Build an AI ROI Evidence Dossier: Turn Vague Spend into Defensible Numbers

The task

HR leaders now get pulled into the quarterly AI-spend review alongside IT and Finance. You're asked what the recruiting copilot, the policy-drafting assistant, and the interview-scoring tool actually returned — in dollars, not vibes. This workflow turns your messy pile of pilot notes, time-study jots, and headcount figures into a one-page evidence dossier your CFO can defend to the board.

The pressure is real: the Forbes read on the shift is that tracking usage isn't enough anymore — you need to track what it earns.

Before AI

Today this usually means a scramble the week before the review: pinging team leads for anecdotes, guessing at hours saved, pulling license invoices from Procurement, and hand-building a slide in PowerPoint. Between the chasing, the reconciling, and the rewriting after Finance pushes back on soft numbers, most HR ops leaders lose a full day — often two. And half of what lands in the deck is still anecdotal, which is exactly the problem the CFO is trying to fix.

The workflow

The idea: feed the model your raw notes and headcount data, force it to separate hard savings (reallocated hours × loaded cost) from soft claims (quality, morale, candidate NPS), and make it flag anything that isn't defensible. Then generate a one-pager and a rebuttal sheet for the questions Finance will ask.

Step 1 — Structure the raw evidence

Paste your notes, invoice lines, and headcount info into this prompt. It normalizes everything into a table before any math happens.

You are an HR operations analyst preparing an AI ROI evidence dossier for a quarterly finance review. You will receive raw, messy input: pilot notes, time-study jottings, license costs, and headcount context.

Your job in this step is ONLY to structure the evidence. Do not calculate ROI yet.

Produce three tables in markdown:

TABLE A — Tools & Costs
Columns: Tool name | Use case (recruiting, policy, L&D, etc.) | Annualized license cost (USD) | Implementation/training cost (USD, one-time) | Notes on cost confidence (High/Med/Low + why)

TABLE B — Claimed Time Savings (per tool, per role)
Columns: Tool | Role affected | Headcount using it | Hours saved per person per week (as reported) | Source of estimate (self-report, time study, system log) | Confidence (High/Med/Low)

TABLE C — Soft/Qualitative Claims
Columns: Tool | Claim (e.g., better candidate experience, fewer biased screens, faster policy turnaround) | Evidence type (survey, anecdote, metric) | Measurable? (Y/N) | What it would take to verify

Rules:
- If a number is missing, write "NOT PROVIDED" — do not invent one.
- If a claim is vague ("saves a lot of time"), put it in Table C, not Table B.
- At the bottom, list every assumption you had to make to categorize an item.

Input follows below.

Q3 AI tool review — HR — notes dump

Tools in use:
1. HireLoop copilot (resume screening + interview scheduling). Licenses: 12 recruiters @ $180/mo. Rolled out Feb. Training: two 2-hr sessions, ~$4,000 all-in with the consultant.
2. PolicyDraft AI (drafts + redlines HR policies). Team license flat $28,000/yr. No formal training.
3. LearnPath (personalized L&D nudges to employees). $6/employee/mo, 1,400 employees on it. Rolled out May.
4. InterviewFair (structured interview scoring, bias flagging). Pilot only — 40 hiring managers, $95/seat/mo since March.

What team leads told me:
- Recruiting lead Priya says HireLoop "saves each recruiter maybe a day a week, honestly more during high-volume weeks." Time study we ran in June across 5 recruiters showed 6.2 hrs/week average reduction in screening + scheduling time.
- Comp & policy lead Marcus: PolicyDraft cut policy turnaround "in half." First drafts that used to take him ~10 hrs now take ~4. He's the only heavy user. Two other analysts use it lightly, maybe 1 hr saved per policy, ~3 policies/mo each.
- L&D director Jenna: no time savings claim, but completion rates on compliance modules went from 71% to 84% since May. Attributes it to LearnPath nudges. She thinks engagement scores are up but hasn't pulled the number.
- Talent acquisition head Devon: InterviewFair flagged 23 potentially biased interview questions in Q3, hiring managers "seem more consistent." No hours saved claim. Candidate NPS up 4 points but he admits other things changed too (new careers site).

Headcount / loaded cost context:
- Recruiter fully loaded cost: $95/hr
- HR analyst fully loaded cost: $72/hr
- Comp & policy lead loaded cost: $110/hr
- Hiring managers are not HR headcount — their time is billed to their business unit.
- Assume 46 working weeks/year for savings calculations.

Step 2 — Calculate defensible ROI, separate hard from soft

Now run the numbers, but with guardrails so nothing sneaks in unsupported.

