Capital budgeting is where finance teams spend their most valuable hours on their least valuable work. Every acquisition, equipment purchase, facility expansion, or software investment requires the same ritual: build a cash flow model, discount it at the right rate, calculate NPV and IRR, work out the payback period, then rebuild the whole thing when someone asks “what if revenue is 15% lower?” A mid-sized company might evaluate thirty capital requests a year. A large one might process hundreds. Each model takes hours to construct, is prone to formula drift between versions, and often arrives at the committee meeting with assumptions nobody documented.
The friction compounds in three places. First, consistency: two analysts modeling similar projects will structure terminal values and working capital differently, making comparisons unreliable. Second, speed: when a CFO wants a sensitivity table by tomorrow, the analyst rebuilds rather than analyzes. Third, documentation: the rationale behind a discount rate or a salvage value assumption lives in someone’s inbox, not in the model. These are not analytical problems. They are process problems, and they are exactly the kind of work large language models handle well when given the right structure.
An AI assistant does not replace the judgment behind a capital decision. It replaces the mechanical work of building, checking, and re-running the analysis. With a well-constructed prompt, you can go from a project brief to a complete NPV, IRR, and payback breakdown in minutes, with assumptions stated explicitly, sensitivities included, and the output formatted for a decision memo rather than a spreadsheet dump. The two prompts below show how to set this up inside Claude, with the first covering single-project analysis and the second covering comparison across a portfolio of competing requests.
Why structure matters more than model choice
A generic request like “calculate NPV for this project” produces generic output. The difference between a useful result and a useless one is the context you supply: the discount rate policy, the treatment of inflation, the required payback threshold, the audience for the output. This is why both prompts below follow the anatomy-of-a-prompt structure: task, success criteria, reference material, context file, and a mandatory clarifying-questions step before execution. That final step is not ceremony. It forces the model to surface ambiguities (Is the discount rate nominal or real? Are cash flows mid-year or year-end?) before it commits to numbers you might otherwise trust without checking.
First, read these files completely before responding:
capital_budgeting_policy.md — our firm’s discount rate policy, hurdle rates by project category, and required payback thresholds
modeling_standards.md — our conventions for cash flow timing, terminal value treatment, inflation handling, and rounding
Here is a reference for what I want to achieve:
[Upload a prior approved capital request memo as markdown, or describe its structure — sections, table layout, and level of detail]
Here’s what makes this reference work:
– Assumptions are stated explicitly in a dedicated table before any results appear
– Cash flows are shown year by year, never summarized into a single figure
– NPV, IRR, and payback each appear with a one-line interpretation, not just a number
– Sensitivities cover at least three variables with a clear base case marked
– Tone is factual and decision-oriented; no hedging language, no filler
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: structured analysis memo, 600-900 words plus tables, in markdown
Recipient’s reaction: the committee should see the assumptions, trust the math, and know exactly what to approve or reject
Does NOT sound like: a textbook, a sales pitch, or a spreadsheet exported without commentary
Success means: the memo passes review with zero requests to re-run the model before the meeting
My context file contains my standards, constraints, audience. Read it fully before starting.
Project details for this analysis:
[Project name and description]
[Initial investment and timing]
[Projected annual cash flows or revenue/cost drivers]
[Project life in years]
[Discount rate or instruction to use policy rate for category]
[Salvage or terminal value assumptions, if any]
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
Note the deliberate sequencing in that prompt. The clarifying-questions gate catches the errors that make capital models embarrassing: mismatched discount rate conventions, cash flows entered as profit rather than free cash flow, or a project life that conflicts with the depreciation schedule. Once those are resolved, the model produces results you can defend line by line.
From one project to a ranked portfolio
Single-project analysis is only half the job. The harder question most finance teams face is allocation: given a fixed capital envelope and eight credible requests, which ones get funded? That requires consistent methodology across projects, a ranking view, and — critically — an honest treatment of IRR’s known weaknesses, such as reinvestment rate assumptions and misleading results for projects with non-conventional cash flows. The second prompt handles this comparison layer, and it depends on the first prompt’s output being standardized, which is why the two work best as a pair.
First, read these files completely before responding:
capital_budgeting_policy.md — our discount rate policy, hurdle rates, and payback thresholds
modeling_standards.md — our conventions for cash flow timing, terminal value, and rounding
Here is a reference for what I want to achieve:
[Upload a prior portfolio ranking or funding recommendation as markdown, or describe its structure — ranking table, rationale column, and decision summary]
Here’s what makes this reference work:
– All projects are evaluated on identical assumptions and timing conventions
– The ranking table shows NPV, IRR, and payback side by side, with the primary ranking criterion stated
– Conflicts between metrics (for example, highest IRR versus highest NPV) are flagged and resolved explicitly
– A funding recommendation respects the budget cap and explains what is excluded and why
– No project is described in promotional language; every claim ties to a number
Here’s what I need for my version / SUCCESS BRIEF:
Type of output + length: portfolio comparison memo, 700-1000 words plus a ranking table, in markdown
Recipient’s reaction: leadership should be able to approve the slate on the spot and defend the exclusions if challenged
Does NOT sound like: a list of summaries stapled together, or a recommendation with hidden assumptions
Success means: the memo identifies the optimal funding combination within budget and documents every tie-break decision
My context file contains my standards, constraints, audience. Read it fully before starting.
Portfolio details:
[Total capital budget available]
[Number of projects and brief description of each, or attach the individual analyses]
[Any strategic constraints — mandatory projects, category limits, timing restrictions]
DO NOT start executing yet. Ask clarifying questions first.
Give me your execution plan (5 steps max) before you begin.
A practical tip before you run either prompt: keep your policy and standards files short and current. A two-page modeling standards document that states timing conventions, rounding rules, and hurdle rates explicitly will improve output quality more than any amount of prompt tweaking. Update it once a quarter, and both prompts stay accurate as your policy evolves.
What to try next: run the single-project prompt against a request you have already approved and compare its output to your original memo. Where the AI’s version differs — a different payback treatment, a missing sensitivity — you have found either a gap in your standards file or a gap in your original analysis. Both are worth knowing about. Once the single-project output matches your house style, move to the portfolio prompt and test it on last year’s funding round. If the ranking it produces would have changed a decision, you have a strong case for making this part of your standard process.
Published on 10 November 2026 on growwithgpt.com