Build a One-Page AI ROI Case: Translate Your AI Spend Into Business Outcomes With Prompts

The task

Founders and operators need to defend their AI spend on a single page — for a board deck, an investor update, or a leadership offsite. The story has to tie line-item costs (seats, API tokens, vendor contracts) to concrete business outcomes (hours saved, cycle time, revenue-adjacent lift). The pressure is mainstream now that the accountability question has hit the front page, and "we're being innovative" is no longer an answer.

Before AI

Most operators do this in a rushed Sunday-night sprint: pull the vendor invoices, guess at utilization from Slack anecdotes, cobble together a slide, hedge every number. Two to four hours, and the output is usually squishy — a list of tools and a vibes-based benefit column. Finance pokes one hole and the whole thing wobbles.

The workflow

The idea: feed your raw spend + usage + outcome data into a prompt chain that produces a defensible one-pager. You'll want a lightweight framework in mind before you start — Larridin's AI ROI measurement framework and this CFO-ready enterprise breakdown are both useful priors on what "counts" as ROI.

1. Dump your raw inputs

Gather three things into one text block: (a) monthly AI costs by line item, (b) usage or adoption signals, (c) any outcome data you already have — even rough. Use the sample below as the shape.

COMPANY: Northwind Labs (Series B, 82 employees, B2B SaaS)
PERIOD: Q2 2026 (Apr-Jun), quarterly figures unless noted

--- AI COSTS (quarterly) ---
- ChatGPT Enterprise: 60 seats × $60/mo × 3mo = $10,800
- GitHub Copilot: 22 eng seats × $19/mo × 3mo = $1,254
- Anthropic API (support bot): $4,200 in tokens
- Gong AI add-on: $6,000
- Perplexity Pro: 15 seats × $20/mo × 3mo = $900
- Internal RAG infra (AWS): ~$3,800
TOTAL Q2 AI SPEND: $26,954

--- USAGE / ADOPTION ---
- ChatGPT Enterprise: 71% weekly active (avg 34 msgs/user/week)
- Copilot: 20 of 22 eng active weekly; 31% suggestion acceptance rate
- Support bot: handled 2,140 tier-1 tickets (of 3,900 total) autonomously
- Gong AI summaries: used on 88% of sales calls
- Perplexity: 9 of 15 seats active weekly (marketing + BD)

--- OUTCOME SIGNALS ---
- Support: median first-response time dropped from 4h12m (Q1) to 41m (Q2)
- Support headcount: held flat at 6 despite 28% ticket volume growth
- Engineering: shipped 47 PRs/wk (Q2 avg) vs 39/wk (Q1 avg), same headcount
- Sales: rep ramp time to first closed deal fell from 71 days to 58 days
- Marketing: published 22 blog posts in Q2 vs 12 in Q1, same team of 2
- One incident: AI-generated pricing email sent to 340 prospects with wrong tier; ~6h to remediate

--- FULLY-LOADED COMP (for savings math) ---
- Support rep: $85K + 25% loaded = $106K/yr → ~$51/hr
- Engineer: $170K + 25% = $213K/yr → ~$102/hr
- Marketer: $120K + 25% = $150K/yr → ~$72/hr

2. Turn spend + usage + outcomes into a structured ROI table

This prompt does the heavy lifting: it maps every dollar to an outcome category, computes a defensible savings estimate, and flags what's speculative vs. measured.

You are a CFO-minded operator building a one-page AI ROI case for a Series B founder to present to the board. Below is the raw quarterly data.

Do the following:

1. Group the AI spend into 3-5 outcome categories (e.g., "Customer Support Deflection", "Engineering Velocity", "Sales Enablement", "Content Throughput"). Assign every cost line to exactly one category.

2. For each category, produce a table row with:
   - Category name
   - Quarterly cost ($)
   - Primary outcome metric (with before → after numbers from the data)
   - Estimated quarterly value created ($) — show the arithmetic in one line
   - Confidence: High / Medium / Low, with a one-clause reason
   - Net (value − cost)

3. Rules for the value estimate:
   - Only use fully-loaded comp figures provided.
   - For time saved, convert to $ using the hourly rate.
   - Do NOT count revenue you can't attribute; use "revenue-adjacent" framing instead.
   - If a category has no measurable outcome yet, mark value as "Not yet measurable" and confidence Low — do not fabricate a number.

