What should a finance team actually use AI for?
Every finance leader we speak to has already tried ChatGPT or Claude for analysis. Most stopped within a week, and almost none could say precisely why. This guide explains the mechanism behind that failure, sets out the twelve criteria that separate approaches that work from approaches that look like they work, and scores all six honestly, including where general-purpose chat and a human analyst beat everything else.
Key takeaways
- Accuracy is an architecture question, not a model question. A language model generates arithmetic as text rather than computing it. The only reliable fix is to move the calculation out of the model and into code you can inspect.
- Context beats capability. The reason a general-purpose model gets your numbers wrong is rarely reasoning. It is that it does not know your fiscal calendar, your metric definitions, or the two customers everyone internally excludes.
- Verifiability decides whether anything gets used. If checking a number means rebuilding the analysis, the tool has saved nothing, and a confident wrong number is more dangerous than an obviously broken one.
- The categories solve different problems. An FP&A platform fixes consolidation and reporting. Agentic analysis fixes the unplanned question. Buying the second when you needed the first is the most expensive mistake in this market.
- Nothing here replaces judgement. A person who knows the business is still the best answer for genuinely novel strategic work. The realistic goal is removing the assembly, not the analyst.
Why ChatGPT and Claude fail at finance, specifically
Both are excellent at language and unreliable at arithmetic, for a mechanical reason. A language model predicts the next token. When you ask it to total a column, it is not running an addition , it is predicting what the answer probably looks like. Most of the time that prediction is right. Occasionally it is close and wrong, and you can watch it happen: ask for a table with a total row and the total will sometimes fail to equal the rows directly above it, in the same response, with nothing indicating a problem.
The deeper failure is context. Ask a general-purpose model for your net revenue retention and it will produce a defensible number using a standard formula. Whether it is your number depends on things it cannot know: that your fiscal year starts in April, that you classify customer success in COGS, that October included a one-time migration fee everyone excludes. It will not flag the gap. Faced with ambiguity it resolves it and continues.
In most domains a wrong AI answer announces itself. Code fails to compile, a summary contradicts the document. Financial output has none of those tells: it arrives formatted, internally consistent, and unhedged. That makes a wrong AI number more likely to reach a board than a wrong hand-built one, because the analyst who built it knows which assumption is shaky and the model expresses no uncertainty at all.
The question that settles it in a demo. Ask any vendor: “when this tool produces a number, where does the arithmetic physically execute?” If the answer is “the model”, you are buying predicted arithmetic. If it is “generated code running in a sandbox, and here is the code”, you are buying computed arithmetic. Everything else in this guide matters less than that.
The 12 criteria that separate approaches that work
Each criterion below carries the questions to ask a vendor and the red flag that should stop the conversation. They apply equally to a $200,000 platform and a free chat window.
Who performs the arithmetic
A language model predicts text token by token. Asked to sum a column it predicts what the total probably looks like rather than computing it, so a total can fail to equal the rows above it with nothing in the output indicating a problem. This is the single most important question to ask, because every other quality is worthless if the numbers are wrong. OpenAI's own guidance is to route any computation involving maths through its code interpreter rather than answering directly, because executing a calculation is more reliable than generating one: the vendors do not dispute this.
Ask vendors
- Does a language model generate the numbers directly, or does it write code that executes?
- Can I see the code or query that produced this figure?
- What happens if the model and the calculation disagree?
Red flag. Any vendor who answers "our model is very accurate" rather than describing where the calculation physically runs. Accuracy is a property of architecture here, not of model quality.
Company-specific context
Your fiscal year may start in April. Customer success may sit in COGS. Two accounts may be excluded from cohort analysis for a reason agreed years ago. None of this is written down, and a tool that does not hold it produces answers that are correct in general and wrong for you.
Ask vendors
- How does the system learn our metric definitions and fiscal calendar?
- Where are those rules stored, and who can change them?
- If I ask the same question in six months, will it apply the same definitions?
