Click for Takeaways
- An LLM working on an exported file performs a one-time transformation; nothing in the file tells it your mappings, definitions or reporting rules.
- AI-built reports break in month two because an account, a definition or the model structure moved and the model rebuilt from scratch.
- AI becomes trustworthy with financial data when definitions are stored once, outside the prompt, and every number traces back to transactions.
Can ChatGPT or Claude transform finance data?
Yes, for a single file at a single point in time. ChatGPT and Claude read a trial balance export or a pipeline extract and reshape it into a variance table or a segment summary in under a minute. The result is only as correct as the context typed into the prompt, and it does not persist.
Where it goes wrong is usually a join made with the wrong rule. For example, in February an analyst pasted a Salesforce pipeline export and a NetSuite revenue export into ChatGPT and asked for bookings against revenue by segment. The table came back clean. Salesforce segmented customers by employee count and NetSuite by customer class, and 61 accounts sat in different segments in each system. The model picked the Salesforce grouping without saying so. The board deck showed enterprise revenue up 14 percent while the CRO’s dashboard showed 6 percent, and two teams spent a week finding out why.
Neither model did anything wrong. Nothing in either file said which segmentation finance treats as authoritative or that the intercompany management fee nets to zero at group level. A finance professional carries those rules in their head. An LLM reading an export has none of them unless a person types them in, every time, in the same words.
Can I use Claude to write my financial formulas?
Yes, and for most spreadsheet formulas it will be faster than writing them by hand. Claude produces working SUMIFS, XLOOKUP, INDEX-MATCH and LAMBDA formulas from a plain-English description. The risk sits in the ranges and assumptions the formula hard-codes, which hold in the month you asked and fail later without an error.
In Gartner’s 2025 survey of 183 CFOs and senior finance leaders, code generation ranked highest for impact of any AI use case (Gartner, 2025).
Watch how an AI-written formula behaves on refresh. In May, an analyst asked Claude for a SUMIFS that totaled paid media spend from a HubSpot export by month. The formula referenced rows 2 to 3,000 because that was the size of the paste. June’s export ran to 3,240 rows. Nobody saw an error, and $86,000 of paid media dropped out of the total. Blended customer acquisition cost (CAC) looked 9 percent better than it was, and the improvement made it into the investor update.
The same pattern shows up with date criteria typed as text that stop matching once the header becomes a real date, and with column letters that shift when accounting inserts an account. Ask Claude for the formula. Then ask it to use whole-column or structured-table references, to read criteria from cells rather than literals, and to explain each assumption in a comment. Better still, point the formula at a governed data source instead of a paste, so the range is no longer the analyst’s problem.
Can I use Power Query instead of FP&A software?
For a single-entity company with one accounting system and one analyst who owns the refresh, often yes. Power Query handles extraction, cleanup and merging of exports well and costs nothing extra. Its permissions stop at the workbook, and it has no audit trail, built-in consolidation logic or shared definitions, so it stops being enough at the second entity or the second user.
Power Query is a transformation tool that lives in one workbook and refreshes under one person’s credentials. Financial planning and analysis (FP&A) software is a shared system with a model, permissions and history. The difference matters most when the business changes shape.
| Capability | Power Query | FP&A software |
| Data refresh | Manual or scheduled from one workbook, under the owner’s login | Scheduled from a central connection shared across users |
| Multi-entity consolidation | Written in M code per entity; foreign exchange (FX) and eliminations coded by the analyst | Consolidation rules, FX rates and eliminations stored once |
| Permissions | Workbook level; anyone with the file sees every row | Row and column level, by user or role |
| Change history | File versions only | Audit trail on data, model and definitions |
| Definitions | Inside each query’s applied steps | Stored centrally, reused across reports |
| New account or entity | Query errors or silently drops the new rows | New members flagged for mapping |
When a company adds an entity, the account mapping breaks before anything else, and intercompany eliminations break right after. Take a typical case: in August, a four-entity group on QuickBooks Online added an SAP Business One subsidiary. The Power Query merge keyed on account numbers, and SAP’s chart of accounts used a different numbering scheme. The query loaded without error, matched 0 of 212 SAP accounts, and consolidated revenue understated the group by the entire new entity for one close. A shared mapping table would have flagged 212 unmapped members before anyone opened the report. The FP&A automation evaluation framework lays out the questions to ask before choosing either path.
Why does AI-generated reporting work the first month and break the second?
Because the report was built from a snapshot and nothing carried its assumptions forward. In month two an account was added, a cost was reclassed or a column moved, and the AI either rebuilt the report in a different structure or applied last month’s logic to new data. Both produce a number that looks right and is wrong.
Three things move between closes, and an export shows none of them.
The chart of accounts (CoA) moves. In month one, an AI built a correct gross margin report from a NetSuite general ledger export. In month two, accounting opened a new Stripe processing fees account and reclassed a $57,000 hosting subscription from general and administrative expense into cost of revenue. Gross margin fell 2.3 points. The AI’s commentary called it a cost of delivery increase. Nothing about delivery had changed.
The definitions move. Revenue operations redefines a qualified pipeline stage in Salesforce and every conversion metric shifts while the ledger stays still.
The structure moves. Asked for the same report again, a model regenerates it. Columns land in a different order and the summary tab that pointed at cell F42 reads blank. The model can build a workbook. It has no memory of the one it built last month.
