No wrong number reaches a decision unflagged.

In finance, a fast answer is worthless if you can’t trust it. We built an AI co-pilot that lets finance teams ask questions of their financial files in plain English, and get answers where every number can be traced back to the exact file, sheet and cell it came from. The AI is never allowed to do the maths. When the system isn’t sure, it says so. Tested on 100 financial files before delivery, 93% of answers matched the source exactly, and every other answer was flagged for human review. No wrong number got through silently.

• ClientA technology partner’s client in financial services (names under NDA)
• IndustryFinance & accounting
• What we builtAn AI assistant that answers questions on financial files, with every number traced to its source
• Files it readsExcel, CSV, PDF and Word, including messy files with mixed units, total rows and negatives in brackets
• Accuracy in testing93% of answers matched the source exactly, across 100 financial files
• SafetyEvery other answer was flagged for human review; no wrong number got through unflagged
• SpeedUpload in 7–12 seconds; full file processing in about 1 minute
• Data privacyRaw files stay on the user’s machine
• Team and timeline3 people, 13 weeks
• StatusDelivered; LatentIQ provides ongoing bug-fix support

Numbers arrive as Excel workbooks, CSV exports from accounting systems, PDF statements and reports, and Word documents, each prepared by a different person, in a different format, sometimes in different units.

Every day, someone needs an answer from those files: a total for a period, a comparison between two entities, a figure for a report or a decision. Finding that answer means opening the right file, finding the right sheet, understanding how it was built and checking that the number is the right one.

The client wanted their team to simply ask a question in plain English and get an answer they could rely on.

1. 01

   Problem 1: Answers buried in files

   Finding and checking one number can mean searching across several files and sheets, then cross-checking it against another source. It is slow, repetitive work, and it pulls skilled people away from analysis.
   What it cost: hours of finance time spent locating numbers instead of using them.

2. 02

   Problem 2: Real financial files are messy

   Financial files are not clean databases. The ones this team worked with mixed:
   • Different units: lakhs, crores and USD thousands, sometimes in the same set of files
   • Total and subtotal rows sitting inside the data, which double-count if they are added again
   • Negatives written in brackets, such as (1,250) meaning minus 1,250
   • Tables inside PDFs and Word documents, not just spreadsheets
   What it cost: any tool that reads these files naively gives wrong totals, and nobody notices.

3. 03

   Problem 3: General AI chatbots invent numbers

   General AI chatbots answer quickly, but they sometimes produce numbers that look right and aren’t in the source at all. In most work, that is an annoyance. In finance, a single invented number can lead to a wrong decision: a wrong forecast, a wrong report or a wrong call on money.
   What it cost: the team couldn’t use AI for the work where speed would help most, because they couldn’t trust it.

4. 04

   Problem 4: Sources that disagree

   When two files report different figures for the same thing, someone has to decide which is right. A tool that silently picks one hides the problem instead of raising it.
   What it cost: errors that only surface later, in a report or a review.

5. 05

   Problem 5: Sensitive data

   Financial files are among the most sensitive data a business holds. Any solution had to keep raw files under the user’s control.

The obvious solution was to hand the files to an AI model and let it answer. We rejected that, because it fails exactly where finance can’t afford it.

• The AI doesUnderstands the user’s question in plain English · Decides which data and calculation the question needs · Writes a clear reply in plain language
• Ordinary code doesReads and cleans the files · Checks that totals add up and units make sense · Runs every calculation · Records exactly where every number came from

Five principles shaped every decision:

1. Separate understanding from calculating. The AI interprets; code computes. A language model never produces a number on its own.
2. Check before trusting. Data is validated before it is used: totals must add up and units must make sense.
3. Never guess silently. When sources disagree, ask the user. When the system isn’t sure, flag the answer for a person.
4. Every number shows its work. Each figure can be traced to its exact file, sheet and cell.
5. Data stays with the user. Raw files stay on the user’s machine.

1. 01

   4.1 Reads messy financial files

   The co-pilot reads Excel, CSV, PDF and Word files and turns them into clean, consistent data. It handles mixed units (lakhs, crores and USD thousands), recognises total rows so they aren’t counted twice, and reads negatives written in brackets correctly.
   Who uses it: analysts upload files as they are, with no reformatting.
   Why it matters: the team doesn’t have to clean files before asking a question, and the messiness doesn’t turn into wrong answers.

2. 02

   4.2 Checks the data before using it

   Before any data is used, the system checks it: totals must add up to their parts, and units must make sense. Data that fails a check is not passed off as reliable.
   Who uses it: runs automatically on every file.
   Why it matters: errors in the source files are caught at the start, not discovered in a finished report.

