New York family office · $500m–$1bn AUM · fewer than 10 staff
The figures were in the report. The risk was what happened on the way into the model.
Practical Claude and Cowork training gave the investment team a higher-accuracy PDF-to-model workflow and a structured way to build, test and verify reusable agents.
What changed
From prompt to process
The team learned how to separate extraction, reasoning, model updates and verification instead of asking one model to do everything at once.
- Built around live investment work
- 3 sessions
- Typical raw LLM extraction on the tested task
- 80–85%
- Specialist extraction on the relevant workflow
- ~99%
- Reusable research skills delivered
- 2 tools
The team was already experimenting with AI, but capability varied. Some people used Claude as a more capable search box. Others had built detailed prompts and skills, yet still rebuilt the same instructions in each new conversation.
The most valuable tasks were also the hardest to trust. Analysts worked with dense PDF reports, adjusted financial statements and Excel models carrying years of formulas, comments and judgment. A plausible-looking answer was not enough if the period was wrong, a row was skipped or an existing formula was replaced.
Altitude 7 organized three working sessions around the firm's own reports, spreadsheets and research questions. One central example was a recurring update: extract monthly retail or store-sales data, map each figure into the existing model and populate the new period without damaging the workbook.
Inside the working method
Turn a financial PDF into a model update an analyst can inspect.
Each stage removes a different source of error, while the route from source document to final workbook remains visible.
Document-to-model workflow
Specialist extraction · controlled mapping · analyst verification
A repeatable path from source to workbook
EXTRACT → CHECK → MAP → WRITE → VERIFY
Financial PDF
Dense tables, repeated headings, footnotes and several reporting periods.
Raw LLM extraction: 80–85%
Structured extraction
LlamaIndex and LlamaParse convert the report into inspectable rows and columns.
~99%
Relevant extraction workflow
Check before mapping
Period
Correct month or quarter
Units
Currency and scale retained
Basis
Reported vs adjusted
Ambiguous or missing fields stop here for review.
Map source to model
Store sales
→Model row 42
Same-store sales
→Model row 47
Store count
→Model row 53
The approved mapping becomes reusable for the next reporting period.
Approved workbook
Update only the intended cells, then inspect the model before use.
Analyst approved
Source visible
Every figure can be traced to the report
Change visible
Updated cells and exceptions can be reviewed
Logic preserved
The analyst retains control of the model
Accuracy before automation
Extraction and reasoning are different jobs.
A general-purpose language model can read a document and return a table that appears complete. On the type of financial PDF the team was testing, direct extraction could sit around 80–85% accuracy. That may be enough for a summary. It is not enough for a financial model.
Altitude demonstrated a specialist document-processing approach using LlamaIndex and LlamaParse. On the relevant extraction workflow, the dedicated tooling reached around 99% accuracy. The output still required review, but the analyst could check structured figures instead of reconstructing an unreliable table.
The broader lesson was tool choice. A specialist service turns the PDF into structured data. Claude then interprets that data, applies firm context and helps decide where it belongs. Each tool handles the part of the job it is better suited to perform.
Map once, review every run
Finding a number is only half the task.
The firm's models used their own row labels, adjusted measures and historical conventions. We showed how to create a repeatable mapping from each source line to the correct model row, so the relationship did not have to be guessed again every month or quarter.
That was particularly useful for recurring reports such as retail and store-sales updates. Once an analyst had reviewed the mapping, later periods could follow the same structure while missing or ambiguous fields were surfaced as exceptions.
Claude and Claude in Excel could then assist with approved model updates. The team learned to constrain which cells could receive values, compare the new period with history and verify the workbook afterwards. Formulas, comments and formatting were treated as part of the model's logic, not incidental presentation.
From using AI to building agents
A reusable agent is more than a long prompt.
Altitude introduced its structured agent-building framework through live examples. The team learned to define the outcome, provide the right sources and context, constrain what an agent could change, test it against known examples and verify the result before broader use.
Testing was part of the build. The examples checked for missing data, wrong periods, incomplete rows and unintended changes to spreadsheet logic. That showed staff how a useful workflow moves from an idea to a tool the team can inspect and improve.
The engagement also delivered two reusable research tools: a company evidence pack that attaches source, period, units and status to key figures, and a signal filter that tests outside commentary against the firm's investment thesis and watch points.
What the team could do afterwards
Use AI as a controlled research workflow, not an isolated answer.
Choose between Claude, Cowork, Claude in Excel and specialist document extraction.
Turn difficult PDF tables into structured data before asking an AI model to reason over them.
Map recurring retail and store-sales figures into the correct rows of an existing model.
Protect formulas, comments and formatting while updating approved cells.
Build, test and verify reusable agents against real investment work.
Treat missing or ambiguous data as visible exceptions rather than plausible answers.
This was a training and enablement engagement rather than an unattended production deployment. The team left with the methods, examples and tools to automate repeatable work while retaining analyst approval over the final model and investment conclusion.
The service behind this work
AI Training & Fluency
Role-specific training, live-work clinics and power-user tracks that build practical capability across the firm.
See how AI training works →Read next
AI training
A firm-wide AI day with a defined next step
Investment, technology, operations and compliance worked through live workflows, reusable skills and the decisions behind a coordinated programme.
AI training
Building agents the team could challenge
A Claude Code programme on agent scoping, sourced evidence, testing, multi-agent design and the decisions behind shared agents.
Does your team have an AI workflow it still cannot trust?
Bring one recurring workflow and the files it depends on. We'll shape the training around the work your team needs to complete and show staff how to build, test and verify the tools that support it.