By Dylan Parkes | August 6, 2026
The Interview
Below is an “interview” with Anthropic’s Claude Sonnet 5 model. My questions and its (their?) answers are reproduced without editing. Thus, the liberal use of the em—dash and the overt formatting when I asked to reproduce the discussion as an interview.
I found the responses to be generally faithful to the questions and not distasteful or unreasonable. That fact in and of itself is a bit alarming but that may be me assigning self-awareness to the tool more than it being demonstrated.
I decided not to edit the answers as I feel it risked my own bias which may already be present in the prompt questions themselves. I leave it with the reader to decide. I hope you found this as bizarre and interesting as I did.
An Interview on Modeling, Judgment, and Career Development with Claude’s Sonnet 5 model
On Overall Capability
Question
There is a lot of chat and rhetoric out there about your capacity both to support project finance modelers in their role and to replace them in building and analyzing models. In your estimation and considering the power of your other learning models (such as Opus and Fable), how would you rate your capacity to perform financial modeling tasks?
Answer Claude
Where the value is genuine
● Building and auditing the mechanics of a project finance model: circularity between the cash flow waterfall and debt sizing, DSCR/LLCR/PLCR calculations, sculpted debt repayment, reserve account logic, tax equity structures (partnership flips, inverted leases), MACRS depreciation, ITC/PTC mechanics.
● Catching structural errors — sign errors, off-by-one period issues, circular reference breaks, inconsistent day-count conventions, timing mismatches between construction and COD.
● Explaining or reproducing standard industry conventions (six-month DSCR conventions, minimum DSCR covenant testing, cash sweep mechanics, debt service reserve sizing off P50/P90).
● Drafting formulas and VBA/Python for scenario and sensitivity automation, or translating model logic into a memo a credit committee could read.
● Sanity-checking outputs against sponsor-side heuristics.
Where the risk is real
● No live access to market data — PPA pricing benchmarks, merchant power curves, interconnection queue costs, current debt pricing/spreads, or tax equity market terms.
● No independent verification the way a second analyst with domain judgment provides — plausible-looking formulas can encode a wrong assumption.
● No accountability. There is no E&O insurance behind an AI-generated number; the modeler who signs off remains responsible.
● Novel deal structures and jurisdiction-specific tax quirks raise the risk of confidently plausible but subtly wrong logic.
On the question of “replace”: the parts of the job that are genuinely valuable — knowing which assumptions a lender will push back on, structuring around a counterparty’s credit concerns, negotiating covenant packages, knowing when a model’s elegant logic doesn’t match how a term sheet will actually be negotiated — are not replaced by AI. The more accurate framing is force multiplier on the mechanical and drafting load, not an unsupervised source of bankable output.
On Workflow Integration
Question
If I were planning to use you as a tool alongside junior analyst modelers, where in the process (document review, template buildout, modeling, VBA, scenario and sensitivity analysis, review and audit, etc.) do you believe you have the greatest value add and why?
Answer Claude
- Highest Document review & term extractionPulling structured terms from PPAs, credit agreements, EPC contracts, and term sheets pricing curves, escalators, distribution conditions, covenant thresholds. High leverage, but load-bearing terms still need a targeted human check.
- Highest — Formula audit and review Systematically walking a full model for consistency across periods, circularity breaks, and sculpting math. A strong first pass so senior reviewers check findings rather than startingfrom zero.
- High — VBA/Python for scenario automation Coding is a strength independent of finance-specific judgment; errors surface quickly since the macro either runs correctly or it doesn’t
- Medium — Template buildout Fast for standard tabs and boilerplate structure, but letting AI build the skeleton risks juniors skipping the rep count on core logic they need to own and defend.
- Lower, needs heavy oversight — Core modeling judgment & deal-specific structuring The highest-risk area: plausible-looking but potentially wrong logic on novel structures. Should remain analyst-driven with AI as a drafting accelerant only.
One point worth naming directly: the areas of greatest AI value-add (document extraction, audit) are also where junior analysts build the pattern recognition that eventually makes them senior. Deploying AI there without care risks faster model turnaround today at the cost of slower analyst development a case for using it alongside the analyst’s own work early in a career, not instead of it.
