Best AI Tools for Financial Modeling in 2026

The best AI tool for financial modeling in 2026 is Kepler, because the historicals in your model come from the filings’ own XBRL with every cell cited to its source, calculated cells carry live formulas over those cited inputs, and the model follows your firm’s template and definitions. When new results land, Kepler’s agents roll the model forward, flag anything the company restated, and keep every number traceable.

This list covers AI for building and updating models in Excel: three-statement models, DCFs, LBOs, comps and the quarterly roll-forward. It’s for analysts and associates in equity research, investment banking, private equity, credit and corporate development.

We build Kepler, so it’s first and labeled. Facts about the other eight tools come from their own sites, release notes and public reporting as of November 2026, linked inline, and prices appear only where published or reported.

How we picked

A model is a chain of numbers that has to tie. We asked six questions of every tool:

• Where do the historicals come from? Pulled from the filing with a source on every cell, or typed by the model.
• Formulas or hardcodes? Projections and calculations should be live formulas you can audit.
• Your template? The firm’s tab structure, color code, definitions and fiscal calendar.
• Does it refresh? Rolling forward when a 10-Q lands, with restatements and segment recasts caught.
• Is it reproducible? The same request next week should return the same model.
• Where does it work? Natively in Excel, in a web app, or both.

The order reflects these six questions, with Wall Street Prep’s scores as one input.

What Wall Street Prep’s 2026 test found

The most useful public test so far comes from Wall Street Prep, which trains analysts for banks and funds. In its comparison (published February 2026, updated June 2026), four tools got one identical prompt: build Apple’s integrated three-statement model, three historical and four forecast years, from the 10-K, the fourth-quarter release and consensus, in banking format with comments, citations and a sources tab. Graders scored the first attempt only, on the same scale Wall Street Prep uses for its trainees, across eight subtasks.

Tool tested	Overall score (of 10)
Shortcut (Excel add-in)	5.9
Claude in Excel (Opus 4.6)	5.5
Copilot Agent Mode (GPT-5)	4.4
ChatGPT 5.2 (outside Excel)	2.5
Low-rated junior analyst, for comparison	6.4

Three findings matter more than the ranking. Every tool scored zero on handling circularity, none modeled debt properly (most balanced the sheet with plugs), and none got shares outstanding right. On historical data, two of the four tools “hallucinated significant portions of historical data,” in line items that still summed to the correct subtotals, which is the kind of error a reviewer checking totals never sees. Wall Street Prep’s advice was to use these tools to get a model started and to upload the source PDFs and spreadsheets yourself rather than let the agent find the data.

The models tested have since been replaced by newer ones, and the article says rankings will be refreshed. The lesson about inputs holds: when a model types the historicals, someone has to tick and tie them.

Quick comparison

Tool	Best for	Where it works	Where the historicals come from
Kepler	Models a firm defends, on its own template	Excel, app, Claude, ChatGPT, API	The filings’ XBRL, every cell cited
Claude in Excel	A frontier model building and editing workbooks	Excel, PowerPoint, Word	Connectors and uploads, typed by the model
Shortcut	A dedicated Excel modeling agent	Excel add-in, web, Windows desktop	Uploads and connected data, typed by the agent
Daloopa (Scout)	Source-linked historicals and KPIs	Excel, MCP	Extracted data points, each linked
Microsoft Copilot in Excel	Workbook mechanics inside Microsoft 365	Excel	Your workbooks and web search
ChatGPT for Excel	Firms on ChatGPT, and Google Sheets users	Excel, Google Sheets, ChatGPT	Included and connected data, typed by the model
Rogo	Bank and PE templates on licensed data	Rogo app, Excel add-in	Licensed data from 51 integrations
AlphaSense (Canalyst)	Analyst-built company models	AlphaSense, Excel, PowerPoint	Canalyst models and standardized financials
Model ML	Deliverables across Excel, PowerPoint and Word	Web, Office plugins, Teams, Slack	Firm and licensed data, through agents

The 9 best AI tools for financial modeling

1. Kepler (our pick)

We build Kepler.

Best for: analysts and associates whose models go to a PM, an investment committee or a client, and who need every historical cell sourced, every calculation auditable and the firm’s own template and definitions applied every time.

