ChatGPT for Finance in 2026: 7 Jobs It Does Well, and the Numbers It Gets Wrong

ChatGPT does not perform arithmetic when it answers you in the chat window. It predicts the next piece of text, and text that looks like a calculation is still text: the digits it produces are the digits that fit the pattern, not the output of an operation performed on your numbers. Sometimes the pattern and the correct answer coincide. Sometimes they do not, and nothing in the response tells you which case you are in. The exception is Advanced Data Analysis, available from Plus upward, which writes actual Python and executes it against a file you uploaded. That is real computation, with a real result, on data the model can actually see.

Hold onto that distinction and everything else falls into place. Language, structure and pattern recognition are where the tool is strong, which covers a surprising amount of a finance job: the commentary behind a variance, the standardization of a filthy vendor export, the logic of a forecast, the memo that turns four numbers into an argument, the questions an investor is going to ask before they ask them. Arithmetic on a table pasted into a chat, or any figure recalled rather than supplied, is where it stops being useful and starts being a risk with your name on the document.

Seven jobs follow, ordered the way the finance calendar actually runs: close first, then planning, then the people you have to explain it all to. The plan question and the failure modes come after, because both make more sense once you can see what you would be typing.

During the close

Close work is where the ratio tips furthest in the tool's favor, because most of what makes close slow is not calculation. It is turning numbers you already trust into sentences other people understand, and wrestling exports into a shape that can be reconciled.

Variance commentary

Variance write-ups eat FP&A time mainly because writing clear, non-defensive prose about a miss is genuinely hard. Hand it clean numbers and the drivers you already know, and it drafts. Hand it a table and the word "analyze" and it invents.

I'm attaching our budget-vs-actual for Q2 by department. For each line
item where the variance is more than 8% (favorable or unfavorable), write
a 2-sentence explanation a non-finance department head would understand,
using the driver I list next to each line. Flag any line where I did not
give you a driver, so I know to fill it in myself. Do not estimate or
invent numbers that aren't in the data I gave you.

[paste budget/actual/variance/driver table]

The "flag what I didn't explain" instruction is the whole safety mechanism. Without it, an unexplained variance comes back with a plausible, fluent, entirely fabricated cause attached, and plausible causes are exactly the ones that survive review. With it, you get a list of the lines you still owe an answer on, which is genuinely useful output in its own right.

Cleaning a messy export

One of the highest-value, lowest-risk jobs in the set, because standardizing vendor names and bucketing transactions is text pattern matching rather than arithmetic. This is Advanced Data Analysis territory: upload the file so real Python runs against it.

I'm uploading a CSV of 400 expense transactions with inconsistent vendor
names (e.g. "AWS", "Amazon Web Svcs", "AMAZON WEB SERVICES INC" all mean
the same vendor). Standardize vendor names into a clean list, categorize
each transaction into: Software, Travel, Payroll Services, Professional
Services, Office, Other. Output a pivot of total spend by clean vendor
and category, and show me the mapping from messy name to clean name so I
can spot-check it.

Always demand the mapping, not just the cleaned output. Fifteen rows checked against the mapping catches the misclassifications before they reach a spend report, and the mapping itself is reusable next month. The same first-pass reading discipline pays off on the contract sitting behind an expense line you cannot explain, which at close is usually an auto-renewal nobody diaried.

I'm pasting the pricing and termination sections of a vendor contract.
Flag: auto-renewal terms and required cancellation notice, any price
increase cap (or lack of one), minimum commitment length, and any
clause letting the vendor change terms unilaterally. List each flag
with the exact clause quoted, and rate overall vendor-friendliness as
low, medium, or high risk to us.

[paste contract text]

Requiring the exact clause quoted rather than paraphrased is what makes that output checkable, and it is triage for deciding which contracts need a real legal read, never a replacement for one.

Debugging a spreadsheet formula

Everyone in finance has stared at a #REF! or a number that is obviously wrong and not obviously why. The model reads formula logic quickly and is decent at locating the actual reference or logic error rather than the symptom.

