Need Sign-Off for AI Spend? How to Build an AI Business Case That Passes Approval

How to Build an AI Business Case

An annotated one-page business-case template walkthrough sits directly beneath this claim, and its first on-screen graphic names “AI business case”.

The technology quote rarely sets the biggest line in the paper. Info-Tech’s business-case research reaches the same finding from its own analyst blueprint: the biggest cost is probably not the AI itself.

Why Technology Cost Comes Second

The reason the priced-process-first rule holds is financial discipline, not house style. MIT Sloan Management Review’s 2026 research finds most AI guidance measures inputs, naming what to invest in, rather than outputs, and that few organisations apply the same financial discipline to AI as to a new factory. A factory carries a costed baseline before it is commissioned; an AI business case needs the same.

The most common root cause of AI project failure sits upstream of the tool. RAND’s 2024 research cites an external estimate exceeding 80% for AI project failure, around twice the rate of non-AI IT projects, and from its own interviews names lacking the necessary data as the second of five root causes; its study focused on AI/ML development and excluded projects using only pretrained LLMs. A business case built on real process records catches this before spend, not after it.

Unbaselined cases are harder to defend at later review. MIT NANDA’s 2025 research, a preliminary report, finds that about 95% of organisations obtain no measurable profit-and-loss impact from generative AI, a measurement finding as much as an outcome one; that missing baselines cause cases to fail review is this page’s inference, not a finding of the report. Without a pre-priced process there is nothing to measure the return against.

The paper itself carries five sections, and the technology quote is one line inside the third section, the change cost, never the anchor of the case. The case reads bottom-up: process cost today, then change cost, then benefit, then risk, so the technology price never leads the argument.

An AI readiness assessment measures the baseline your business case stands on, supplying the process, data and cost records this section requires before the paper is drafted.

Data You Need Before Writing a Word

The baseline needs four numbers from real records: process volume, cycle time per work item, error and rework rate, and staff time per item, each with its source named, not estimated.

Volume comes from process logs, cycle time from timing samples, error rate from quality and error registers, staff time from time-recording data. Estimates that cannot be traced to a record are the first thing finance challenges. RAND’s root-cause analysis shows AI failures trace to problem misalignment and missing data, both detectable only from real inventories, never from a guess.

Access and Permissions to Secure First

Three access items unlock the baseline: time-recording data, finance’s loaded cost rates, and the process records themselves. Each has an owner who signs off release. Securing the owner sign-off list before pricing starts prevents the baseline stalling halfway for a permission nobody requested.

The request list runs in order: request time-recording data, request loaded cost rates from finance, request process records, then confirm the owner sign-off list.

Skills at the Table

Three roles price the baseline: the process owner who knows the real workflow, the finance partner who supplies loaded cost rates, and a technical estimator who scopes the change cost. An AI readiness assessment supplies the missing baseline data when those records do not yet exist.

Pricing the Process Baseline

Step 1: Selects the Single Process in Scope

Scope creep across processes sinks the case. The paper prices one process, so a multi-process ask splits into separate cases.

Step 2: Counts the Volume Over a Defensible Window

A seasonal window distorts the count. Counting over a full cycle keeps the volume from being challenged later.

Step 3: Times the Work Item End to End

Self-reported estimates inflate the time. Sampling real items end to end, including queue time, keeps the figure honest.

Step 4: Prices the Time at Loaded Cost Rates

Base salary understates the true cost. Loaded cost rates supplied by finance set the real figure, not a payroll line.

Step 5: Records Error and Rework Cost

Rework hides in other teams’ queues. Tracing it prices the error pool instead of assuming it away.

The priced baseline is non-negotiable for the same reason MIT Sloan Management Review names: applying factory-grade financial discipline to AI means a costed baseline sits before the asset, not after it. Data quality is the failure surface RAND names as its second root cause, and a baseline built on real logs and registers is the mitigation, captured before any build starts.

The cost of skipping the baseline shows up later, in the abandonment curve. S&P Global Market Intelligence’s 2025 research a survey of 1,006 North American and European professionals, found the share of organisations abandoning the majority of their AI initiatives jumped from 17% to 42% year on year, and the average organisation scrapped 46% of its proof-of-concepts. That a priced baseline is what a survivor points to at review is this page’s inference, not a causal finding of that survey.