Red flag. A tool that requires you to restate your definitions in every prompt. That is not context, it is a workaround, and it fails the moment someone forgets.
Verifiability
Finance output that cannot be checked cannot be used. If verifying a number means rebuilding the analysis yourself, the tool has saved you nothing, and a wrong number that arrives formatted and confident is more likely to reach a board than a hand-built one, because the analyst who built it knows which assumption is shaky.
Ask vendors
- Show me the formula, the query, and the source rows behind any number in this demo.
- How many clicks does that take?
- Can I export the calculation logic to Excel or SQL?
Red flag. A demo where the presenter cannot reach the underlying rows within two clicks. If they cannot do it on their own data, you will not do it on yours.
Data access and freshness
An analysis is only as current as the data underneath it. Pasting an export into a chat window means the answer is a snapshot that goes stale immediately, and it caps the size of the question you can ask at whatever fits in a context window.
Ask vendors
- Does it connect directly to our ERP and CRM, or do we paste exports?
- How often does data refresh, and is that push or pull?
- What is the largest dataset it can reason over in one question?
Red flag. Copy-paste as the only ingestion path. It is fine for a one-off and unworkable as a monthly process.
Write access and blast radius
A system with write access to your ledger can corrupt your system of record. Read-only access removes that entire category of risk, and there is rarely a good reason for an analysis tool to need more.
Ask vendors
- Is access read-only? Can you prove it at the connection level rather than by policy?
- What is the blast radius if the model produces something unexpected?
- Who can grant write access, and is it logged?
Red flag. Write access requested "for convenience" during setup. Convenience is not worth the ability to modify your books.
Handles unplanned questions
Templated reporting is a solved problem. The work that actually consumes a finance team is the question nobody predicted, arising from something that just happened. A tool that only answers pre-built questions leaves the expensive part untouched.
Ask vendors
- Ask it a question you did not brief them on, using their demo data.
- How long does a genuinely novel question take end to end?
- What proportion of questions require someone to build something first?
Red flag. A demo that only follows a scripted path. Insist on going off-script; the difference between a report builder and an analyst shows up immediately.
Audit trail
Finance functions get audited, and boards ask how a number was produced months after the fact. A system that cannot reconstruct what it did, on what data, at what time, creates a documentation gap you will have to fill manually.
Ask vendors
- Is every query and result logged with a timestamp and the data version?
- Can I reproduce an analysis from three months ago exactly?
- Is the log exportable if we leave?
Red flag. Chat history offered as an audit trail. A transcript is not a lineage record: it shows what was said, not what was computed against which data.
Security and data handling
Financial data is among the most sensitive a company holds. Consumer AI tiers have historically differed from enterprise agreements on whether inputs may be used for training, and that distinction matters before anything leaves your network.
Ask vendors
- Is our data used for model training? Get this in the contract, not the sales call.
- Where is data stored and processed, and in which jurisdiction?
- Do you sign a DPA? What is your incident response commitment?
Red flag. Vague answers about training data, or a policy that differs between the tier being demoed and the tier being sold.
Time to first useful answer
Every week of setup is a week still doing the work manually. But speed alone is misleading: a tool that works in five minutes because it has no access to your data will still be useless in month three.
Ask vendors
- What is the realistic time to a genuinely useful answer on our data?
- What do we have to do, and who from our side is needed?
- What breaks when our chart of accounts changes?
Red flag. An implementation timeline quoted without naming what is in scope. Custom connectors, data cleanup, and definition capture all add time and are routinely excluded from the quote.
Total cost, honestly counted
The licence is rarely the largest cost in year one. Implementation, internal time, and the analyst hours still being spent afterwards all belong in the comparison, as does the cost of the status quo, which is usually invisible because it sits inside salaries already being paid.
Ask vendors
- What is the all-in year-one cost including implementation?
- How much of our current analyst time does this actually remove?
- What does the cost look like at twice our current volume?
Red flag. Pricing that cannot be discussed until late in the process. Information asymmetry late in a negotiation protects the vendor, not you.