In an AI-built model, the inputs are the hard part. A general assistant reads the filing, chooses the figures and types them into cells, so a transposed digit, a year-to-date number used as a quarter or a recast segment table looks exactly like a correct one. The analyst then ticks and ties every cell by hand, which is the work the tool was meant to remove.

We built Kepler so the inputs arrive proven. Statements assemble from each filing’s own XBRL with no model in the path, as filed or as restated, on any date, and every historical cell links to the page, line or cell it came from. Extraction models we train read every page of 26M+ SEC filings and 40M+ IR documents and earnings call transcripts at ingest; in held-out tests they read tagged facts at the right scale 99.4% of the time (Stop Paying to Forget). Companies reshape their statements constantly: 194 of the 200 we studied reclassified a line in ways XBRL tags can’t follow, and our paper Reclassified to Conform recovers the comparable history. Where a filing doesn’t disclose a line, the cell stays empty and Kepler says why.

Calculated cells carry live formulas over cited inputs, so a margin, a multiple or an EV bridge opens to its formula and the source of each input. Our forward-deployed engineers, who built this layer before at Palantir, build your firm’s ontology, so “adjusted EBITDA” means your add-backs, peer sets are yours, and your template is saved once as a playbook and applied to every model. When a 10-Q lands, an agent adds the quarter, flags any line that moved against your estimate and shows both versions of anything restated. The workbook arrives in Excel, in the Kepler app, or inside Claude and ChatGPT, and engineers can pull the same data through our API.

Why it’s our pick for financial modeling

• Every historical cell cited. Click any number and land on the filing line it came from.
• As filed or as restated. Recasts and reclassifications caught, with both versions shown.
• Live formulas over cited inputs. Every calculation is auditable in the workbook.
• Your template and definitions. House tab structure, color code, fiscal calendar and adjusted EBITDA, every time.
• Models that update themselves. Agents roll the model forward the morning a filing lands.
• Reproducible. The same request returns the same model with the same sources.
• Excel and everywhere else. Excel, PowerPoint, Word, Claude, ChatGPT, Slack and the API.

Models analysts ask Kepler to build

• “Three-statement model for NVIDIA, FY2024 to FY2029, historicals from the 10-Ks and 10-Qs with every cell sourced, segment revenue broken out, projections on our template.”
• “Trading comps for US vertical software, EV/revenue and EV/EBITDA on each filer’s own adjusted EBITDA, every multiple traced to its line items.”
• “Roll the model forward for the 10-Q filed this morning and flag every line that differs from our estimate by more than 3%.”

“Kepler’s numbers match the filings. My team stopped re-checking every figure before it goes to a client.” (Vice President, Global Investment Bank)

Not the right fit if: you’re a student practicing model builds or an individual investor valuing your own portfolio.

Build your first model at app.kepler.ai, see how comps work in Kepler, or book a demo.

2. Claude in Excel (Claude for Financial Services)

Best for: analysts who want the strongest general model working directly in their workbooks, on data licenses the firm already holds.

Claude in Excel is generally available on paid Claude plans and can answer questions about any cell, update assumptions without breaking formulas and build models from scratch; context carries into PowerPoint and Word. Anthropic’s financial services repository adds a Model Builder agent for DCF, LBO, three-statement and comps work live in Excel, plus /dcf and /comps commands, and connectors bring in FactSet, S&P Capital IQ, PitchBook, LSEG, Daloopa and others through your own licenses. In Wall Street Prep’s test it placed a close second and won the sourcing and commenting subtask. Claude Opus 5.5 also holds the top Financial Modeling score on Vals AI’s Finance Agent benchmark, at 57.0%.

How it compares: excellent at structure, formulas and explanation, with historicals the model types from what it reads. Kepler runs inside Claude, so the same chat can return a workbook with every historical cell cited. See how to use Claude for financial analysis.

3. Shortcut

Best for: analysts who want a dedicated AI agent for building models in Excel.

Shortcut from Fundamental Research Labs runs as an Excel add-in, a web app, a Windows desktop agent (ShortcutXL) and a command-line tool. Its 2026 releases added skills, a vault for live-data API keys and, in July, Pivot 0.5, its first in-house model (Shortcut blog). Shortcut topped Wall Street Prep’s test at 5.9 out of 10 and won the formatting subtask; the same test found the top tools, Shortcut included, hallucinated significant portions of the historical data on the graded first attempt, though Wall Street Prep reports Shortcut’s second attempt returned almost no mistakes. Pricing is published: a free tier with 20 weekly credits, Pro at $100 a month billed annually with 1,000 monthly credits, and Teams at $320 a month plus $100 per seat (Shortcut pricing).

