# The Frozen Excel Preference There is a pattern showing up in finance teams that almost nobody is writing about, because it runs counter to the narrative. Finance leaders at organisations with messy, ungoverned data have made a deliberate choice: they prefer working with a frozen Excel snapshot over connecting AI to their live financial data. Not because they are afraid of technology. Not because they distrust AI as a concept. Because they cannot trust what the AI will do with data that has no governed definitions, no intercompany reconciliation rules, no clean chart of accounts underneath it. **The frozen snapshot has known inputs. The live data has unknowns. And in finance, unknown inputs are not a theoretical problem. They are a liability.** ## **The rational logic behind a seemingly irrational choice** When you hear that finance teams are reverting to snapshots, the instinct is to frame it as conservatism. That reading is wrong. The teams doing this are applying the same logic they use when signing off on a quarterly close: if you cannot validate the inputs, you cannot stand behind the output. A frozen Excel file can be audited. You know exactly what was in it when you ran the model. You can reproduce the answer. Live financial data connected directly to an AI model is a different proposition. If your chart of accounts has duplicate cost centre codes, if your intercompany eliminations are handled inconsistently across entities, if EBITDA means three slightly different things depending on which business unit built the report, then the AI is not working with your financial reality. It is working with your financial noise. **Limitations on one side creating preferences on the other. The preference for frozen data is not irrational. It is a precise response to a data governance problem that nobody has solved yet.** ## **What CFOs actually said** At a recent CFO Exchange event, the gap between the public narrative and the private conversation was striking. Every panel discussion was about agentic AI. Every hallway conversation was about data quality, ERP consolidation, and getting close time down. Three things came up repeatedly: we want AI, our data is a mess, we know. ERP migration is eating the budget. Our teams need to become data-literate before we can automate anything meaningful. Around half of the attendees were already using AI models in some capacity. But they had found them fragile at scale. Try feeding 50 tables into a language model and there are no guarantees on context handling, deduplication, or repeatability across runs. The outputs were inconsistent enough that teams stopped trying to connect AI to live data and went back to prepared extracts. One CFO asked their AI model what the cash position would be at year-end. **The answer was approximately 30% off.** That is not a model failure in isolation. **That is what happens when an AI model meets ungoverned financial data without a semantic layer between them.** ## **The problem is not the model** The model is not the bottleneck. Claude, Copilot, Gemini: none of them were designed to interpret company-specific financial definitions on the fly. They do not know that your organisation records intercompany loans differently in two entities. They do not know that your EBITDA excludes one-off restructuring charges in the board pack but not in the management accounts. **They will give you an answer. It will be confident. It may be 30% off.** The problem is structural: AI needs governed, validated, semantically consistent data to produce reliable financial outputs. Without that foundation, you are not doing AI-assisted finance. You are doing plausible-looking approximation with extra steps. The frozen Excel preference is, in a strange way, the more rigorous choice. At least the CFO knows what they fed in. ## **Five signs your financial data is not ready for AI** These are the conditions that produce the frozen Excel preference. They are also the checklist worth running before any AI pilot in finance: 1. 01 ### Your EBITDA definition varies by entity, report, or team. If different business units exclude different items, or if the board pack and management accounts use different adjustments, an AI model querying EBITDA will retrieve whichever version the data surface returns first. The answer will be confident and wrong in ways that are hard to detect without knowing which definition was applied. 2. 02 ### Intercompany transactions are not consistently coded for elimination. Ungoverned intercompany flows are the most common source of consolidated numbers that do not reconcile. An AI model cannot identify or eliminate intercompany balances it cannot see. The result is group-level figures that are inflated or deflated by whatever intercompany volume is present. 3. 03 ### A number in last month's board pack cannot be traced back to its source in under ten minutes. Auditability is the threshold for AI-readiness. If your finance team cannot reproduce a figure from source data quickly, an AI model operating on the same data will produce outputs that cannot be validated. The AI is not the problem; the traceability gap is. 4. 