3. 03

   4.3 Asks when sources disagree

   When two sources give different figures for the same thing, the co-pilot asks the user which one to use, instead of silently choosing one.
   Who uses it: analysts and managers, who decide which source is authoritative.
   Why it matters: conflicts become visible decisions, made by a person who understands them.

4. 04

   4.4 The AI understands; code calculates

   The AI model’s only jobs are to understand the question and to write the reply. Every calculation is run by ordinary code on the cleaned data, through tool-calling: the AI asks for a calculation, and code performs it.
   Who uses it: everyone asking questions; this is how every answer is produced.
   Why it matters: the numbers in the answer are computed, not generated, so the AI can’t invent them.

5. 05

   4.5 “Show your work” on every number

   Every number in every answer has a “show your work” button that traces it to the exact file, sheet and cell it came from.
   Who uses it: analysts checking an answer, managers relying on it, and reviewers auditing it.
   Why it matters: anyone can verify a figure in seconds, which is what makes it safe to use in a report or decision.

6. 06

   4.6 Flags anything uncertain for a person

   When the system isn’t confident an answer is right, it doesn’t guess. It flags the answer for human review.
   Who uses it: the finance team’s reviewers.
   Why it matters: an uncertain answer becomes a task for a person, not a hidden error.

7. 07

   4.7 Keeps raw files on the user’s machine

   Raw financial files stay on the user’s machine.
   Who uses it: the client’s leadership and anyone responsible for data security.
   Why it matters: sensitive financial data stays under the client’s control.

An example of the flow; the question is illustrative.

1. The analyst uploads files: say, two Excel workbooks and a PDF statement. Upload takes 7–12 seconds, and full processing about a minute.
2. The files are read and cleaned. Units are converted to one standard, total rows are identified, and bracketed negatives are read as negatives.
3. The data is checked. Totals are matched against their parts, and units are checked for sense.
4. The analyst asks: “What were total expenses for the quarter across both entities?”
5. The AI works out what is needed: which figures, from which data, and which calculation.
6. Code runs the calculation on the cleaned data.
7. If two sources disagree, the co-pilot asks the analyst which one to use.
8. The AI writes the answer in plain English, with the computed numbers.
9. Every number has “show your work”, linking to its file, sheet and cell.
10. If the system isn’t sure, the answer is flagged for human review instead of being presented as fact.

1. 01

   Finance analyst

   Before: searches files and sheets by hand, converts units, re-checks totals.
   After: asks in plain English and gets a traced answer.

2. 02

   Finance manager

   Before: waits for answers, and can’t easily check where a number came from.
   After: gets answers faster, and can check any number in one click.

3. 03

   Reviewer or auditor

   Before: rebuilds the trail behind a figure manually.
   After: follows “show your work” straight to the source cell.

4. 04

   Leadership

   Before: can’t use general AI on financial data, because it can’t be trusted.
   After: uses an AI assistant that never lets a wrong number through unflagged.

• Financial files tested100
• Answers that matched the source exactly93%
• Remaining answersAll flagged for human review
• Wrong numbers that got through silentlyNone
• Upload time7–12 seconds
• Full file processingAbout 1 minute

Why the 93% matters less than the 100%. No system is right every time. What matters in finance is what happens when it isn’t. Here, every answer was either exactly right or clearly flagged. A person always knew which answers to trust and which to check. That is the difference between an AI a finance team can use and one it can’t.

A team of 3 people built and delivered the co-pilot in 13 weeks. LatentIQ continues to provide bug-fix support.

• PythonThe core of the system
• PolarsFast reading and cleaning of tables
• PyMuPDFReading text and tables from PDF files
• DuckDBRunning calculations on the cleaned data
• ChromaDBFinding the relevant parts of documents for each question
• LLM tool-callingLetting the AI request calculations from code instead of doing them itself
• StreamlitThe interface where users upload files and ask questions

1. It designed for trust, not just speed. The goal was answers a finance team could act on, so every design choice protected accuracy first.
2. It gave each job to the right tool. AI handles language; code handles numbers. That one decision removed the risk of invented figures.
3. It treated messy files as normal. Mixed units, total rows and bracketed negatives were handled by design, not left to break things.
4. It made uncertainty visible. Flagging and asking replaced silent guessing, so errors couldn’t hide.
5. It made every number checkable. “Show your work” turned each answer into something a reviewer can verify in seconds.

The same design (AI to understand, code to calculate, every number traced) fits any team that needs answers from financial or operational files it can’t afford to get wrong:

• Accounting and CA firms working across many clients’ files
• Finance teams preparing monthly, quarterly and board reports
• Businesses consolidating figures across entities, currencies and units
• Review and audit preparation, where every figure needs a source

AI in finance isn’t useful because it is fast. It is useful when every number can be trusted, or is clearly flagged when it can’t.