How it compares: the most model-focused agent on this list, and fast at a first draft. Sourcing the inputs stays with the analyst; upload the filings and tick the historicals.

4. Daloopa (Scout)

Best for: model builders who want every historical data point linked to the filing it came from.

Daloopa covers 6,000+ public companies with 14 years of history, each data point linked to its source document. Scout, its Excel agent, has moved from beta to general availability: it builds and updates models, compares companies and fills templates, and Daloopa says values are mapped to line items and periods “using fixed rules, rather than relying on an LLM.” Scout is included in the Core and Premium plans, and a read-only MCP server carries the same data into Claude, ChatGPT and other assistants.

How it compares: the closest to Kepler on source-linked historicals, as a data layer for public companies. Your firm’s private documents, definitions and agents live elsewhere. See Daloopa alternatives.

5. Microsoft Copilot in Excel

Best for: teams standardized on Microsoft 365 that want help with the mechanics of a workbook.

Copilot’s Agent Mode reached general availability in Excel for the web in December 2025 and on Windows in January 2026. Microsoft’s help now covers it as Copilot in Excel, with edit, plan and chat modes: it edits sheets, formulas and formatting, builds charts, PivotTables and dashboards, imports data from other workbooks, searches the web with citations and follows custom skills and workbook rules (Microsoft). It runs on OpenAI’s GPT models and Anthropic’s Claude models, with Claude enabled by admins for enterprise customers (Microsoft FAQ). The preview =COPILOT worksheet function was retired on September 14, 2026. In Wall Street Prep’s test it ranked third overall and won two subtasks, historical accuracy and flowthrough, the second largely because it built the simplest model.

What users say: “When I feed Copilot a PDF contract and ask it what is my monthly minimum I can charge this client and it tells me $1000 … when I ask it to cite where it finds that information … it points me to a paragraph that clearly says $1500.” (a Hacker News commenter testing Microsoft Copilot on a contract PDF, November 2025)

How it compares: strong on workbook mechanics, and it works on whatever data you bring. Finance-specific content comes from partners plugged into it, such as Rogo.

6. ChatGPT for Excel (and ChatGPT for Financial Services)

Best for: firms that work in ChatGPT, and teams that model in Google Sheets.

OpenAI opened ChatGPT for Excel as a beta in March 2026 and made it generally available to every plan in May, along with a version for Google Sheets. Analysts describe the model they want and ChatGPT writes it into the workbook with Excel’s own formulas; it traces how an assumption moves through the sheets, points to the exact cells behind each answer, and asks before it edits. Firms on the ChatGPT for Financial Services plan add included data from SEC filings, Quartr, Daloopa (delayed), Fiscal.ai and PitchBook Essentials, plus admin-published Excel templates for house-format models. Wall Street Prep’s February test predates the Excel add-in: it ran ChatGPT 5.2 outside Excel, where it placed fourth.

How it compares: a capable modeling assistant inside Excel and Sheets; the historicals come from data the model reads and types. Kepler runs inside ChatGPT too. See ChatGPT vs Claude for financial analysis.

7. Rogo

Best for: investment banks and PE deal teams that want models built from templates on licensed data.

Rogo’s native Excel add-in lets Felix, its agent, populate financials, build analysis, link and audit tabs and stress-test assumptions, and agents built in Rogo carry over into the spreadsheet (Rogo, May 2026). Its Agent Library includes DCFs and LBOs, public comps, precedent transactions, quality-of-earnings adjustments and quarterly earnings models, built by Rogo’s own finance team. Data comes from 51 integrations, including S&P Capital IQ, FactSet, LSEG, Daloopa, PitchBook and Preqin, and Rogo was a launch partner for Copilot in Excel built for finance. Sacra estimates about $3,300 per seat a year.

What users say: “It also hallucinates sometimes so you have to check anything related to searching for deal values, etc.” (an M&A analyst at a bank that uses Rogo, on Wall Street Oasis, 2025)

How it compares: built around the banker’s finished file. See Rogo vs Claude.