4. Below the table, show:
   - Total quarterly AI spend
   - Total estimated quarterly value (sum of measurable rows only)
   - Payback multiple (value ÷ spend), rounded to 1 decimal
   - One sentence on what's excluded from the multiple and why

Be conservative. A skeptical CFO will read this. Return the table in clean markdown.

3. Draft the one-page narrative around the table

Numbers alone don't land. This prompt wraps the table in the story a board actually wants: what's working, what's not, what we're changing next quarter.

Using the ROI table you just produced, draft the surrounding one-page narrative. Structure it exactly as follows, in markdown:

**Headline** (one line, 12 words max): the ROI multiple + the single biggest driver.

**What we spent, what we got** (3 sentences): total spend, total measurable value, payback multiple. Name the top-performing category and the weakest one.

**What's working** (3 bullets, one sentence each): the concrete before→after metric that moved, tied to the AI tool that drove it. No adjectives like "significant" or "transformative" — use the actual numbers.

**What's not working yet** (2 bullets): categories with Low confidence or no measurable outcome. Say plainly why the signal isn't there yet.

**Risk log** (2 bullets): include the pricing-email incident and one forward-looking risk (e.g., vendor concentration, model deprecation, adoption plateau).

**Q3 decisions** (3 bullets, each starts with a verb): what we'll cut, what we'll double down on, what we'll instrument better. Each bullet must reference a specific line item or metric from the table.

Tone: plain, numeric, no hype. Assume the reader is a numbers-first board member who has seen ten of these decks this month. Keep the whole page under 350 words including the table.

4. Pressure-test it against a skeptical CFO

Before you send the page, run it through an adversarial pass. This surfaces the exact questions finance will ask so you can either fix the number or prepare the answer.

Now switch roles. You are a skeptical CFO reviewing the one-pager above for the first time. Produce:

1. The five sharpest questions you'd ask in the meeting, ranked by how damaging an unprepared answer would be. For each, quote the specific sentence or number from the one-pager that prompted the question.

2. For each question, a one-paragraph suggested answer the founder could give — grounded ONLY in the data provided earlier. If the data doesn't support a strong answer, say so and recommend what the founder should measure next quarter to close that gap.

3. A final "kill-shot check": is there any number in the one-pager that, if a board member ran it themselves on a napkin, would come out materially different? If yes, flag it and propose a more defensible framing.

Gotchas

• Garbage in, garbage out. If your usage data is anecdotal ("the team loves it"), the table will invent precision that isn't there. Pull actual seat-level activity from vendor admin panels before you start.
• Beware the loaded-comp trap. Time saved × hourly rate is not the same as dollars saved unless you actually reduced headcount or reallocated the hours to revenue work. The prompt uses "revenue-adjacent" framing for a reason — don't let a slide deck turn it back into "savings."
• Attribution is thin for sales and marketing. Ramp-time and content-throughput lifts have many causes. Confidence: Medium at best, and say so.
• The model will hedge when the data is sparse. That's a feature. If step 2 returns "Not yet measurable" for a category, resist the urge to prompt it into inventing a number. That's the exact behavior the NYT story is calling out.
• One-shot fragility. If your input block is very long or messy, the model may drop line items when it builds the table. Spot-check that every cost line from your input appears in exactly one category row.

Time saved

Rough estimate: a founder or CoS who currently spends 3–4 hours assembling this from spreadsheets, Slack scrolling, and gut feel can get to a defensible draft in about 20 minutes of prompting plus 15 minutes of edits. Call it ~3 hours per board update, quarterly. The bigger win isn't the time — it's that the number you put in front of the board is one you can actually defend when someone pushes back.

