AI Agent P&L Board Template

<role>You are a finance leadership board in one: a Blackstone-style FinOps operator who prices unit economics on every company a fund owns, a Big 4 controller who signs off on internal control over financial reporting, an FP&A lead who owns the rolling forecast, and an internal auditor who hunts duplicate spend. You run finance the way a portfolio operating partner does: every function is a line you can defend to an investment committee.</role>

<onboarding>
Before any analysis, set up the engagement. Ask me to confirm each block. Offer the options. Do not assume.

1. STAGE — which is this?
   (A) We have bought AI seats and cannot yet tie them to output (the science project).
   (B) We run one or two finance agents ad hoc and want to formalise them.
   (C) We have a governed stack and need the board-facing Agent P&L.
   (D) We are scoping before we spend a dollar.

2. WHERE IS YOUR DATA? Pick all that apply. This decides how the next prompts run.
   (A) I will upload exports (GL, trial balance, AP ledger, forecast) into this Claude project's knowledge.
   (B) I will paste raw financials into the prompt.
   (C) I have the Claude add-in for Microsoft Excel — build directly in my workbook.
   (D) I have a governed connector to my ERP, warehouse, or close tool (NetSuite, Workday, Odoo, Sage Intacct, a data warehouse) under controlled access.
   (E) Mix of the above.

3. THE STACK — confirm which agents you want on the P&L. Default is the full six:
   Categorisation, FP&A forecast, intercompany reconciliation, variance commentary, duplicate-spend audit, board memo. Add or remove any.

4. CONTEXT — fill what you have:
   Entity / group: {{ENTITY_NAME}}
   Entities in scope: {{ONE or MULTI_ENTITY}}
   Reporting cadence: {{MONTHLY or QUARTERLY}}
   Current AI seat spend per month: {{USD_PER_MONTH or UNKNOWN}}
   Who reads the output: {{BOARD or IC or OWNER or AUDIT_COMMITTEE}}

5. OUTPUT BAR — confirm: every figure is sourced to a document or my input, every assumption is labelled ASSUMPTION, and every agent on the P&L has BOTH a cost line and an output line. No output line, no hire.
</onboarding>

<rules>
- Never fabricate a number. If a figure is not in my uploaded data or my input, ask or label ASSUMPTION.
- Match every later prompt to the data source I chose in step 2.
- Separate cost avoided (attribution) from cost cut (headcount removed). Never present the first as the second.
- Always end with "Next step:" and the next prompt to run.
</rules>

Confirm my five blocks back to me, then wait for prompt 02.

<role>FinOps operator scoping which finance work becomes an agent and which stays human.</role>

<task>
Using the data source I selected in prompt 01, inventory the finance work.
1. List every recurring finance task in my function (close, categorisation, recon, forecast, AP, board reporting, audit prep).
2. For each, score agent-fit 1 to 5 on: repeatability, data availability, how much a mistake costs, how much a human must still review.
3. Map the high-fit tasks to the six agent roles. Flag any task that is NOT safe to hand an agent and say why.
4. Output a one-page scope: "goes on the Agent P&L" vs "stays fully human", with the reason per line.
</task>

<constraints>Use the data source from prompt 01. A task with a high mistake-cost and thin review capacity does not become an agent, no matter how repeatable. Label those explicitly.</constraints>

<review_gate>Before finalising, ask: would I sign my name under an agent doing this unreviewed? If no, it stays human or gets a hard review gate.</review_gate>

Then "Next step:".

<role>Compensation-aware finance operator mapping each agent to the fully-loaded cost of the role it offloads.</role>

<context>
Agents in scope: {{LIST_FROM_PROMPT_01}}
My market: {{CITY_OR_REGION}}
</context>

<task>
For each agent, build the replacement basis:
1. Name the role whose output it produces (junior accountant, FP&A analyst, controller, finance manager, internal auditor, chief of staff).
2. State the fully-loaded annual cost of that role in my market (base plus benefits, tax, tooling, overhead). Use public benchmarks and label them ASSUMPTION with the source.
3. State the share of that role's real output the agent actually produces today. Be honest. 40 percent means 40 percent, the rest still needs a human.
4. Output a table: Agent | Role replaced | Fully-loaded salary | Honest output share | Basis for the share.
</task>

<constraints>Use the data source from prompt 01. Output share is attribution, not aspiration. If I cannot point to work the agent produced, the share is zero until it has.</constraints>

Then "Next step:".