On Competitive Positioning
Question
How would you honestly rank yourself amongst your peers (GPT, Grok, Gemini, DeepSeek, Meta) in this space? Where are your peers more or less effective?
Answer Claude
No rigorous, project-finance-specific benchmark exists across model providers, and any self-assessment from an interested party deserves real skepticism.
What follows is inference from general-capability benchmarks (coding, long-document reasoning, structured extraction) as a proxy — a reasonable starting prior, not a substitute for a direct head-to-head pilot on an actual template and deal.
General benchmark landscape (mid-2026)
The frontier has fractured rather than consolidated, with several model families sitting within a few benchmark points of one another. Claude’s newer models lead frontier coding benchmarks; Gemini 3.1 Pro leads scientific and abstract reasoning; GPT-5.4 leads computer-use benchmarks; DeepSeek and lighter-tier models anchor the low-cost end for bulk processing.
Translated to project finance sub-tasks
● Document review/extraction: roughly a toss-up with Gemini — Gemini may have an edge on clean tabular extraction and Google Workspace integration; Claude may hold an edge on dense narrative contract language and cross-referenced covenant packages.
● VBA/Python/scenario automation: the area of highest relative confidence, tracking directly with coding-benchmark strength.
● Formula audit: similarly strong by extension from coding ability, though this is extrapolation rather than a direct measurement.
● Novel deal structuring and judgment calls: not meaningfully differentiated across providers this isn’t a skill current benchmarks capture well.
The honest caveat: this ranking is built on general-capability proxies, not a direct project-finance-modeling benchmark. Testing two or three models head-to-head on an actual (anonymized) deal template is the only way to get an answer reliable enough to base a workflow decision on.
On Career Development
Question
If you were making a recommendation on how to pursue project finance and modeling as a career to a young person, where would you recommend they focus their learning and how would LLM/AI models fit into that scope?
Answer Claude
Priority learning areas
● Accounting and credit fundamentals, cold — understanding why a DSCR covenant, cash sweep, or tax equity flip point is structured the way it is, not just where to look it up.
● Building models by hand before building them with AI — circularity, sculpting, and waterfall logic learned through the debugging process itself, not just the finished output.
● Contract literacy — reading PPAs, EPC contracts, credit agreements, and tax equity documents directly, not just working from extracted summaries.
● Market and structuring literacy — current tax equity structures, interconnection queue risk allocation, merchant tail pricing; the fastest-aging part of the job, requiring active engagement with deal flow.
● Communication — explaining a model’s assumptions and sensitivities to a credit committee concisely; a ceiling on career progression independent of technical skill.
Where AI fits into that learning path
● As a tutor for concept-checking, not a crutch for concept-skipping — asking for explanations or critiques of self-built models rather than outsourcing the build.
● Heavily for audit and review practice once a baseline exists — build, then have AI find errors, then find what it missed.
● For volume on tedious early-stage work (contract term extraction, VBA scaffolding), freeing time for judgment-heavy tasks — without letting it replace direct document reading entirely.
● Treated the way one would treat a smart but unaccountable junior’s output — always independently verified, since that verification habit is itself the skill lenders and investment committees are paying for.
The people who do best in this field over the next decade are unlikely to be those who avoid AI, and equally unlikely to be those who lean on it uncritically. The likeliest advantage belongs to those who use it deliberately enough during training that their own judgment gets sharper rather than outsourced.
Pivotal180 Courses
If you want to learn and expand your current abilities to understand, analyze, and use financial models, come take a class with us!
Whether you’re analyzing data centers, renewable energy projects, or transport infrastructure, Pivotal180 teaches you how to build complete financial models from scratch and apply these skills to real-world projects.
Our courses break down the essential project finance concepts through hands-on, practical examples, so you don’t just learn, you do. Choose from self-paced online programs, live instructor-led sessions, or tailored corporate training designed to fit your organization’s needs.
Check out the links below to learn more about Pivotal180 and our financial modeling courses. Experience the Pivotal180 difference.
