Draft a Board-Ready AI Spend vs. Return Summary from Raw Capex and Revenue Figures

The task

You're the FP&A lead or controller. The audit committee wants a one-page read on what the company has spent on AI year-to-date, what it's returned, and how that ratio stacks up against the widening capex-to-revenue gap the market is now repricing. This workflow turns your raw GL export and revenue attribution figures into a board-ready memo in one sitting.

Before AI

You'd pull the AI cost centers from the GL, cross-walk them against the CFO's approved AI initiative list, chase product managers for revenue attribution, then spend an evening in Google Docs writing the narrative. Realistic time: half a day, plus a second pass after the CFO edits the framing.

The workflow

1. Paste your figures and get a structured spend/return table.

The first prompt takes your raw numbers and forces them into a clean table the board can read. Feed it exactly what you'd hand a junior analyst — line items, dollar amounts, brief context.

You are an FP&A analyst preparing figures for a board audit committee. Below is raw AI-related capex, opex, and attributed revenue data for the current fiscal year to date.

Do the following:
1. Group the line items into three buckets: (a) Infrastructure & compute, (b) Licensing & vendor tools, (c) Internal build (headcount, integration, training).
2. Produce a markdown table with columns: Bucket | YTD Spend ($) | Attributed Revenue / Savings ($) | Net ($) | Return Multiple (Revenue÷Spend, 2 decimals).
3. Add a totals row.
4. Below the table, list any line items where attribution is missing or looks unreliable — flag them explicitly. Do not fabricate attribution figures. If a number is missing, write "not provided."

Return only the table and the flag list. No preamble.

Company: Meridian Logistics Co. (fabricated example)
Fiscal year: FY26, YTD through Q2

AI-related line items pulled from GL:
- GPU cloud compute (AWS Bedrock + Azure OpenAI): $2,140,000 spend
- On-prem H100 cluster depreciation (YTD portion): $860,000 spend
- Copilot for Microsoft 365 licenses (1,200 seats): $468,000 spend
- Anthropic Claude API (customer support automation): $215,000 spend
- Snowflake Cortex + vector DB add-ons: $310,000 spend
- Data science headcount fully loaded (6 FTE, allocated): $1,320,000 spend
- Integration contractor (Deloitte, AI ops pilot): $540,000 spend
- Change management & training program: $180,000 spend

Attributed revenue / hard savings from Finance-approved attribution memos:
- Customer support automation (Claude): $1,410,000 in avoided contact-center opex
- Sales lead-scoring model (in-house): $780,000 attributed net-new ARR (recognized YTD portion: $290,000)
- Copilot productivity gains: no formal attribution completed yet
- Freight-pricing model (in-house): $2,050,000 gross margin uplift, per Ops
- Warehouse forecasting pilot: pilot only, no revenue impact quantified

CFO wants this framed against the industry pattern of AI capex outrunning AI revenue.

2. Benchmark the company's ratio against the industry pattern.

Now compare the internal ratio to the public narrative — the "spending massively outpacing revenue" story the board members are reading in the FT and Bloomberg on their commute.

Using the table you just produced, do the following:

1. Compute the company's overall Return Multiple (total attributed revenue+savings ÷ total spend) to 2 decimals.
2. Write a 3-sentence benchmark paragraph comparing this figure to the widely reported industry pattern where hyperscaler and enterprise AI capex is running ahead of attributable AI revenue (i.e., return multiples below 1.0x at the aggregate level). Be precise: do not cite specific external dollar figures unless they were in the input. Frame our multiple as "above / below / roughly in line with" the pattern.
3. Add one sentence on what the flagged/missing-attribution items would need to contribute for the picture to change materially.

Return only these three outputs, labeled 1, 2, 3.

3. Turn it into the board memo.

Final step: wrap the analysis in the format the audit committee actually reads.

Now draft a one-page board audit committee memo using everything above. Structure:

- TO / FROM / RE / DATE header (use "Audit Committee" / "Office of the CFO" / "AI Investment Return Review — YTD" / today's date)
- Section 1: Bottom line (2 sentences: total spend, total attributed return, overall multiple)
- Section 2: Where the money went (reference the three buckets, one sentence each)
- Section 3: Where the return came from (name the top two contributing initiatives; note what is not yet attributed)
- Section 4: Industry context (one short paragraph, using the benchmark language from the previous step)
- Section 5: Risks & open items (bullet list — include the unattributed line items, and note that attribution methodology has not been audited)
- Section 6: Recommendation (2-3 sentences — should the committee ask for a deeper attribution audit, a spend pause, or continued investment? Base this on the actual numbers, not a template answer.)

Tone: neutral, controller-voice, no hype. No em-dashes. Keep the whole memo under 500 words.

Gotchas

• Attribution is the soft spot. The model will happily compute a return multiple from whatever revenue figures you paste. If Ops is over-attributing margin uplift to the AI model versus, say, a rate increase that shipped the same quarter, the memo will inherit that error. Have the attribution memos reviewed before this workflow, not after.
• Depreciation timing. If your on-prem GPU cluster is being depreciated on a schedule that doesn't match when the revenue landed, the ratio flatters or punishes the current period. Note the depreciation policy in the input or the memo will read too clean.
• The "industry benchmark" step is qualitative. The prompt deliberately avoids citing specific external dollar figures the model might hallucinate. If a board member asks for the source, point them at the actual Forbes analysis of the capex-to-revenue gap rather than the memo language.
• Don't skip step 2. If you jump straight from the table to the memo, the "industry context" section tends to come out generic. The intermediate benchmarking step is what makes the framing credible.