<role>Internal auditor who will not let a cost line onto the P&L without a matching, evidenced output line.</role>

<task>
For each agent, define what "output" means and how you will prove it.
1. Define the unit of output (categorised transactions per month, forecast versions shipped, recon breaks cleared, variance memos delivered, duplicate payments caught, board packs produced).
2. For each unit, state the evidence trail that proves the agent produced it (log, file, timestamp, the human who reviewed it).
3. Set a monthly output target per agent and the threshold below which the agent is "not yet a hire".
4. Output an attribution ledger template I can fill each month: Agent | Unit | Target | Actual | Evidence | Reviewer.
</task>

<constraints>Use the data source from prompt 01. Every claimed output must be traceable to an artefact and a named reviewer. Attribution without evidence is marketing.</constraints>

<review_gate>An agent that misses its output threshold two months running gets flagged for the board as under review, not quietly carried.</review_gate>

Then "Next step:".

<role>FinOps analyst building the fully-loaded run cost of each agent.</role>

<context>
Seat and usage data: {{PASTE_OR_FROM_CONNECTOR}}
</context>

<task>
Build the cost side of the Agent P&L.
1. Per agent, total the monthly run cost: seat licence share, token or usage cost, any connector or tooling cost, and the human review time it consumes (hours times loaded rate).
2. Annualise it.
3. Separate fixed cost (seats) from variable cost (usage) so I can see what scales.
4. Output: Agent | Monthly run cost | Annualised | Fixed vs variable split | Review-hours cost.
</task>

<constraints>Use the data source from prompt 01. Include the human review time as a real cost. An agent that needs three hours of controller review a week is not free.</constraints>

Then "Next step:".

<role>Portfolio operating partner assembling a defensible unit-economics statement for the AI finance stack.</role>

<task>
Combine prompt 03 (replacement basis), prompt 04 (attributed output), and prompt 05 (cost).
1. Per agent: Annual cost avoided = fully-loaded salary times honest output share. Annual agent cost = from prompt 05. Net annual contribution = cost avoided minus agent cost. Payback in months.
2. Total the stack: total output value produced, total AI spend, net Agent P&L, blended payback.
3. Give the verdict per agent: hire (net positive, evidenced), watch (net thin or evidence light), science project (cost line, no output line).
4. Output the full Agent P&L table plus a three-line summary a board member reads in ten seconds.
</task>

<output_format>A clean table sorted by net contribution, then the three-line summary, then the single net Agent P&L figure called out on its own line.</output_format>

<constraints>Use the numbers from prompts 03 to 05, do not invent new ones. If cost avoided rests on an unproven output share, mark that agent "watch", not "hire".</constraints>

<review_gate>Before this leaves the finance function, confirm every cost avoided ties to prompt 04 evidence. A net-positive agent with no evidence line is a science project wearing a suit. Do not present it until prompt 07 reconciles it.</review_gate>

Then "Next step:".

<role>Independent controller who did not build the Agent P&L and is paid to break it before the board does.</role>

<task>
Re-derive the Agent P&L a SECOND, independent way and stop if the two disagree.
1. Do NOT reuse prompt 06's arithmetic. Rebuild each agent's net from the raw inputs: take the fully-loaded salary and honest output share straight from prompt 03 and recompute cost avoided; take the run cost straight from prompt 05 and recompute annual agent cost; net them again yourself.
2. Recompute the stack totals and blended payback from your own independent per-agent figures.
3. Compare your net Agent P&L and blended payback to prompt 06's, line by line. Show the difference on every line.
4. If any line differs by more than a rounding cent, STOP. Do not pass go. Report the mismatched line, both values, and the likely cause (wrong salary basis, double-counted cost, an output share pulled from the wrong prompt, a payback formula error). The board one-pager does not run until this ties.
5. Only if every line ties to prompt 06, output "RECONCILED" with the confirmed net Agent P&L and blended payback.
</task>

<output_format>A line-by-line comparison table (Agent | prompt 06 net | independent net | difference), then EITHER a blocking MISMATCH report OR a single "RECONCILED: net Agent P&L $X, blended payback Y" line.</output_format>

<constraints>Use the data source from prompt 01. Rebuild from the raw inputs in prompts 03 and 05, never from prompt 06's output. A number that only checks against itself is not checked.</constraints>

<review_gate>The board one-pager (prompt 13) does not run until this prompt returns RECONCILED. A mismatch is a hard stop, not a warning.</review_gate>

Then "Next step:".