8. AlphaSense (Canalyst models)

Best for: equity analysts who want analyst-built, drivable company models alongside a research library.

AlphaSense still offers Canalyst models, which came with its 2024 Tegus acquisition: more than 4,200 KPI-rich, drivable models, plus standardized financials on 18,000+ public companies. AlphaSense for Excel and PowerPoint (July 14, 2026) edits models in plain language without breaking formulas, pulls cited financials and checks assumptions against research, and its SuperAnalyst agent can keep Excel models updated through Skills.

What users say: “There’s also less emphasis on structured financial models and datasets, which makes it not ideal if your work relies heavily on detailed financial modeling.” (a G2 reviewer, June 2026)

How it compares: strong if you want a pre-built model on a covered name and the research library around it. Your own template and definitions stay a manual adaptation.

9. Model ML

Best for: banking and advisory teams producing models, decks and documents together.

Model ML describes itself as an agent harness for financial services that produces client-ready PowerPoint, Excel and Word deliverables, across models from OpenAI, Anthropic and Google. Its 2026 releases include Operator, an agent for Office work, plugins for Word, Excel, PowerPoint and Teams, and a September update that shows before-and-after cell changes in Excel. Its forward-deployed engineers build custom agents for each firm. Pricing isn’t published.

How it compares: focused on the deliverable package more than the model’s source data.

How to test an AI modeling tool in an hour

1. Rebuild a model you know cold. Same company, same template, same periods. Differences show up fast.
2. Tick and tie 20 historical cells. Open the filing for each one and check the figure, the period, the unit and the sign. The errors in Wall Street Prep’s test hid below the subtotals, so check every line item as well as the totals.
3. Pick a restated quarter. Ask for the year as filed and as restated, and see whether the tool notices the difference. Our guide to what XBRL is explains why this breaks so often.
4. Look for plugs. Trace how the balance sheet balances and whether projections are formulas or typed numbers.
5. Run it twice. The same request should produce the same historicals.
6. Roll it forward. Give it the next 10-Q and see what it updates, what it misses and how it shows changes.

For comps specifically, see how to build trading comps with AI; for the wider analyst toolkit, our list of the best AI tools for financial analysts; and for checking any figure, how to check AI-generated numbers.

Build a model on your own coverage at app.kepler.ai, or book a demo and we’ll rebuild one of your existing models with every cell cited.

Frequently asked questions

What is the best AI tool for financial modeling?

For models a firm will defend, our pick is Kepler. Historicals assemble from the filings' own XBRL with every cell cited to the exact page, line or cell, as filed or as restated; calculated cells carry live formulas over cited inputs; the model follows your firm's template and definitions; and agents update it when new results land, so the same model rebuilt next quarter ties out.

Can AI build a three-statement model in Excel?

Yes, with review. Claude in Excel, Copilot in Excel, ChatGPT for Excel and Shortcut all build integrated models with live formulas. In Wall Street Prep's 2026 test the best of four tools scored 5.9 out of 10, below the 6.4 for a low-rated junior analyst, and none handled circularity. In Kepler, historicals build from the filings' XBRL with every cell cited, so review time goes to the assumptions.

Is AI accurate enough for financial modeling?

Accurate enough to start a model, and not yet reliable for inputs the model types itself. In Wall Street Prep's test, two of the four tools hallucinated significant portions of the historical data while subtotals still tied, and the best Financial Modeling score on Vals AI's Finance Agent benchmark is 57.0%. The fix is historicals pulled by code from the filing with a source on every cell, which is how Kepler builds them.

Should historicals in an AI-built model be hardcoded or linked?

Hardcode historicals, but give every one a source: the filing, the page and the line, with its unit and period. Projections should be formulas only, driven by labeled assumptions. An AI tool that types historicals without sources leaves the analyst to tick and tie every cell, so ask for a source column or a tool that cites each cell automatically.

How do I keep an AI-built model up to date after earnings?

Roll the model forward from the new 10-Q or 10-K rather than the press release, check whether the company recast any prior period, and compare every updated line against your estimate. In Kepler, an agent does this when the filing lands: it adds the new quarter from the filing's XBRL, flags restated lines with both versions shown, and cites every cell.