The five steps above produce one number: cost per work item. That figure becomes the denominator every benefit claim in the paper is read against.

Structuring the Approval Paper

Step 6: States the Decision Requested, Staged Not Total

An all-or-nothing ask tends to fail the board. Staging the commitment funds a bounded first phase with a defined go or no-go point.

Step 7: Presents the Costed Baseline as the Anchor

Burying the baseline invites challenge. Leading the paper with the cost-per-work-item keeps every later number read against it.

Step 8: Prices the Change, Build and Run, Honestly

An omitted run cost surfaces at post-approval review. Listing build cost and ongoing run cost as separate lines gives operations the run figure it owns.

Step 9: Claims Three Benefit Categories With Owners

Three benefit categories carry the case: time released, error reduced, and capacity created, each with a named owner sitting against it. Every numeric benefit example a reader plugs into this template stays specific to their own measured baseline; no vendor multiple, no headline return appears anywhere in this paper. How those benefit categories get measured after approval, in four value pools, is a separate question this page hands over rather than answers.

Step 10: Registers Risks With Treatments, Not Disclaimers

Governance risks written as disclaimers get escalated. A treatment per risk answers the objection before a reviewer raises it, and the client’s regulatory set gives every treatment its reference anchor, to be backed by implementation evidence rather than the mapping alone.

Why the ask is staged at all traces to the same expectation gap the board already carries. IBM’s 2025 CEO study reports that only around 25% of AI initiatives deliver the expected ROI, and just 16% have scaled enterprise-wide. A staged decision limits the board’s exposure to the majority that under-deliver against an over-set expectation. Unpriced run cost is fatal for the same reason the abandonment surge names: cases die after approval when the run cost was trimmed to make the numbers work.

The risk register hosts governance risk, not just delivery risk, and the treatments map to named standards. ISO/IEC 42001 places AI risk treatment inside a management system, with Clause 6 covering planning and risk treatment and Annex A setting the controls. The NIST AI Risk Management Framework, a voluntary framework, names four functions, GOVERN, MAP, MEASURE and MANAGE, that a risk register can sort its treatments against. Singapore’s AI Verify testing framework assesses documented implementation against eleven governance principles, giving each treatment an evidence requirement rather than a promise. Australia’s Guidance for AI Adoption, published 21 October 2025, sets six voluntary essential practices (evolving the earlier Voluntary AI Safety Standard’s ten guardrails), and names which one a treatment answers. Naming the standard a treatment maps to structures the treatment and its evidence requirement; the mapping is not itself proof that a control operates effectively, and compliance with applicable Singaporean or Australian law needs implementation evidence and a separate legal assessment.

The three benefit categories above are the paper’s fixed structure, not a claimed return, and the count stays the same on every case. The paper is complete when each of the five readers’ sections answers their one question without needing a follow-up meeting.

Presenting for Sign-Off

The check maps one-to-one to the five sections built above. A case that fails a row here fails on the construction step that produced it, not on presentation. With only around 25% of initiatives delivering the expected ROI, a board’s default position is scepticism, and a paper that pre-answers each objection converts that scepticism into a decision.

Once sign-off lands, the AI strategy roadmap sequences the funded build once the board signs, ordering the approved work in payback order.

Where Business Cases Go Wrong

The three rejection modes each trace to a specific step: benefits claimed without a baseline trace back to steps 2 through 5 being skipped, tool-first framing that leads on the technology price traces back to step 7 being inverted, and unpriced run cost that surfaces at review traces back to step 8 being trimmed.

The abandonment data explains why these three dominate. The 42% of organisations abandoning the majority of their initiatives, and the 46% of proof-of-concepts scrapped, are the market backdrop; that these reflect cases which cleared a weak internal bar and then reverted once the run cost or missing baseline caught up is this page’s inference, not something the S&P survey attributes. The fix path names the step to redo, not the whole paper.

From Paper to Measured Baseline

Because the whole paper stands on the baseline, an AI readiness assessment measures the process, data and cost baseline your case needs before you write section one. The baseline is the single point every reader tests, and the assessment measures that baseline directly, so the case gets built on measured records rather than estimates. Enquiries route through the enquiry form; Synthesised.ai does not publish a rate card.

The AI resources hub catalogues every other resource this page connects to, and citable AI adoption and failure statistics supply the wider market context behind the baseline argument above.

An estimated baseline is a claim. A measured baseline is a case.