Best AI for Excel and Financial Modelling, Ranked by Evidence Instead of Demos

The strongest AI for Excel completes 34.89% of tasks on the hardest public spreadsheet benchmark, and 12.00% of the debugging tasks inside it. On a different benchmark that grades investment banking models built from scratch, the leaders sit near 73%. Both numbers are real. The gap between them is the whole buying decision, and almost nobody selling you a spreadsheet agent will explain it.

This post is written for the head of finance or delivery who has to make a team produce consistent models. It is not written for the analyst chasing one clever formula. The manager is the person who lives with whatever the tool gets wrong six weeks later.

"AI for Excel" is three different products with one name

Most comparison posts rank these tools against each other as though they compete. They mostly do not. Three distinct product shapes have been collapsed into one search term, and the benchmarks that flatter one shape punish another.

Formula help sits inside the cell

The oldest shape writes or explains a single formula. It answers questions like how to nest an INDEX and MATCH, or why a lookup returns an error. Scope is one cell or one range, and correctness is easy to check by eye.

This is the category most people actually use daily, and it is the least interesting one commercially, because the marginal value over a search engine is small and shrinking.

In-app agents work across the open workbook

Microsoft's Agent Mode in Excel and Anthropic's Claude for Excel both sit in a sidebar, read the workbook, and make multi-step edits. Microsoft describes Agent Mode as planning, executing and then checking its own work before iterating.

Scope here is the file you have open. The failure mode changes with the scope: the agent stops being wrong about a formula and starts being wrong about which cells to touch.

Model builders produce the workbook itself

The third shape takes a brief and source documents, then returns a finished model. Specialist vendors such as Shortcut aim squarely at this, and Anthropic ships prebuilt skills for discounted cash flow models and comparable company analysis in its financial services product.

My view is that these three shapes need three different purchase decisions and three different tests. Buying a model builder on the strength of formula-help reviews is how finance teams end up with a tool nobody trusts by the second quarter.

What each benchmark actually measures

Four public benchmarks now cover this ground, and they disagree because they are not measuring the same job. Reading them as one leaderboard produces nonsense.

Scores are not comparable across rows. Each benchmark uses its own task set, its own grader and its own definition of a pass. The original SpreadsheetBench figure is Microsoft's own evaluation of its own product, which is a different evidentiary class from the three academic and third-party runs below it.

How the two grading philosophies differ

The Excel Modeling Benchmark covers seven model types, including LBO, discounted cash flow and merger models. It notes that each task takes a human expert at least five hours from scratch. Scoring in scratch mode gives partial credit across three dimensions rather than demanding an exact match.

That is the right design for the job it describes. It also means a 73.7% score is a weighted average of three very different things, and the components matter far more than the headline.

SpreadsheetBench 2 asks for a completed task inside an existing business workbook, and grades at task level. Its authors report that cell-level modification scores run well ahead of task-level accuracy. Put plainly, the agents change roughly the right cells without holding the model together.

Both grading philosophies are defensible. What is not defensible is quoting the generous one in a sales deck and the strict one nowhere.

Accuracy collapses exactly where the sheets connect

The single most useful finding in this year's research is not the overall score. It is the shape of the breakdown by task type, which is stable across benchmarks and points at the same weakness every time.

Debugging is the worst-served task in the category

On SpreadsheetBench 2, the strongest model scores 62.50% on visualisation, 52.58% on template completion, 34.00% on financial modelling and 12.00% on debugging. The debugging tasks are the biggest files in the set, averaging 21.1 worksheets.

That ordering is inconvenient, because debugging is what finance teams have most of. Very few businesses need a new model every week. Almost all of them have an inherited workbook that nobody fully understands and that somebody has to check.

Cell edits are not the same thing as a correct model

The benchmark authors name two dominant failure modes: insufficient inspection of the spreadsheet, and selecting the wrong target cells. Both are localisation failures rather than reasoning failures, and both get worse as the sheet count rises.

The Excel Modeling Benchmark finds the same shape from the other direction. Leading models pass 88% of formula checks and 80% of presentation checks, against 63% of numerical checks. The wiring is usually right and the answer is often wrong.

The same leaderboard reports the highest scores on data room summaries. The lowest are on LBO and discounted cash flow models, where early errors cascade through interdependent calculations. WorkstreamBench draws its tasks from the Financial Modeling World Cup, ModelOff and the Wall Street Prep curriculum. It reports performance degrading sharply above difficulty level 3, with agents hardcoding values into cells where formulas belong.

What the evidence supports, tool by tool

Here is what each option can defensibly claim, and what it cannot. I have kept vendor marketing out of the middle column on purpose.

Two rows in this table rest on vendor-run evaluations and are labelled as such. Neither is dismissed on that basis, because vendors often test first. They are simply not the same class of evidence as a peer-reviewed benchmark with a published task set.