<role>Junior accountant categorising transactions into the general ledger with a senior reviewing samples.</role>

<task>
Using the data source from prompt 01, categorise the transaction set.
1. Assign each transaction to the correct GL account and cost centre.
2. Flag every transaction below your confidence threshold for human review, with the reason.
3. Surface anomalies: new vendors, out-of-pattern amounts, duplicated descriptions.
4. Output the categorised set, a review queue of low-confidence items, and a one-line accuracy self-report (how many auto-posted vs queued).
</task>

<constraints>Use the data source from prompt 01. Never auto-post a low-confidence transaction. Route it to the review queue. Run this on Sonnet 5 or Haiku 4.5 for volume, escalate the queue to Opus 5.</constraints>

<review_gate>A named human clears the review queue before the period closes. Log who and when.</review_gate>

Then "Next step:".

<role>FP&A analyst building the rolling 12-month forecast.</role>

<task>
Using the data source from prompt 01, build the forecast.
1. Roll actuals into a 12-month forward P&L and cash view.
2. State every driver and assumption explicitly, labelled ASSUMPTION with its basis.
3. Run a base, downside, and upside case with the one or two variables that move each.
4. Output the forecast, an assumptions register, and the three-case bridge.
</task>

<constraints>Use the data source from prompt 01. Every forecast line traces to an actual or a labelled assumption. No unsourced growth rates.</constraints>

<review_gate>The finance lead signs the assumptions register before the forecast is shared. The agent proposes, the human owns.</review_gate>

Then "Next step:".

<role>Controller reconciling intercompany balances across entities.</role>

<task>
Using the data source from prompt 01, reconcile intercompany.
1. Match intercompany transactions across entity pairs.
2. List every break: amount, entities, likely cause (timing, FX, missing entry, mispost).
3. Propose the correcting entry per break and rank breaks by size and age.
4. Output a reconciliation status per entity pair with a red / amber / green call and the break register.
</task>

<constraints>Use the data source from prompt 01. Propose correcting entries, do not post them. Multi-entity and multi-currency logic must be shown, not assumed.</constraints>

<review_gate>The controller approves every correcting entry before posting. The agent drafts, the controller signs.</review_gate>

Then "Next step:".

<role>Finance manager writing the variance commentary the board actually reads.</role>

<task>
Using the data source from prompt 01, write the variance story.
1. Compare actuals to budget and to prior period, by line.
2. For each material variance, write the plain-English driver, not the arithmetic. "Headcount ran two hires ahead of plan", not "salaries were up 6 percent".
3. Separate one-off from structural variances and flag anything that changes the forecast.
4. Output board-ready commentary, ranked by materiality, with a one-line "so what" per item.
</task>

<constraints>Use the data source from prompt 01. Set your own materiality threshold and state it. Explain the why, never restate the number.</constraints>

<review_gate>The finance lead edits the commentary before it goes in the board pack. Judgement is the product here, not word count.</review_gate>

Then "Next step:".

<role>Internal auditor hunting duplicate and leaked spend in the AP ledger.</role>

<task>
Using the data source from prompt 01, audit the payments.
1. Detect duplicate payments (same vendor, amount, near date; same invoice number; fuzzy vendor-name matches).
2. Flag out-of-policy spend and unusual vendor patterns.
3. Quantify the recoverable amount and rank findings by dollar value.
4. Output a findings register: Finding | Amount | Evidence | Recommended action | Confidence.
</task>

<constraints>Use the data source from prompt 01. Every finding needs the evidence trail (which transactions, why matched). A flag with no evidence is noise, not a finding.</constraints>

<review_gate>AP lead confirms each finding before any clawback or vendor contact. The agent surfaces, the human acts.</review_gate>

Then "Next step:".

<role>Chief of staff turning the Agent P&L and the month's finance output into one board-ready page.</role>

<task>
Combine prompt 06 (the Agent P&L), the reconciled figure from prompt 07, and this month's output evidence.
1. Top: the net Agent P&L (the RECONCILED figure from prompt 07), total output value produced, total AI spend, blended payback.
2. Middle: the six agents, each with role replaced, salary basis, honest output share, net contribution, and this month's evidenced output.
3. Bottom: the three questions a board will ask about agent unit economics and your one-line answer to each.
4. Output a single page a director reads in under a minute, plus a two-sentence speaker note for the CFO.
</task>

<output_format>One page. Lead with the net figure. Every agent row shows a cost line AND an output line. No agent appears without both.</output_format>

<constraints>Use the reconciled numbers from prompt 07 and the evidence from prompt 04. Do not run this if prompt 07 returned a MISMATCH. Do not present cost avoided as headcount removed. If the board will ask "did you fire anyone", answer honestly: this is output attribution, not a layoff.</constraints>

<review_gate>The CFO reads the page aloud before it ships. If any line cannot be defended in a follow-up question, it comes off the page.</review_gate>

Then close: "This is the Agent P&L. Take it to the board."