Build an AI ROI One-Pager: Turn Vague Spend Into a Defensible Business Case

The task

You walk into a steering committee with a client who spent seven figures on AI last year and can't tell the board what they got for it. You need a one-page ROI narrative — usable in a deck, defensible to a CFO — for a specific deployment. This is the artifact you produce before the renewal conversation, the expansion pitch, or the "should we kill this" review.

Before AI

Today this means pulling license invoices from procurement, chasing usage data from the platform team, asking three business leads for anecdotes, and stitching a story in slides. You end up with either a vanity metric ("40,000 prompts run last quarter") or a hand-wavy claim ("meaningful productivity uplift"). Neither survives a CFO's second question. Half a day, minimum, and the output is usually thin.

The pressure is real: an MIT study found 95 percent of corporate generative AI pilots delivered no measurable return, pointing at habits over the technology. Boards read those headlines. So do your clients.

The workflow

The Forbes framing on measurement is the right frame here: that return could be measured in one of three ways: hours saved, errors prevented or dollars earned. Everything below forces the narrative onto one of those three axes.

Step 1 — Extract the raw facts and force them into a value hypothesis

Give the model whatever mess you have — invoices, usage exports, interview notes — and make it separate cost from benefit before it tries to be clever.

You are a consulting analyst preparing an AI ROI one-pager for a client steering committee. I will paste a mixed dump of information about a client's AI deployment (invoices, usage stats, interview quotes, adoption numbers).

Do NOT write the one-pager yet. First, produce a structured intake in this exact format:

1. DEPLOYMENT SUMMARY
   - Tool / vendor
   - Business function it serves
   - Population of users (headcount, role)
   - Time in production

2. TOTAL COST OF OWNERSHIP (annualised)
   - Licenses / subscription
   - Implementation & integration
   - Internal FTE time (training, admin, support)
   - Model / token / compute costs if separate
   - TOTAL (state assumptions inline)

3. CLAIMED BENEFITS (as reported, unverified)
   - List each claim, tag as: hours saved / errors prevented / dollars earned / vanity
   - Flag any claim that is a usage metric (prompts run, seats active) as VANITY

4. EVIDENCE GAPS
   - What you would need to verify each non-vanity claim
   - Rate confidence: High / Medium / Low

Be blunt. If a number looks self-reported and unverified, say so. If costs are missing categories, list what's missing. Do not invent figures — mark them "not provided".

Here is the dump:

CLIENT: Meridian Health Partners (regional payer, ~3,200 employees)
DEPLOYMENT: "ClaimsCopilot" — GenAI assistant for claims adjudicators, built on Azure OpenAI, rolled out Q2 last year.

INVOICE SUMMARY (last 12 months):
- Azure OpenAI consumption: $412,000
- Vendor implementation (Northlake Consulting): $680,000 one-time, $95,000/yr managed service
- Internal build team (2 FTE data eng, 1 FTE PM for 9 months): ~$540,000 loaded cost
- Compliance & security review: $70,000
- Training (all 340 adjudicators, 4 hrs each): estimated $180,000 loaded time

USAGE (from platform telemetry):
- 340 licensed adjudicators, 218 active monthly (64%)
- 47,000 assistant sessions last quarter
- Median 3.2 sessions per active user per day

INTERVIEW NOTES:
- VP Claims Ops (Sandra Ruiz): "Adjudicators tell me they save maybe 15 minutes a case on complex appeals. We do about 92,000 complex appeals a year."
- Team lead (Marcus Chen): "Honestly some people just use it to rewrite emails. But the appeals summarization is legit."
- QA lead (Priya Anand): "First-pass accuracy on coding decisions went from 88% to 93% since rollout. We haven't formally attributed it though — could be the new training curriculum too."
- CFO (Dan Whitaker): "I need to know if this thing pays for itself before renewal in November. Right now I see $1.9M out the door and a lot of enthusiasm."

BENCHMARKS:
- Loaded cost per adjudicator hour: ~$68
- Average cost per complex appeal (labor portion): ~$47
- Rework cost per miscoded claim: ~$310

Step 2 — Build the defensible value model

Now translate the verified pieces into a range, not a point estimate. CFOs trust ranges.