Using the three tables you just produced, calculate the ROI dossier. Follow these rules strictly:

1. HARD SAVINGS ledger — only include a line if ALL of the following are true:
   - Hours saved has a non-self-report source (time study or system log), OR is a self-report explicitly corroborated by a second data point.
   - Headcount and loaded cost are known.
   - Formula: hours/week × weeks/year × headcount × loaded $/hr = annualized $ saved.
   - Show the arithmetic inline.

2. PROVISIONAL SAVINGS ledger — self-reported hours with no corroboration. Same math, but label each line "PROVISIONAL — requires validation" and cut the estimate by 50% as a discount for unverified self-report.

3. SOFT / STRATEGIC value — list qualitative wins from Table C. Do NOT assign dollar values. State what would need to be true to monetize each.

4. COSTS — sum annualized license + amortize one-time implementation over 12 months (unless useful life is stated).

5. NET ROI TABLE — per tool: Hard savings | Provisional savings | Annual cost | Hard-only net | Hard+Provisional net | Payback months (hard-only).

6. CONFIDENCE FLAGS — for each tool, one sentence on the weakest link in the calculation.

7. THINGS I REFUSED TO CALCULATE — list any claim you dropped and why. Better to omit than to fabricate.

Do not round aggressively — show real numbers, then a rounded summary line.

Step 3 — Produce the one-pager and the rebuttal sheet

The dossier is only useful if it survives contact with Finance. This step generates both the summary and the pushback prep.

Now produce two deliverables, in this order:

DELIVERABLE 1 — Executive one-pager (max ~350 words, markdown, no tables wider than 5 columns):
- Headline number: total hard annualized savings vs. total annual cost, expressed as net $ and payback months.
- Second number: hard + provisional, clearly labeled as the upper bound pending validation.
- Per-tool line summary (one row each).
- Three-bullet "What we are NOT claiming" section — this is the credibility move; name the soft wins we chose not to monetize.
- Two-bullet "What we need to validate next quarter" section — the specific data that would move provisional lines into hard.

DELIVERABLE 2 — Rebuttal sheet for the CFO meeting:
For each of the likely challenges below, write a 2-3 sentence honest answer grounded ONLY in the dossier data. If the honest answer is "we don't know yet," say that.

  Q1: "How do you know these hours saved aren't just absorbed as slack instead of reallocated to higher-value work?"
  Q2: "Self-reported hours are notoriously inflated. What's your discount?"
  Q3: "If we cut the tool with the worst payback tomorrow, what breaks?"
  Q4: "What's the bias/candidate-experience risk you're accepting by scaling these tools, and who owns it?"
  Q5: "Which of these numbers would you stake your credibility on, and which wouldn't you?"

Keep the tone plain and non-defensive. A good rebuttal sheet concedes weak points early.

Gotchas

• Self-report inflation is real. The 50% discount on provisional hours is a starting convention, not gospel. If your org has done a proper time study before and knows the local inflation factor, override it.
• Loaded cost ≠ headcount reduction. Hours saved are only cash-real if the freed hours actually get reallocated or a role gets un-backfilled. The rebuttal sheet forces you to be honest about this; don't skip Q1.
• The bias/candidate-experience line is not a rounding error. Tools like the interview-scoring one in the sample carry policy risk. Keeping their value in the "soft" column is deliberate — don't let the model be talked into monetizing candidate NPS lifts that co-occurred with other changes.
• The model will occasionally try to sum "hard + provisional" as the headline. Read the one-pager and make sure hard-only is the top-line number. That's what survives CFO scrutiny.
• Garbage in still applies. If your team leads gave you round numbers with no time-study backing, most of your ledger will end up provisional. That's the correct output — it tells you what to instrument next quarter.

Time saved

Manual assembly of a defensible AI ROI packet — chasing leads, normalizing invoices, doing the arithmetic, drafting the one-pager, prepping for Finance questions — typically eats 4-6 hours for an HR ops lead running this quarterly. This workflow compresses the structuring, math, and drafting to roughly 30-45 minutes of paste, review, and correction. The estimate assumes you already have the raw notes in hand; gathering them is unchanged.