How do I know an AI used my company’s definition of a metric?
You know when you can click the number and see the definition and the transactions behind it. If the answer arrives with no lineage, assume the model used the most common public definition, because that is what it was trained on. Company definitions reach an AI only when they are stored somewhere it must read.
Annual recurring revenue (ARR) is the usual casualty. Sales counts a contract as ARR at signature, the date the customer relationship management (CRM) system records. Finance counts it at service start. In June, a $240,000 contract signed June 28 with a July 15 start landed in the HubSpot ARR dashboard and not in the finance roll. The board deck and the sales dashboard disagreed by 3.4 percent, and each was right by its own rule. An LLM given both tables picks one convention, usually the CRM’s, and says nothing about the choice.
Two tests expose the definition used. Ask the model to state the formula and the source fields. Then change one definition and ask again; if the answer holds still, the definition was never wired in.
The fix is a definition stored once, outside the prompt, read through a governed connection. The Model Context Protocol (MCP), the open standard Anthropic published in November 2024 (Model Context Protocol specification, 2024), lets a model call such a source instead of reading a paste. What sits behind that connection decides the answer, as why the data layer matters for a finance MCP server explains.
What has to be true before I can trust AI with finance data?
Six conditions, all of them about the data underneath the model. Definitions are stored once and enforced. Mappings and consolidation rules live outside the workbook. Permissions apply at the data level. Every number traces to transactions. The model reads and drafts while a named person approves. Outputs survive a change in the source without a rebuild.
Gartner’s 2025 survey put finance AI adoption at 59 percent, level with the year before, with 91 percent of respondents reporting low or moderate impact at first (Gartner, 2025). Inadequate data quality was one of the two largest obstacles it found. Score your own setup with the Month-Two Test.
| Condition | Test | Points |
| Definitions | ARR and gross margin each have one written formula the AI reads from a source rather than the prompt | 2 |
| Mappings | Account and department mappings exist outside any workbook, with a change log | 2 |
| Permissions | A regional manager asking the AI sees only the entities their login allows | 2 |
| Lineage | Any number in an AI answer drills to the general ledger lines behind it | 2 |
| Human review | AI writes drafts; a named person approves journal entries and commentary before posting | 1 |
| Refresh | Last month’s report reproduces this month’s numbers without regenerating the workbook | 1 |
Scoring: 9 or 10, ready for AI on recurring reporting and close work, with a person approving every output. 6 to 8, ready for internal analysis but not for anything reported externally. 5 or below, fix the data before spending on the model.
An auditor asks for two things when AI touched a reported number: the tie-out from that number to the ledger, and who approved the mapping change if one occurred. A pasted export answers neither. The 3-layer test every CFO should run on AI numbers is a shorter version of the same discipline.
How does Datarails give AI tools finance-ready data instead of a raw export?
Datarails does it through FinanceOS®, the governed data layer between your source systems and the AI tools your team uses. It holds what an export leaves out: mappings, definitions, permissions and lineage, stored once and refreshed every period. The AI reads that governed context instead of rebuilding it, so the second month’s report is the first month’s report with new data.
Take the gross margin case. Had the new Stripe fees account arrived as an unmapped member in a mapping layer, it would have been mapped to cost of revenue first, and the $57,000 hosting reclass would have carried a name and a date. The commentary would have read: gross margin down 2.3 points, of which 2.1 is reclassification. A controller can sign that sentence.
The existing Excel model stays the operating model; the Excel add-in and DR.GET formulas point it at governed data, so the workbook stays put and the numbers refresh. The same context reaches Claude or ChatGPT through the FinanceOS AI Connector, so an ARR question gets the finance definition and drills to the transactions. For teams tempted to build that connection in-house, the build-versus-buy question is worth settling first.
If you want to see what the AI tools your team already uses return when they are reading a governed finance layer, a FinanceOS demo is the fastest way to find out.
Where to start
Pick the report that broke last month and list what moved between the two closes: accounts, definitions or structure. That list is your context gap, and closing it is worth more than any model choice. Bring one metric your team argues about to a Datarails demo and watch it resolve.
Check with IT before anyone uploads a trial balance, because the answer depends on your plan and settings. Consumer accounts and enterprise workspaces handle data retention differently. When AI works inside Datarails, it reads your data for analysis only, and customer data is not used to train models.
No. AI in Datarails reads governed data and produces drafts, such as variance commentary or a proposed journal entry. A person reviews, edits or discards each draft before anything is posted, so approval stays with a named owner on the finance team.
Not necessarily. A warehouse organizes data, but it does not hold your finance definitions, mappings or consolidation rules. If you already run one, FinanceOS works alongside it and adds that financial context; if you don’t, FinanceOS connects to your source systems directly.
The FinanceOS AI Connector works with MCP-compatible tools, including ChatGPT and Claude. Because MCP is an open standard, more tools can connect as they adopt it. Each tool sees only what the user’s permissions allow, under the same rules that apply to Datarails reports and models.
Your existing models stay in place. The Excel add-in and DR.GET formulas connect the workbooks your team already uses to governed data. The structure, formulas and formatting stay as they are, and each period the numbers refresh from FinanceOS instead of from a new paste.
Finance usually owns the reported definition, because it ties to the ledger and has to survive an audit. Sales can keep an operational version, such as ARR at signature, as long as both are named, written down and stored where every report and AI tool reads them.