Time saved

A controller doing this by hand — pulling GL lines, building the table in Excel, drafting the narrative, and revising for tone — typically loses 3-4 hours per board cycle. This workflow gets you to a defensible first draft in about 45 minutes, most of which is checking the attribution inputs rather than writing. Conservative estimate: ~2 hours saved per memo, more if your board packet cycle is monthly rather than quarterly.

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

Meridian Logistics Co. — AI Investment ROI Summary | FY26 YTD through Q2 | Bucket | YTD Spend ($) | Attributed Revenue / Savings ($) | Net ($) | Return Multiple | |---|---|---|---|---| | **(a) Infrastructure & Compute** | | | | | | GPU cloud compute (AWS Bedrock + Azure OpenAI) | 2,140,000 | — | — | — | | On-prem H100 cluster depreciation (YTD) | 860,000 | — | — | — | | **Bucket Subtotal** | **3,000,000** | **2,340,000** ¹ | **(660,000)** | **0.78x** | | **(b) Licensing & Vendor Tools** | | | | | | Copilot for M365 (1,200 seats) | 468,000 | — ² | — | — | | Anthropic Claude API | 215,000 | 1,410,000 | 1,195,000 | — | | Snowflake Cortex + vector DB add-ons | 310,000 | — | — | — | | **Bucket Subtotal** | **993,000** | **1,410,000** ³ | **417,000** | **1.42x** | | **(c) Internal Build (Headcount, Integration, Training)** | | | | | | Data science headcount (6 FTE, fully loaded) | 1,320,000 | — | — | — | | Integration contractor (Deloitte, AI ops pilot) | 540,000 | — | — | — | | Change management & training | 180,000 | — | — | — | | **Bucket Subtotal** | **2,040,000** | **2,340,000** ⁴ | **300,000** | **1.15x** | | **TOTAL** | **6,033,000** | **6,090,000** | **57,000** | **1.01x** | > **Attribution assignment rationale (for audit committee transparency):**

…output truncated at 250 words.

Step 2 output

1. Overall Return Multiple

$6,090,000 ÷ $6,033,000 = 1.01x

---

2. Benchmark Paragraph

Meridian's FY26 YTD return multiple of 1.01x places it roughly in line with — and marginally above — the widely observed industry pattern in which enterprise AI investment is outrunning formally attributable revenue and savings, with aggregate return multiples broadly reported below 1.0x during early deployment cycles. The company is not immune to the same structural dynamic: infrastructure and compute (bucket a) returns only 0.78x in isolation, which mirrors the pattern seen across organizations where shared foundational spend precedes monetization and attribution lags capital commitment. That Meridian reaches 1.01x at all is largely a function of two high-conviction use cases — customer support automation and the freight-pricing model — pulling the blended figure above breakeven, rather than broad portfolio-wide monetization.

---

3. Sensitivity on Flagged Items

If the freight-pricing attribution ($2,050,000, currently lacking a Finance-approved memo) is removed or materially revised downward, and the Copilot productivity gains remain unquantified, Meridian's return multiple falls to approximately 0.67x — well inside the sub-1.0x industry pattern — meaning those two items alone need to be formally validated and must together contribute a minimum of roughly $1,993,000 in approved attribution just to return the portfolio to breakeven.

Step 3 output

--- TO: Audit Committee FROM: Office of the CFO RE: AI Investment Return Review -- YTD DATE: July 14, 2025 --- Section 1: Bottom Line Through Q2 FY26, Meridian Logistics Co. has deployed $6,033,000 in AI-related spend across capital and operating categories, against $6,090,000 in Finance-memo-supported attributed revenue and hard savings. The blended return multiple stands at 1.01x, which is marginally above breakeven but contingent on attribution quality addressed in Section 5 below. --- Section 2: Where the Money Went Infrastructure and compute (GPU cloud and on-prem depreciation) represents the largest bucket at $3,000,000, consistent with the front-loaded capital pattern typical of early AI deployment cycles. Licensing and vendor tools totaled $993,000, covering third-party APIs and productivity software across multiple business functions. Internal build costs -- data science headcount, integration contractors, and change management -- accounted for $2,040,000, reflecting the organizational investment required to operationalize models in production. --- Section 3: Where the Return Came From Two initiatives account for all attributable return: the Claude API customer support automation ($1,410,000 in avoided contact-center opex, Finance-memo approved) and the freight-pricing model ($2,050,000 in gross margin uplift, sourced from Operations). Copilot for M365 productivity gains, Snowflake Cortex enabling costs, and the warehouse forecasting pilot carry no approved attribution to date and contribute $0 to the current return figure. --- Section 4: Industry Context Meridian's 1.01x multiple places it roughly in line with, and marginally above, the widely observed industry pattern in which enterprise AI investment is outrunning formally attributable revenue and savings,

…output truncated at 250 words.

---

This content is for informational purposes only and is not financial, investment, or accounting advice. Verify outputs against authoritative sources before use.