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

# AI ROI Evidence Dossier — HR Q3 Pilot Review ### Structured Evidence Tables (Pre-Calculation) --- ## TABLE A — Tools & Costs | Tool Name | Use Case | Annualized License Cost (USD) | Implementation / Training Cost (USD, one-time) | Notes on Cost Confidence | |---|---|---|---|---| | HireLoop | Recruiting (resume screening + interview scheduling) | $25,920 | $4,000 | High — per-seat price confirmed ($180/mo × 12 recruiters × 12 mo); training cost invoiced via consultant, all-in figure provided | | PolicyDraft AI | Policy drafting & redlining | $28,000 | NOT PROVIDED | Med — flat annual license figure stated directly; no training cost mentioned and none assumed; unclear if renewal pricing is fixed | | LearnPath | L&D (personalized compliance/learning nudges) | $100,800 | NOT PROVIDED | Med — per-employee rate stated ($6/mo × 1,400 × 12 mo); headcount fluctuation could shift this; no implementation cost provided; rollout was May so current-year cost is partial (~8 mo = $67,200), but annualized figure used here per instructions | | InterviewFair | Recruiting / interview quality & bias mitigation | $43,320 | NOT PROVIDED | Low — pilot only, 40 seats since March (~10 mo to year-end = $38,000 partial-year); annualized at $95/seat/mo × 40 × 12 for comparability; no implementation cost provided; seat count and pricing may change if pilot expands or is discontinued | --- ## TABLE B — Claimed Time Savings (per tool, per role) | Tool | Role Affected | Headcount Using It

…output truncated at 250 words.

Step 2 output

# AI ROI Evidence Dossier — HR Q3 Pilot Review ### ROI Calculation Document (Finance Review Draft) --- ## PRELIMINARY NOTES ON SCOPE - Loaded hourly rates and 46-week working year are taken directly from the input; no substitutions made. - Hiring manager time is explicitly out of scope (billed to business units). No hiring manager savings are calculated anywhere. - "Annualized license cost" figures are run-rate annuals from Table A. Partial-year actuals are noted where relevant but ROI is calculated on annualized basis for comparability. - Implementation costs are amortized over 12 months (no useful-life figure was stated for any tool). --- ## SECTION 1 — HARD SAVINGS LEDGER Criteria: time-study or system-log source, known headcount, known loaded cost rate. --- ### Hard Saving H-1: HireLoop — Recruiters Basis: June time study across 5 recruiters showing 6.2 hrs/week average reduction in screening and scheduling time. Headcount: 12 recruiters. Loaded cost: $95/hr. Working weeks: 46/yr. Arithmetic: `` 6.2 hrs/week × 46 weeks/year × 12 recruiters × $95/hr = 6.2 × 46 = 285.2 hrs/year per recruiter = 285.2 × 12 = 3,422.4 hrs/year across team = 3,422.4 × $95 = $325,128.00 `` Annualized Hard Saving — HireLoop: $325,128.00 --- ### Hard Saving H-2: LearnPath — Compliance completion rate lift Evaluated for inclusion. The 71% → 84% completion rate improvement is a system metric, but it represents a compliance outcome, not a labor-hour saving. No hours-saved figure is attached to it. It is therefore moved to Section 3 (Soft/Strategic Value) rather

…output truncated at 250 words.

Step 3 output

# DELIVERABLE 1 — Executive One-Pager ## HR AI Tools: Q3 ROI Summary --- ### Headline Numbers Hard savings (time-study verified only) Annual savings: $325,128 | Annual cost: $202,040 Net: +$123,088 | Portfolio payback: 7.5 months Upper bound (hard + provisional, pending validation) Net: +$125,680 The $2,592 difference reflects discounted self-report estimates for PolicyDraft AI analyst users. Do not present this as confirmed. --- ### Per-Tool Summary | Tool | Hard Savings | Provisional | Annual Cost | Net (Hard Only) | Payback | |---|---|---|---|---|---| | HireLoop | $325,128 | $0 | $29,920 | +$295,208 | 1.1 months | | PolicyDraft AI | $0 | $2,592 | $28,000 | −$28,000 | Unproven | | LearnPath | $0 | $0 | $100,800 | −$100,800 | Unproven | | InterviewFair | $0 | $0 | $43,320 | −$43,320 | Unproven | | Total | $325,128 | $2,592 | $202,040 | +$123,088 | 7.5 mo | Three of four tools currently show no calculable financial return. The portfolio's positive net is driven entirely by HireLoop. --- ### What We Are NOT Claiming - LearnPath's compliance lift is not monetized. Completion rates rose 71%→84% — a real metric — but we cannot yet isolate LearnPath as the cause or attach a verified dollar value to the improvement. - InterviewFair's bias flags and NPS gain are not monetized. 23 flagged questions and a 4-point candidate NPS rise are logged as strategic value; the NPS result has a known confound (new careers site) and litigation-risk modeling

…output truncated at 250 words.