This formula returns #DIV/0! for some rows but not others:
=IFERROR((C2-B2)/B2,"") copied down through row 500. Explain what causes
this, and rewrite it to show "N/A" when B2 is zero but calculate
normally otherwise. Also check whether IFERROR here would silently hide
a different kind of error I should know about.

Include that last line every time. Error-wrapping is how analysts accidentally mask genuine problems for years: the model of the business quietly stops reflecting reality, the cells stay clean, and nobody notices until someone rebuilds the file from scratch. Asking what the wrapper is hiding turns a formatting fix into an actual review.

In planning

Planning work is structural, and structure is what the model is good at. The numbers stay in your spreadsheet, where you can audit them. What you are extracting from the chat is the skeleton and, more importantly, the assumptions made explicit.

Scaffolding a forecast and its assumptions

It will not build your three-statement model, and you should not want it to. What it produces fast is a defensible first-pass structure with the mechanics written out, which you then populate yourself.

Help me build a 12-month revenue forecast structure for a B2B SaaS
company: current MRR $180,000, net revenue retention 104%, average new
logo ACV $9,600, roughly 6 new deals closed per month. Show the
month-by-month formula logic (not final numbers) for how MRR compounds
from retention plus new bookings. Add a churn-adjusted version and an
aggressive version assuming NRR reaches 110% by month 9. Output as a
table with the formulas written out in plain language next to each row.

Asking for formula logic rather than final numbers is deliberate, and it sidesteps the arithmetic problem entirely: plain-language mechanics are a language task, compounding twelve months of MRR in a chat window is not. Take the logic into a spreadsheet, build it there, and audit every cell yourself. A useful follow-up is to ask the model to list every assumption its structure relies on and mark which ones it inferred rather than took from you, since inferred assumptions are the ones that end up quietly baked into a board number.

Structuring scenario and sensitivity work

Board decks and fundraising conversations both need a base, upside and downside view eventually. Generate the grid fast, then spend the time you saved interrogating the assumptions rather than formatting cells.

Build a sensitivity table for our EBITDA forecast. Base case: revenue
$8M, COGS 32%, opex $4.2M. Show EBITDA across 3 revenue scenarios
(-10%, base, +15%) crossed with 2 opex scenarios (flat, +8% from a
planned hire). Present as a 3x2 grid with EBITDA and EBITDA margin per
cell, plus one sentence per scenario on what would need to be true for
it to happen.

The "what would need to be true" clause is the reason to run this at all. It drags the narrative logic behind each scenario into the open, where someone can disagree with it, instead of leaving it implicit in a percentage nobody remembers choosing. Recompute the cell values yourself, or upload the base numbers as a file so Advanced Data Analysis does the maths properly. A downside case with an arithmetic error in it is worse than no downside case, because people plan against it.

For the audience above you

The last two jobs are pure communication, and this is where the tool is at its least dangerous and most immediately valuable. Nothing here computes anything. It converts numbers you have already verified into the register your board, your investors or your CEO actually read in.

A board memo from your own numbers

Board writing has a specific register: direct, numbers-forward, no cushioning. Give it your figures and the shape of the argument, and ban the hedges explicitly, because hedging is the default it reaches for.

Draft a board memo section on Q3 performance. Structure: headline number
first, then 3 supporting points, then one paragraph on what we're doing
about the pipeline miss. Facts to use, don't add others: revenue $2.4M
vs $2.6M plan (8% miss); gross margin 71%, up from 68% last quarter;
pipeline generated down 15% QoQ, tied to the SDR team being down 2
headcount for 6 weeks; cash runway 14 months at current burn. Tone:
direct, no hedging, no "we are pleased to report." Under 300 words.

The same technique compresses further when the board pack needs a front-page summary rather than a section. Boards do not want a dashboard, they want to know which number to worry about, so name the jargon you refuse to see.

