Token Efficiency for Finance: 8 Habits That Cut Your AI Cost
Most finance teams waste AI tokens without realizing it. Not because they use AI too much, but because they use it inefficiently. The largest cost in a tool like Claude is rarely the answer you get back. It is everything the model has to re-read before it can answer: your old messages, uploaded files, previous corrections, a messy workbook, a vague prompt.
Token efficiency is quietly becoming a real finance and FP&A skill, and most teams are not thinking about it yet, even as nearly all of them now use AI in their day-to-day work (AI in finance statistics). The good news is that most of the waste comes from a handful of habits, and each has a fix. Here are eight that make an immediate difference, and one that matters more than all of them combined.
Where the tokens actually go
Every time you send a prompt, the model re-reads the context it needs to answer. On a clean, focused task that context is small. On a typical finance session, it is enormous: a full workbook uploaded "just in case," a long chat history, a heavy PDF, and an instruction like "analyze this." The model dutifully ingests all of it before doing the thing you asked. You pay for the re-reading, not the reasoning.
So token efficiency is mostly context hygiene. Give the model exactly what the task needs and nothing more, and the cost collapses toward the size of the actual question.
The eight habits
1. Don't upload the whole workbook
Upload only the tabs or tables the analysis needs. A 20-tab model dumped in whole means the model re-reads all 20 tabs to answer a question about one. Extract the relevant range and paste that instead.
2. Convert PDFs to clean text or markdown
A heavy PDF is expensive to parse and full of layout noise the model has to wade through. A clean text or markdown version of the same content is usually far cheaper to process and easier for the model to read correctly.
3. Match the model to the task
Not every finance task needs the most powerful model. Use a fast, light model like Haiku for quick lookups and extraction, Sonnet 5 as your default workhorse for most analysis, and save Opus 4.8 or Fable 5 for genuinely hard reasoning: a tricky valuation, a subtle reconciliation, ambiguous drivers. Reaching for the heaviest model on every task is one of the quietest ways to overspend.
4. Be specific
A vague prompt forces the model to guess, over-read, and over-produce.
Bad: "Analyze this file."
Better: "Compare Actual vs Budget for Revenue, Gross Margin, and EBITDA. Highlight only variances above 5%, explain the likely drivers, and draft CFO-ready commentary in five bullets."
The second prompt scopes the work, so the model reads what it needs and writes what you want, once.
5. Ask for shorter outputs first
Start with an executive summary. Then ask for detail only where you actually need it. Generating a full report you skim and discard burns output tokens on text no one reads.
6. New topic, new chat
A long-running chat carries its entire history into every new prompt. When you move to a genuinely new topic, start a fresh session so you are not paying to re-read an unrelated conversation on every turn.
7. Summarize long sessions, then continue fresh
After 15 to 20 messages, ask the model to summarize the work so far, then continue in a new chat from that summary. You keep the thread of what you decided without dragging the full transcript along behind you.
8. Redo only the section that is wrong
If one paragraph or one line is off, ask the model to fix that section. Regenerating the entire report because a single part is wrong pays full price to reproduce everything that was already correct.
The habit that beats all the habits
Every tip above is context hygiene at the margin. It helps, and you should do it. But if you build and edit financial models with AI, there is a deeper source of waste that no amount of prompt discipline touches: you are asking the model to be your spreadsheet.
When the thing the model works on is a raw grid, it has no map of what matters. To change one assumption safely, it re-reads large parts of the model to work out the structure first, then rebuilds formatting and formulas it has to re-derive every turn. The cost tracks the size of your whole model on every edit, not the size of the change. That is not a prompt you can shorten your way out of. It is structural.
The fix is to stop making the model rebuild your spreadsheet and let it reason on top of a structure that already exists. Keep the model's logic, named variables, explicit dependencies, and your conventions, in a dedicated layer the AI reads once and edits surgically. Then changing an assumption is a targeted operation on one node, not a re-read of the sheet, and the context persists between sessions instead of being re-explained each time. This single shift changes the order of magnitude, not just the margin. (The full mechanism is in How to Save Tokens When Building Financial Models with AI.)
The bottom line
If your tokens vanish in a handful of prompts, the answer is not to stop using AI. It is to stop making it re-read things it does not need. The eight habits above cut the everyday waste: upload less, convert PDFs, right-size the model, prompt specifically, output short first, keep chats scoped and summarized, and fix only what is broken. And if your work is building models, the biggest lever is structural: give the AI a model to reason over instead of a grid to rebuild, and the cost drops to the size of the edit you actually made.
For the deeper structural picture, see How to Save Tokens When Building Financial Models with AI and Claude for Financial Modeling.
Layerz holds your model's structure so the AI reasons on a real workspace instead of re-reading a grid on every edit. Named variables, surgical edits, persistent context, conventions in a FINANCE.md, and clean Excel export anytime. Fewer tokens, more model. Explore Layerz →