Using the structured intake you just produced, build a VALUE MODEL with three scenarios: Conservative, Base, Optimistic.

Rules:
- Every scenario must show: Gross Annual Benefit, Total Annual Cost, Net Value, ROI %, Payback (months).
- For each benefit line, show the formula in plain arithmetic (e.g., "92,000 appeals × 0.25 hrs × $68/hr × 60% attribution = $X"). No black boxes.
- Attribution haircut: Conservative applies 40% attribution to AI (rest goes to training, process changes, etc.), Base 60%, Optimistic 80%. State this explicitly.
- Exclude any VANITY claim from the model. Note it separately as "Not modeled — usage metric".
- If a claim was rated Low confidence in Step 1, exclude from Conservative, include at half-weight in Base.
- Include a "Sensitivity" note: which single assumption most changes the answer.

Output as a clean table plus a 3-sentence plain-English interpretation a CFO would accept.

Step 3 — Compress into a one-pager narrative

Now write the client-facing ONE-PAGER. Constraints:

- Fits on a single page (≤ 450 words).
- Structure: (1) Headline finding in one sentence with a number and a range. (2) What we spent. (3) What we got — in hours saved, errors prevented, or dollars earned, never in usage metrics. (4) What we're NOT counting and why. (5) The decision this supports (renew / expand / restructure / kill), with the condition that would flip it.
- Tone: consulting-neutral. No hype words ("transformative", "game-changing", "unlock"). No emojis. No em-dashes in the final copy.
- Include one explicit caveat about attribution.
- End with a "What we'd need to tighten this by the next review" list — max 4 bullets.

Do not restate the value model table; reference it as "see appendix". Write for a CFO who will read this in 90 seconds before a steering meeting.

Gotchas

• Garbage in, confident out. If the intake dump is thin, the model will happily invent tidy numbers. The Step 1 "Evidence Gaps" section is the guardrail — read it before trusting Step 2.
• Attribution is the whole ballgame. Almost every client will have a co-occurring change (new training, process redesign, team reshuffle) that inflates the AI story. The attribution haircut in Step 2 is deliberately blunt; adjust the percentages to what you can defend to this CFO, not a generic one.
• Vanity metrics sneak back in. Watch Step 3 for phrases like "adoption reached 64%" — that's a usage stat, not a return. AI spend does not become ROI simply because usage goes up. Strip them.
• Token costs are now a line item, not a rounding error. "Token costs are now a standard line item in our ROI calculations," said Sancho. Those costs are treated as part of the company's cost of goods alongside labor and infrastructure. If your intake dump doesn't have them broken out, flag it — the total cost figure is wrong without them.
• Don't skip the "condition that would flip it" line. A one-pager that only says "renew" reads like advocacy. Naming what would change your mind is what makes it consulting work, not marketing.

Time saved

A defensible ROI one-pager built the manual way — chasing data, reconciling numbers, drafting, redrafting after the partner review — is a 4-6 hour job for a senior consultant, and often longer if the client's data is scattered. This workflow gets you to a first draft in about 30 minutes of model time plus your review, and the structured intake step catches the evidence gaps you'd otherwise discover the night before the meeting. Estimate, not measurement: the quality of the input dump is still the ceiling.

What we got when we ran it

We executed every prompt above, in order, against the sample input (model: claude-sonnet-4-6). Outputs below are real and unedited apart from length truncation.