Behaviour when it does not know
The correct response to an unanswerable question is to say so. A system that always produces an answer will sometimes invent one, and in finance an invented answer is indistinguishable from a real one until someone acts on it.
Ask vendors
- Ask it something your data genuinely cannot answer. What does it do?
- How does it behave when a month is missing from the data?
- Does it distinguish between an assumption and a finding?
Red flag. A system that silently interpolates a missing period. This is the failure mode you are least likely to catch and most likely to be harmed by.
Fit for mid-market reality
Tools built for enterprise finance assume a data team, a analytics engineer, and a multi-quarter implementation. Tools built for very small businesses assume one entity and simple books. A $1M–$50M company sitting between the two gets a bad fit from either direction.
Ask vendors
- Show me a customer at our revenue and headcount, not your largest logo.
- What does this require from us that we do not currently have?
- Who operates this day to day once you leave?
Red flag. Every reference customer is materially larger than you. It usually means the product assumes resources you do not have.
Six ways finance teams are actually doing this
Before comparing products, understand the approach behind them. The architecture sets a ceiling on accuracy, context, and verifiability regardless of how the marketing describes it.
General-purpose AI chat
You paste data into a chat window and ask a question in plain language. The model reasons over what fits in its context and answers in prose.
Examples. ChatGPT, Claude, Gemini, Copilot
Genuine strength
Immediate, free or near-free, genuinely excellent at explaining concepts, structuring an approach, drafting commentary, and reviewing your logic. If you are not yet sure what analysis you need, this is the fastest way to find out.
Structural ceiling
The model generates numbers as text rather than computing them, has no durable knowledge of your business rules, and cannot reach your systems. Every session starts from zero, and verification means rebuilding the analysis yourself.
AI spreadsheet assistants
An AI layer inside or alongside Excel or Sheets that writes formulas, fills ranges, and answers questions about the open workbook.
Examples. Copilot in Excel, Gemini in Sheets, and a range of add-ins
Genuine strength
Works where the data already is, and genuinely good at writing a formula you would have had to look up. Low friction because it requires no new system.
Structural ceiling
Scoped to one workbook, so it inherits whatever errors are already in it. It cannot reconcile across systems, and it has no more durable business context than a chat window.
BI copilots
A natural-language layer over a BI tool or warehouse. You ask a question, it generates a query against a modelled dataset and returns a chart.
Examples. Copilots in Power BI, Tableau, Looker and similar
Genuine strength
Queries live, governed data and returns reproducible results. Where a semantic layer already exists, definitions are consistent by construction.
Structural ceiling
Requires the warehouse and semantic model to exist first, which is a significant prerequisite for most mid-market finance teams. Answers questions the model anticipated; genuinely novel questions still need modelling work.
FP&A platforms with AI features
An established planning and reporting platform with AI added: narrative generation, anomaly flagging, natural-language search over existing reports.
Examples. Datarails, Mosaic, Cube, Drivetrain, Fathom
Genuine strength
Genuinely good at what the underlying platform does: consolidation, recurring reporting, planning workflow, and version control. Mature products with real operating histories and enterprise-grade security posture.
Structural ceiling
The AI sits on top of a report builder, so it is strongest on questions the reports already answer. For the unplanned question, most teams still export to Excel, which is the pattern we heard consistently from finance leaders who already owned one of these.
Agentic financial analysis
Where DataWyse sitsA model plans the analysis but never performs the arithmetic. Calculations execute as code in a sandboxed environment against connected source data, output is validated, and every figure returns with its formula and lineage.
Examples. DataWyse and a small number of early entrants
Genuine strength
Built specifically for the unplanned question. Company-specific rules are captured once and applied automatically, and every number can be verified without rebuilding it.
Structural ceiling
An early category. Operating histories are short, customer counts are small, and enterprise security certifications are generally still in progress. If your procurement process requires a completed SOC 2 Type II today, that is a real gate.
Human analyst or outsourced finance
A person (a hire, a fractional analyst, or an outsourced team) does the analysis.