This month's SaaS metrics: MRR $310K (+4% MoM), churn 1.8% (target
under 2%), CAC payback 14 months (up from 11), NRR 108%. Write a
3-sentence "state of the business" summary for the board that leads
with the one metric that needs attention, in plain language, no jargon
like "north star metric" or "flywheel."

Naming the specific phrases to ban works considerably better than asking for "plain language", which the model will agree to and then ignore. And the "don't add others" constraint on the facts is not optional: unconstrained, it fills narrative gaps with confident connective sentences that read like findings and are not.

Earnings and investor Q&A prep

Before a call, the valuable exercise is stress-testing what you will be asked, in private, while there is still time to have a better answer.

Based on these results (revenue up 6% QoQ but gross margin down 300bps,
churn flat, one customer at 22% of revenue), generate 10 tough questions
an investor might ask next call, ranked by likelihood. For the 3 most
likely, draft a direct, honest answer using only the numbers I've given
you.

Treat the drafted answers as raw material for your own thinking rather than a script. The output that matters is the ranked question list, because the uncomfortable question you had not anticipated is worth more than three polished paragraphs. The preparation extends to reading the documents behind the call, whether that is a competitor's filing or your own, as long as you paste the text rather than asking the model what a filing says.

I'm pasting the MD&A and Risk Factors sections of [Company]'s most recent
10-K. Summarize as: 1) revenue and margin trend over 3 years with actual
numbers, 2) the risk factors that are new or reworded versus standard
boilerplate, 3) any mention of customer concentration, debt covenants, or
going concern language, 4) one paragraph on tone versus a typical filing.
Quote the exact sentence for anything in points 2 and 3, don't paraphrase.

[paste filing text]

Forcing direct quotes on the risk section stops it from softening or sharpening disclosed language beyond what the company actually wrote, which is the specific way filing summaries go wrong.

The plan question is really a data-controls question

Free works for trying prompts on made-up numbers. It is too thin for the work above: tight message caps, limited file handling, and no Advanced Data Analysis, which means no real computation on a file you upload.

Plus ($20/month) is the realistic floor for a finance seat, principally because Advanced Data Analysis is what makes the messy-export and file-based work genuine rather than theoretical. Business or Team (roughly $20 to $30/user/month) starts to matter the moment a second person is putting company figures into prompts. Pro ($200/month) buys higher ceilings and the strongest reasoning models, which is a real benefit if you are in the tool daily and unnecessary for most teams.

What actually separates those tiers, from a finance perspective, is not features. It is the default answer to "what happens to this data afterwards". On personal accounts (Free and Plus), conversations may be used to improve future models unless you go into Data Controls and switch it off, and a default that depends on every individual remembering a setting is not a control you can attest to. Business, Team and Enterprise invert that: training exclusion is the default state, with admin-level retention settings and, at Enterprise, a signed DPA and centralised logs.

Read that as a governance question rather than a procurement one. Retention policy determines what a regulator or an acquirer could theoretically reach, and no tier changes the underlying fact that you have transmitted information to an external processor. So decide in advance which categories never leave the building in any account: unreleased results, cap tables, payroll detail, an unsigned term sheet, anything that would be material non-public information if it moved. Then write that list down somewhere your team can find it at 23:00 the night before a board deck is due, which is precisely when nobody goes looking for the policy. Where placeholder entity names and rounded figures would serve the analysis just as well, use them by default. (Finpresso runs a daily read on AI in finance if this is a beat you follow.)

What it should never be trusted with

Arithmetic on a pasted table. Numbers typed into the chat window and totalled there are pattern completion, not computation. The totals will often be right, which is what makes this dangerous, because a spot-check that passes twice buys unearned confidence in the third one. If a figure needs to be correct, upload the file so Advanced Data Analysis runs Python against it, or compute it in the spreadsheet and use the model only for the words around it.

Accounting standards recalled from memory. Ask how a specific transaction should be treated under a given standard and you will get a fluent, structured, authoritative-sounding answer that may be a version of the guidance from several revisions ago, or a blend of two jurisdictions' treatments. Standards get amended, effective dates stagger, and the model has a training cutoff it will not mention. Paste the relevant text of the standard or your firm's policy and ask it to work from that, exactly as you would with a filing.