Step 1 output

# STRUCTURED INTAKE: ClaimsCopilot — Meridian Health Partners --- ## 1. DEPLOYMENT SUMMARY | Field | Detail | |---|---| | Tool / Vendor | "ClaimsCopilot" — GenAI assistant built on Azure OpenAI; implementation by Northlake Consulting | | Business function | Claims adjudication support — primarily complex appeals summarization; secondary use (unplanned) for email drafting | | Population of users | 340 licensed adjudicators; 218 monthly active (64% adoption). Note: Marcus Chen's comment suggests a meaningful but unquantified subset is using the tool off-label for email rewriting, not core adjudication. | | Time in production | ~12 months (rolled out Q2 last year; evaluating for November renewal) | --- ## 2. TOTAL COST OF OWNERSHIP (ANNUALISED) All figures annualised. Where costs are one-time, they are amortised over an assumed 3-year useful life unless noted otherwise. The CFO's "$1.9M out the door" likely refers to Year 1 actual cash outflow — that figure is reconciled at the bottom. | Cost Category | Annual Figure | Notes / Assumptions | |---|---|---| | Azure OpenAI consumption | $412,000 | Actual invoice. Annualised from last 12 months. Assumed stable; usage growth not modelled. | | Vendor managed service | $95,000 | Recurring annual contract with Northlake Consulting. | | Vendor implementation (amortised) | $226,667 | $680,000 one-time ÷ 3 years. If client expenses this in Year 1 only, Year 1 cash impact is $680K. | | Internal build team (FTE) | $240,000 | $540,000 over 9 months = $720K/yr run rate, but this was

…output truncated at 250 words.

Step 2 output

# ClaimsCopilot Value Model — Meridian Health Partners --- ## Governing Rules Applied (State Once, Not Repeated Per Scenario) | Parameter | Conservative | Base | Optimistic | |---|---|---|---| | AI attribution haircut | 40% | 60% | 80% | | Low-confidence claims included? | No | Half-weight | Full-weight | | Vanity metrics modeled? | No | No | No | | Cost basis | Annualised steady-state ~$1,057K | Same | Same | Claims excluded from all scenarios (VANITY): - 47,000 sessions/quarter — Not modeled — usage metric - 218 MAU / 64% seat utilization — Not modeled — adoption metric - 3.2 sessions/user/day — Not modeled — engagement metric --- ## Benefit Line Definitions (Formulas Shown Once) ### Benefit A — Time Savings on Complex Appeals (Confidence: LOW) Source: Sandra Ruiz self-report; unverified by workflow timestamps. Scope adjustment: Only 218 of 340 adjudicators are active monthly. Appeals worked with tool assistance are not directly measured; we proxy by assuming active users handle a proportional share of annual appeal volume. Adjusted appeal volume exposed to tool: 218 active users ÷ 340 licensed users = 64.1% of workforce active → applied to 92,000 appeals = 59,000 appeals plausibly tool-assisted (Conservative excludes this line entirely per Low-confidence rule. Base and Optimistic include at half-weight and full-weight respectively.) | Scenario | Formula | Value | |---|---|---| | Conservative | Excluded — Low confidence | $0 | | Base | 59,000 appeals × 0.25 hrs × $68/hr × 50% weight × 60% attribution

…output truncated at 250 words.

Step 3 output

# ClaimsCopilot: Return on Investment Assessment Meridian Health Partners — Steering Committee Review Prepared by: [Analyst Team] | Review Date: [Month Year] --- FINDING At current verified performance levels, ClaimsCopilot costs $1.06M annually against a benefit range of $0 to $1.53M per year, depending on how much of the claimed productivity and quality gains are attributable to the tool versus concurrent process changes. The deployment has not yet demonstrated a confirmed positive return, but a plausible positive case exists if two factual questions can be answered before November renewal. --- WHAT WE SPENT Total Year 1 cash outlay was $1.98M, including implementation. On a steady-state annualised basis, the ongoing cost of the deployment is approximately $1.06M per year, covering compute, vendor managed services, amortised build costs, and internal support time. --- WHAT WE GOT Two substantive benefit claims are in scope. First, adjudicators report saving roughly 15 minutes per complex appeal. Applied to the 59,000 appeals plausibly handled with tool assistance, and assuming 60% of that gain is attributable to the AI rather than process changes, that represents approximately $302,000 in recovered adjudicator capacity annually. Second, first-pass coding accuracy has improved five percentage points since rollout. At the same 60% attribution, that implies roughly $274,000 in annual rework cost avoidance. Combined Base case benefit: $576,000 per year. Full scenario range is detailed in the appendix. Attribution caveat: Neither benefit has been isolated to ClaimsCopilot specifically. A new training curriculum launched near the same time, and time savings are drawn from

…output truncated at 250 words.