Examples. In-house FP&A hire, fractional CFO services, outsourced finance
Genuine strength
Judgement. A person can decide what is worth investigating, challenge the question, and be argued with. For genuinely novel strategic work, this is still the highest-quality option and nothing here replaces it.
Structural ceiling
Cost scales linearly and capacity is fixed while demand is variable. A new hire takes months to acquire the business context that makes ad-hoc work possible, and you staff for the average while drowning in the peaks.
All six approaches, scored against the criteria
Scored 1 to 5 on what the architecture can support, not on marketing claims. Note where DataWyse's tier scores worst, setup and cost, because those are real trade-offs against a free chat window.
| Approach | Numerical accuracy | Company context | Verifiability | Speed to answer | Ease of setup | Low cost | Average |
|---|---|---|---|---|---|---|---|
| General-purpose AI chat | 2 | 1 | 1 | 5 | 5 | 5 | 3.2 |
| AI spreadsheet assistants | 3 | 2 | 3 | 4 | 5 | 4 | 3.5 |
| BI copilots | 4 | 3 | 4 | 3 | 2 | 2 | 3.0 |
| FP&A platforms with AI features | 4 | 3 | 4 | 3 | 2 | 2 | 3.0 |
| Agentic financial analysis | 5 | 5 | 5 | 4 | 3 | 3 | 4.2 |
| Human analyst or outsourced finance | 5 | 5 | 4 | 1 | 1 | 1 | 2.8 |
DataWyse built the framework and sits in tier 5, so read the scorecard with that in mind. We have tried to score the criteria we are weak on honestly: general-purpose chat beats us outright on setup and cost, an FP&A platform beats us on security certification maturity, and a human analyst beats every tool on judgement. Verify all of it in a demo rather than taking our word.
How to choose, by where you actually are
You need an answer in the next hour
Use ChatGPT or Claude, and verify every number by hand. Define your metrics inline, state your fiscal calendar, and require the formula behind each figure. It will not be reliable enough to act on unchecked, but it will get you to a structured approach faster than anything else.
Consolidation is your bottleneck
Buy an FP&A platform. If the monthly pack takes a week to assemble and data lives in five systems, that is a reporting problem and this category solves it well. An agentic analysis tool will not fix a consolidation problem.
Reporting works, but every report creates a week of follow-ups
This is the ad-hoc problem, and buying a second reporting tool will not touch it. Look at tier 5, and evaluate hard on arithmetic execution, context capture, and verifiability: criteria 1, 2, and 3.
Procurement requires completed SOC 2 Type II today
That is a hard gate, and it currently rules out most of the agentic category including DataWyse, where certification is in progress rather than complete. An established FP&A platform will clear it first. Say so early rather than discovering it at contract stage.
The work is genuinely strategic and novel
Hire a person. Judgement, challenging the premise of a question, and being argued with are not things any tool in this guide does. The realistic role for automation is removing the assembly work so that person spends their time on judgement.
You already have a warehouse and semantic layer
A BI copilot may be the cheapest path, since the expensive prerequisite is already built. Evaluate how it handles questions your semantic model did not anticipate: that is where the category tends to stop.
Six red flags in any AI finance demo
An accuracy percentage with no methodology
“99% accurate” means nothing without knowing what was measured, against what baseline, over what period. Ask where the arithmetic executes instead: that is verifiable on the call.
Cannot reach source rows in two clicks
If the presenter cannot trace a number back to the underlying data on their own demo environment, you certainly will not on your production data.
Only follows the scripted demo path
Insist on asking a question they did not prepare for. The difference between a report builder and something that handles unplanned questions shows up immediately.
Vague on training data
“Your data is secure” is not an answer. Ask specifically whether inputs are used for model training, get it in the contract, and check the tier being sold matches the tier demoed.
Requests write access for convenience
An analysis tool has no legitimate need to modify your ledger. Read-only should be provable at the connection level, not asserted in a policy document.