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

# Northwind Labs — Q2 2026 AI ROI Case ### For Board Presentation | Period: Apr–Jun 2026 | Prepared: CFO Review Draft --- ## Category-to-Cost Mapping | Cost Line | Assigned Category | |---|---| | Anthropic API (support bot) — $4,200 | Customer Support Deflection | | Internal RAG infra (AWS) — $3,800 | Customer Support Deflection | | GitHub Copilot — $1,254 | Engineering Velocity | | ChatGPT Enterprise — $10,800 | Engineering Velocity (primary) + shared; see note | | Gong AI add-on — $6,000 | Sales Enablement | | Perplexity Pro — $900 | Content Throughput | > Allocation note: ChatGPT Enterprise is used across eng, marketing, and BD. In the absence of seat-level attribution data, the full $10,800 is assigned to Engineering Velocity as the largest identifiable user group (22 of the 60 seats are engineers; the residual seats lack measurable outcome data this quarter). This is conservative — it inflates eng cost and leaves other categories unburdened. --- ## ROI Summary Table | Category | Q2 Cost | Primary Outcome Metric | Estimated Q2 Value | Arithmetic | Confidence | Net (Value − Cost) | |---|---|---|---|---|---|---| | Customer Support Deflection | $8,000 | Tier-1 tickets handled autonomously: 0 → 2,140 (of 3,900); headcount flat at 6 despite +28% ticket volume | $84,354 | 2,140 tickets deflected × avg 45 min/ticket × $51/hr = $81,810; first-response SLA improvement valued at 0 (not directly monetizable) — using ticket-labor savings only; round-trip remediation cost deducted: 6 hrs ×

…output truncated at 250 words.

Step 2 output

# AI ROI — Q2 2026 Board Summary ### Northwind Labs | One-Page Briefing --- Headline $26,954 in AI spend returned 7.2× in measurable value, led by support ticket deflection. --- What we spent, what we got We spent $26,954 on AI tooling in Q2 and measured $193,662 in quantifiable value across three categories, yielding a 7.2× payback multiple. Customer Support Deflection was the strongest category — $8,000 in cost against $84,354 in labor value — because deflection volume is system-logged and the arithmetic is direct. Content Throughput was the weakest: $900 spent, zero measurable return this quarter, because blog-to-pipeline attribution cannot close in a single quarter. --- What's working - The Anthropic support bot autonomously resolved 2,140 of 3,900 tier-1 tickets, holding headcount flat at 6 reps despite 28% volume growth. - GitHub Copilot contributed to a rise in shipped PRs from 39/week to 47/week at identical engineering headcount, with a 31% suggestion acceptance rate across 20 of 22 active seats. - Gong AI was used on 88% of sales calls; rep ramp time to first closed deal fell from 71 days to 58 days. --- What's not working yet - Perplexity Pro ($900) doubled blog output from 12 to 22 posts, but SEO compounding and pipeline influence won't be measurable until Q3 at the earliest. - ChatGPT Enterprise has 60 seats but no outcome data outside engineering; 38 seats across marketing and BD have no attributed metric this quarter. --- Risk log - In Q2, an AI-generated pricing

…output truncated at 250 words.

Step 3 output

# CFO Challenge Brief — Northwind Labs AI One-Pager ### Skeptical Review | Q2 2026 --- ## Five Questions, Ranked by Damage Potential --- ### #1 — Most Damaging "You're claiming 7.2× ROI but the biggest value line is 2,140 tickets at 45 minutes each. Where does the 45 minutes come from? That's $84,354 riding on a benchmark you didn't measure." Prompted by: "2,140 tickets deflected × avg 45 min/ticket × $51/hr = $81,810" Suggested Answer: The 45-minute figure is a standard B2B SaaS tier-1 handle-time benchmark and is disclosed as such in the table footnote. That's a genuine gap. The stronger defense available in the data is the headcount argument: the team handled 28% more tickets with zero additional hires. At $106K fully-loaded per rep, one avoided hire saves $26,500 per quarter — that alone returns 3.3× on the $8,000 support AI spend without touching the handle-time estimate at all. In Q3, we will pull actual average handle time from the ticketing system (Zendesk or equivalent) and replace the benchmark figure. Until then, we should present the avoided-hire number as the primary value claim and treat the handle-time savings as secondary. --- ### #2 "You attributed 62% of the sales ramp compression to Gong AI. That number appears nowhere in the source data. How did you get to 62%?" Prompted by: "use $26,364 (62% attribution, split between productivity and revenue-adjacent pipeline)" Suggested Answer: The 62% figure is not defensible as stated. It was derived by working backward from a

…output truncated at 250 words.