Market and competitor figures. A competitor's revenue, a historical rate, a comparable transaction multiple, a funding round size: asked to recall these, the model can produce a number that is well formatted, plausibly scaled and simply wrong, with no hedge attached. It has no live data unless browsing is enabled, and even then anything time-sensitive (a price, an exchange rate, a just-announced deal) is a starting point to confirm against a primary source, not a citation.

Model logic that looks right. The most expensive failure is not a wrong number, it is a correctly formatted structure with a flawed mechanic buried in it: a retention assumption applied to the wrong base, a growth rate compounded on a figure that already includes it, a scenario that quietly double-counts a headcount cost. It renders beautifully. It reconciles internally. Trace the logic line by line the first time you use any structure it produced, because the presentation quality of the output carries no information about whether the mechanics are sound.

What these tools actually cost

We price every tool we review, so this is measured rather than estimated. Across 429 tools, 293 publish a price and 33% offer a free tier. Among finance tools, the median entry plan is $37 a month, which runs above the $24 median across every category we price.

The spread matters more than the median. Half of the finance tools sit between $25 and $149, and the range runs from $15 to $200. A quoted "starting at" price near the bottom of that range usually means per-seat add-ons land on top of it.

Price point	Finance tools	All tools
Cheapest paid plan	$15	$1
Lower quartile	$25	$10
Median	$37	$24
Upper quartile	$149	$49
Most expensive	$200	$990
Tools measured	16	293

FAQ

Can output from ChatGPT go into a board pack?

The words can, once you have checked them. The numbers should come from your own model or ledger, not from the chat. In practice the split is clean: the model drafts the commentary, the summary and the framing, and every figure inside it is one you produced and verified elsewhere. Read the draft specifically hunting for claims you never made, because gap-filling connective sentences are its most reliable habit and they read exactly like conclusions.

Does using ChatGPT break my audit trail?

Not by itself, but it produces nothing an auditor can follow. There is no lineage from a chat response back to a source system, and a conversation is not evidence of how a number was derived. Keep the derivation where it has always lived, in the model, the ledger or the reconciliation, and treat the chat as a drafting surface. On Enterprise you get centralised conversation logs, which helps with oversight of what staff are sending out, and is a different thing from an audit trail for a figure.

Can it actually do the arithmetic?

Yes when it runs Advanced Data Analysis against an uploaded file, because that executes real Python and returns a real result. No, or not dependably, when it produces a total conversationally in the chat window, because that is text prediction wearing the costume of a calculation. The practical test is simple: if you did not upload the data and the model did not run code, treat the number as a draft.

Excel Copilot or ChatGPT?

They sit at different points in the workflow. A spreadsheet copilot lives inside the file, knows the cells and the ranges, and can act on the actual grid. ChatGPT is stronger on unstructured work around the data: the narrative behind a variance, a memo draft, a filing summary, a messy export restructured into something reconcilable, the investor question you had not thought of. Plenty of teams run both, with the copilot inside the model file and the chat window handling everything that ends up as prose.

Is it safe to upload company financials?

It depends on the account and the sensitivity. Personal accounts may use conversations for training unless you disable that under Data Controls, while Business, Team and Enterprise exclude your data from training by default. Neither changes the fact that the data has left your environment for a third-party processor, so treat material non-public information as a decision requiring sign-off in advance rather than an analyst's judgment call, and redact identifying details wherever the analysis does not depend on them.

How do I decide what is too sensitive to paste?

Use a rule your team can apply without thinking, since the risky moment is always the one where nobody is thinking. A workable version: if the information would move a valuation, identify a specific person, or embarrass you in a disclosure, it does not go into a chat window in any tier without prior approval. Everything else goes in with entity names replaced and figures rounded or indexed where precision is not required for the task. The point is to make the safe path the fast path, so it survives a deadline.