approved appraisal methods and qualified reviewers.

Sources:

HM Treasury, The Green Book 2026 — https://www.gov.uk/government/publications/the-green-book-appraisal-and-evaluation-in-central-government/the-green-book-2026
HM Treasury, Green Book supplementary guidance: optimism bias — https://www.gov.uk/government/publications/green-book-supplementary-guidance-optimism-bias
U.S. Government Accountability Office, Cost Estimating and Assessment Guide — https://www.gao.gov/assets/d20195G.pdf
NIST, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile — https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence

Sources checked 29 August 2026.

Related reading:

How to Create a Cost-Benefit Analysis with AI