Never says “I don't know”
Ask something the data genuinely cannot answer. A system that always produces an answer will eventually invent one, and you will not be able to tell which.
Questions buyers ask
What are the best AI finance tools?
The honest answer depends on the job. For formulas inside a spreadsheet, an Excel or Sheets assistant. For consolidation and recurring reporting, an FP&A platform. For close and reconciliation, accounting analytics. For unplanned questions across systems, agentic analysis. A tool that claims to be best at all four is describing marketing scope, not product scope.
What are FP&A AI tools?
FP&A AI tools add model-driven features to planning and reporting platforms: natural-language search over existing reports, forecast suggestions, anomaly flags, and drafted commentary. They work well within the reports the platform already builds. They generally cannot answer a question nobody built a report for, because the underlying model is scoped to the platform's own data model.
What is AI FP&A software?
AI FP&A software is a financial planning and analysis platform with model-driven features layered on top of consolidation, reporting, and planning. The two questions worth asking of any of them: where does the arithmetic physically execute, and can you trace a number back to its source rows in two clicks. Both are verifiable on a demo call.
What is an AI financial analyst?
An AI financial analyst is a system that plans and runs financial analysis from connected source data rather than only visualising it: connecting to the ERP and CRM, assembling the data, running the calculation, and returning the answer with its working. It does not replace judgement. What it removes is assembly, which is roughly 70% of the time in a typical ad-hoc analysis.
Can ChatGPT or Claude do financial analysis?
They are genuinely useful for structuring an approach, explaining a concept, drafting commentary, and reviewing your reasoning. They are unreliable for producing numbers you intend to act on, because a language model generates arithmetic as text rather than executing it, and it has no durable knowledge of your fiscal calendar, metric definitions, or business exceptions. Use them for language, not for the numbers.
What is the difference between an AI chat tool and agentic financial analysis?
In a chat tool the model produces the answer directly, including the arithmetic. In an agentic system the model plans the analysis and the calculation executes as code in a sandboxed environment against your connected data, with the formula and source rows returned alongside the result. The distinction matters because it moves accuracy from being a property of the model to being a property of the architecture.
Is it safe to put financial data into ChatGPT or Claude?
Check your organisation's policy and the provider's data retention terms before anything leaves your network. Consumer tiers have historically differed from enterprise agreements on whether inputs may be used for training. Many finance teams restrict use to aggregated or anonymised figures until an enterprise agreement with a signed DPA is in place.
Do I need an FP&A platform and an AI analysis tool?
It depends which problem you have. If consolidation and recurring reporting are painful, an FP&A platform solves that and an AI analysis layer will not. If your reporting already works but every report generates questions that send you back to Excel, a second reporting tool will not help. Diagnose the problem before evaluating the category.
How do I evaluate accuracy in an AI finance tool?
Do not accept an accuracy percentage. Ask where the calculation physically executes, ask to see the formula and source rows behind a number in the demo, and ask what the system does when the data cannot answer the question. Accuracy in this category is determined by architecture, and architecture is a question you can actually verify on a call.
What should I ask in a demo?
Ask a question they did not prepare for, using their own demo data. Ask to trace any number back to its source rows in two clicks. Ask what happens when a month is missing. Ask for a reference customer at your revenue and headcount rather than their largest logo. Those four questions separate a report builder from an analyst faster than any feature list.
Is this guide independent?
No, and you should read it accordingly. DataWyse builds an agentic analysis product and sits in tier 5. We have tried to be accurate about where other approaches genuinely win: general-purpose chat is better if you need an answer in the next thirty seconds and cannot wait for setup, an FP&A platform is the right answer if consolidation is your bottleneck, and a human analyst is still better for judgement-heavy strategic work. Where we score ourselves highest is on the criteria we built the product around, which is exactly what you would expect and exactly why you should verify it yourself in a demo.
Not sure which approach fits?
Answer eight questions about your team, your systems, and what is actually breaking. You get a ranked recommendation across all six approaches, including the ones that are not us.