04 ### You have more than two ERPs contributing to group reporting. Each additional ERP adds chart of accounts fragmentation, currency translation complexity, and intercompany coding inconsistency. Without a governed consolidation layer above all of them, an AI model is working from four or five different financial realities simultaneously and averaging across them. 5. 05 ### Your team prepares a 'clean extract' before any analysis or reporting. The existence of a prepared extract is the signal. It means someone already knows the live data is not trustworthy enough to use directly. An AI model connected to the same live data encounters exactly the same problem, without the human judgment that goes into building the extract. ## **When the preference reverses** The teams that have stopped preferring frozen snapshots all have one thing in common: they built the data foundation before they connected the AI. What that foundation looks like in practice is a governed semantic layer sitting between the raw financial data and the model. When the AI queries EBITDA, it queries your company's definition: pre-validated, consistent across entities, auditable to source journal entries. **It does not make its best guess from 50 tables. It retrieves a governed answer.** Home Credit, operating across nine countries, reduced FP&A time by 70% after building this kind of governed data layer. The change their team described was not that AI became more capable. It was that work shifted from data pushing to intelligence and analysis, because the data was finally in a state where AI could be trusted with it. ## **The question worth asking** Before you connect AI to your financial data, there is a prior question worth sitting with. Not which AI model should we use, or what use cases should we pilot. The prior question is: if an AI model queried your EBITDA figure right now, would it get the same answer your CFO would give in a board meeting? **If the answer is no, if the answer is 'it depends on which entity, which system, which definition we are using that day,' then the AI is not your problem yet. The data is.** The finance teams reverting to frozen snapshots already know this. They are not waiting for better AI. They are waiting for their data to be ready for AI. That is a distinction worth understanding before you start the pilot. ## **Frequently asked questions** **What is the frozen Excel preference in finance?** It is a pattern where finance teams deliberately choose to work from a static, prepared data extract rather than connecting AI tools to live financial data. The reason is data governance: when financial data has no governed definitions, inconsistent intercompany coding, or multiple conflicting chart of accounts structures, a frozen extract with known inputs is more auditable and reliable than a live connection where the AI is working with ungoverned data. The preference is rational, not technophobic. **Why does AI give wrong answers on financial data?** AI models do not know your company's specific financial definitions. They do not know that your EBITDA excludes restructuring charges in one report but not another, or that two entities use different depreciation assumptions. Without a governed semantic layer between the raw data and the model, the AI makes its best interpretation from whatever data it receives. That interpretation can be 30% off on something as fundamental as cash position, not because the model is bad, but because the data it was given was not governed. **What is a semantic layer in financial reporting?** A semantic layer is a governed definition layer that sits between raw financial data and the tools or models that query it. When an AI model queries EBITDA through a semantic layer, it retrieves your company's specific definition: the accounts included, the intercompany adjustments applied, the currency methodology used, consistent across all entities. Without a semantic layer, the AI resolves the definition itself, which produces confident answers that may not match the CFO's definition. **How do you know if your financial data is ready for AI?** The fastest test: ask your AI model what your EBITDA was last quarter and compare the answer to the figure in your board pack. If they match, your data governance is likely sufficient for that use case. If they do not, the gap tells you exactly what the governance problem is. A secondary test: ask your team how long it takes to trace a board pack number back to source transactions. If the answer is longer than ten minutes, the traceability gap that makes manual finance hard is the same gap that makes AI unreliable. **Does this mean AI in finance does not work yet?** No. It means AI in finance requires a governed data foundation to work reliably. The organisations that have built that foundation, a clean chart of accounts mapping layer, consistent intercompany elimination rules, and a semantic layer that enforces company-specific metric definitions, are running AI against financial data with repeatable, auditable outputs. The constraint is not the AI model. It is the data infrastructure underneath it. The frozen Excel preference resolves itself once that infrastructure exists.