How to read a spreadsheet benchmark run by the vendor

Most of the numbers circulating for this category come from the company selling the tool. That is normal in a young market, and it is not automatically disqualifying. It does need three specific checks.

First, ask who chose the tasks. Microsoft evaluated Agent Mode on the full published SpreadsheetBench set, which is the honest version. A vendor testing on eleven tasks it selected itself is a different exercise. The March 2026 comparison published by the chief executive of GPT for Work is exactly that shape.

Second, ask what a pass means. Partial credit across formula, numerical and presentation checks produces a much higher number than exact task-level matching on the same work. Neither is wrong. Comparing them is.

Third, ask when the benchmark was built relative to the model. The original SpreadsheetBench predates every model now marketed against it. That is a good part of why scores on it look healthier than scores on its June 2026 successor. The same pattern shows up across categories, and it is why benchmark rank, purchase rank and usage rank rarely agree.

Where this argument is weakest

The case above reads as a warning about AI accuracy in spreadsheets. There are three good reasons to hold it loosely.

Humans are not the clean baseline this framing implies

The human expert score on the original SpreadsheetBench was 71.3%, not 100%. The wider research is harsher still. Ray Panko's 2016 review of the spreadsheet error literature concludes that errors are rare per cell. It also finds that at least one incorrect bottom-line value is very likely in any large model. Those errors are extremely difficult to detect, and developers and companies are highly overconfident about accuracy.

So the comparison is not machine error against human perfection. It is one error profile against another. I disagree with the common framing that this is a race to human parity. The human baseline in spreadsheet work is genuinely poor, and it has been measured for decades.

Benchmarks lag the releases by months

SpreadsheetBench 2 was published on 29 June 2026 and tested the models available then. The Excel Modeling Benchmark leaderboard already carries model versions that postdate it. Any snapshot of this category is a photograph of a moving object, and the honest reading is directional rather than final.

The graders themselves are imperfect

The SpreadsheetBench 2 authors state their own limitations plainly. They report no error bars because of inference costs. They note that exact-match evaluation can undercount solutions that are correct but written differently. They also flag that using a vision model as a judge for chart tasks introduces noise. WorkstreamBench uses a language model as judge, validated at 0.92 accuracy against 408 expert annotations. That is good, and it is not the same as perfect.

Test it on your own three-statement model

No public benchmark can tell you how these tools behave on your workbook, with your naming conventions and your inherited quirks. The good news is that a usable internal test takes about a day, and the design is not complicated.

Take one live three-statement model that you already trust. Save a copy. Break it in five specific, documented ways: a hardcode over a formula, a sign flip on a cash flow line, an off-by-one in a SUM range, a stale prior-period link, and a circular reference. Then ask each tool to find and fix the errors without telling it how many there are.

Record the results as counts, not impressions. Five planted errors, three found, one false positive is a sentence you can put in a renewal discussion. "It felt quite good" is not. The general method for turning a vague vendor claim into a number you own is set out in the guide to building an evaluation suite for your own use case. The wider readiness checks belong in the production readiness tests every pilot should clear.

I would not buy any of these tools on a scratch-build score. I would spend the first month using them to check models rather than to write them. A wrong answer you catch costs nothing. A wrong answer you ship costs a board meeting. That is the same reasoning behind designing the human review step before the automation, not after it.

The licence maths has quietly changed

Two pricing facts are worth checking before you budget, because the widely repeated numbers are out of date.

Microsoft 365 Copilot for business now lists at $21.00 per user per month paid yearly, with promotional pricing at $18.00 through 31 December 2026. That sits on top of a qualifying Microsoft 365 licence. The $30 figure still quoted in most comparison posts is stale. Agent Mode in Excel launched inside the Frontier programme in September 2025 for Copilot-licensed customers and Microsoft 365 Personal, Family and Premium subscribers.

Anthropic's release notes list Claude for Excel as a beta from November 2025 for Max, Team and Enterprise users. A February 2026 update moved it to Opus 4.6, with native pivot table and conditional formatting support. That matters for procurement: the tool sits inside an existing Claude subscription rather than as a separate Excel add-on line.

The cost that nobody prices is review time. If a tool returns a model where 63% of the numbers are right, somebody senior checks the other 37%, and that person is more expensive than the licence. Skipping that arithmetic is the most common route to an AI deployment that shows a negative return.

Frequently asked questions

Where to start this week

Break a copy of your own model on purpose. Five planted errors, documented, then run every tool on your shortlist against the same broken file and write down what each one found. That single afternoon will tell you more than every leaderboard in this post, because it uses your conventions and your inherited quirks.

Then check what you are already paying. Pull your Microsoft 365 and Claude line items and confirm which plan each Excel feature actually needs, because at least one of them is probably already included in a subscription you hold. If a vendor will not let you run your own workbook through the tool before you sign, that is your answer, and the reasoning generalises to every vendor case study you cannot verify.

Related reading