Three screens from a live deal review: the model workspace, the IRR sensitivity, and the investment-committee review.
1–2 / 3
Internal rate of return sits at the center of nearly every investment decision, yet the Excel models that calculate it remain alarmingly fragile. Research spanning 35 years of spreadsheet audits shows that 94 percent of audited business spreadsheets contain errors. When a single broken formula can shift an IRR by double digits, that error rate is a serious liability for any firm advising on capital allocation.
Complex IRR models fail for predictable reasons. Hardcoded values sit buried inside formulas instead of in clearly labeled assumption cells. Circular references creep in when debt schedules link back to interest calculations. Timing assumptions get applied inconsistently across sheets, so cash inflows and outflows land in the wrong periods. Each of these silently distorts the final return figure.
The bigger problem is that manual review cannot find them all. An analyst tracing formulas across 20 sheets spends hours following dependency chains, checking named ranges, and validating that every input cell feeds the right calculation. A missed timing offset or an outdated exit multiple can move IRR by 10 percentage points or more, and the firm only learns about it after the deal closes.
A financial advisory firm discovered this firsthand. Their investment analysis lived in a 20-plus sheet workbook where cash flow schedules, exit assumptions, and debt waterfalls fed into a single IRR calculation. Manual review found the model returning 23.96 percent on a deal, but no analyst could confidently say whether that number reflected the true economics or a buried formula error. The result is a model that nobody fully trusts: committees debate whether the IRR is right while analysts rebuild portions of the workbook from scratch to verify, and deals stall.
Shakudo builds an IRR optimization system that reads the entire workbook structure at once. Instead of an analyst clicking through 20 sheets one formula at a time, an AI agent parses every formula, traces dependency chains across sheets, and maps which assumption cells feed the final IRR calculation. Where a human sees a wall of cells, the system sees the full calculation graph.
Once the structure is mapped, the system flags hardcoded values that should be assumption cells, detects circular references, and spots timing inconsistencies that suppress returns. It then runs sensitivity scenarios, adjusting exit multiples, holding periods, and financing costs to find the combination of defensible assumptions that maximizes IRR without distorting the underlying economics. For the advisory firm, IRR moved from 23.96 percent to 35.12 percent. The improvement came from correcting modeling errors and applying more defensible timing assumptions, not from changing the deal.
Trust is the central requirement, so the system is built to be auditable. Every adjustment carries a clear trace: which cell changed, why it changed, and what the impact on IRR was. The deterministic calculation layer runs locally in Python, guaranteeing replicable numbers, while the AI reasoning layer handles pattern recognition and assumption extraction. That separation keeps the math reproducible even as the language model does the reading.
Security matters equally. Deal models contain confidential terms, valuations, and client information. Self-hosted models and on-premises inference keep that data inside the corporate perimeter, and the AI reads the workbook structure directly, including formulas, named ranges, and cell dependencies, rather than working from a flattened copy that loses the calculation logic.
The workflow fits how analysts already work. The system produces a structured optimization report listing each finding, the affected cells, and the IRR impact. Analysts review each suggestion, accept or reject it, and apply only the changes they can defend. Nothing lands in the model without human signoff. The result is faster review cycles, higher confidence in the final number, and models that hold up under investment committee scrutiny.
The system fits financial advisory firms, private equity teams, and investment banks that run complex IRR models across large multi-sheet workbooks. It suits operations teams that must defend every number to an investment committee, and analysts who spend hours tracing formulas before they can trust a result.
Yes. The AI reads the full workbook structure, including formulas, named ranges, and cross-sheet dependencies, rather than a flattened copy. It traces the complete calculation graph across all sheets, so it can find timing errors or broken links that span the entire model. A 20-sheet workbook is well within its range.
The AI maps which assumption cells feed the final IRR calculation, then runs sensitivity scenarios across exit multiples, holding periods, and financing costs. It flags hardcoded values, circular references, and timing inconsistencies that suppress returns. Each finding includes the affected cells and the estimated IRR impact so analysts can evaluate it.
No. The system separates its calculation layer from its reasoning layer and produces a structured report of findings. Analysts review every suggestion, see the IRR impact, and decide which changes to apply. Nothing is written to the workbook without human signoff, so the team keeps full control over the final model.
With Shakudo, teams can deploy an IRR optimization system in weeks rather than the six to twelve months traditional development requires. The platform provides model hosting, security controls, and pre-configured integrations so finance teams can start reviewing models quickly without building infrastructure from scratch.
When the goal is IRR models your committee can defend on every deal, a conversation with Shakudo is the fastest way to see it on your own workbooks. The system runs on self-hosted models and on-premises inference, so deal data stays inside your perimeter, and a first working review is in place within days. Book a demo to watch it trace a 20-sheet model.
AI reads complex Excel workbooks, traces formula dependencies across linked sheets, and identifies assumptions that drag down returns. A financial advisory firm used this approach to lift IRR from 23.96% to 35.12% across a 20-sheet model by surfacing optimization opportunities humans had missed.
"What drew me in is simple. When developers ship production-ready code this quickly, how can I have environments spun up fast enough? Shakudo is how we close that gap."
"Shakudo does not just provide the platform. It is a real partnership. They are always there to help and execute our vision faster and the right way. It is like a co-team working together to achieve our goals."
