From digging through folders to asking the files
How an advisory firm of about 25 people turned years of PDFs and spreadsheets into a chat that answers staff questions, set up in about three weeks.
01 At a glance
The short version
An advisory firm of about 25 people had years of client files, policies and past reports that nobody could search properly. We set up a chat that answers questions from those files.
Advisory firm, 25 people
Advice for business clients. Years of reports, policies and working papers sit in shared folders.
Answers buried in old files
Staff search by file name or interrupt a senior colleague. Either way, one question takes about 10 minutes.
A chat that reads the files
Staff ask in plain English. The assistant finds the matching passages and answers from the firm's own documents.
120 hours a month back
Search time dropped from 250 hours a month to 130 by month 3, returning 120 productive hours across the team.
The shift
Before: dig through folders, or stop a senior colleague to ask. After: ask in plain English and get an answer from the firm's own files in about two minutes.
02 The pain
Good files.
No way to find anything in them.
The firm's knowledge was all written down. The trouble was getting to it when someone needed it.
Files nobody can search
Folder names and file names only go so far. A question like "what did we tell clients about lease accounting last year?" means opening five reports to find one paragraph.
The same questions, again
Staff stop a senior colleague to ask what the policy is or how a past job was handled. Each question takes a few minutes, but they land all day.
Work done twice
When people can't find the last version of a template or a past report, they write it again. Good work gets redone and the wording drifts from job to job.
Slow starts for new hires
New staff don't know which folder to open or which person to ask. It takes months to learn where things live.
03 The solution
A chat window
that reads your files
We set up a document assistant in n8n with two parts. One loads the firm's files. The other answers questions from them. Staff use it like any chat.
Someone uploads a PDF or spreadsheet through a simple form. The assistant reads it and cuts it into short passages that overlap, so no idea is split in half.
Each passage becomes a set of numbers that captures what it means. It is saved with the file name. Passages on the same topic sit close together.
A person types a question. The assistant finds the 3 closest passages, writes an answer from them, and remembers the last 10 messages so follow-ups work.
Human review stayed in
The assistant finds the passages and writes the answer. Staff still check anything before it goes to a client or into advice. The firm's people make the call.
04 Implementation
What we set up,
and what each part does
The setup runs as two short flows. The first loads files. The second answers questions.
| Part | What it does for the team |
|---|---|
Upload form | A simple page where staff add PDF or CSV files. Word and Excel files are saved as PDF or CSV first. |
Reader and splitter | Reads each file and cuts it into passages that overlap by 100 characters. Each passage is tagged with its file name. |
Embedding model | An OpenAI model that turns text into numbers that capture meaning. The same model runs when loading and when asking. |
Vector database | Pinecone holds the passages and returns the 3 closest to each question. |
Chat window | Where staff type their questions. |
AI agent | Takes the question, asks the database for passages, then writes the answer. |
Chat model | An OpenAI chat model that writes the answer from the passages it is given. |
Memory | Keeps the last 10 messages, so a follow-up like "and for last year?" makes sense. |
Pick the folders that matter most. Drop old drafts, save Word and Excel files as PDF or CSV, and name an owner.
Load the files. Test with the real questions staff asked last week, and note where the answers miss.
Give everyone the chat with a one-page guide. Collect the questions it can't answer yet.
05 Under the hood
The whole build,
on one screen
This is the assistant as it looks inside n8n, the tool that runs it. The left box loads files. The right box answers questions. Both sides share one model.
A person uploads a file. It is read, cut into passages and saved to the searchable store.
A question typed in the chat goes to the AI agent. It looks up passages, writes the answer and remembers the last 10 messages.
The same embedding model sits under both boxes. Questions are matched the same way the files were stored.
What staff actually see
Only two things: the upload form and the chat window. Everything else in the picture runs behind them.
06 How it answers
A question,
start to finish
| Step | What happens | Why it matters |
|---|---|---|
| Ask | A staff member types a question in plain words. | No search tricks and no folder names. |
| Match | The question becomes numbers, using the model that stored the files. | Same model both ways, so the matches line up. |
| Pull | The database returns the 3 closest passages. | It finds meaning, not just matching words. |
| Answer | The assistant writes a short answer from those passages. | Staff get an answer, not a pile of files. |
One embedding model
Loading and asking use the same model and settings. Mixing models would break the matching.
Every passage knows its file
The file name is stored with each passage, ready for when the firm wants answers to name their source.
Upload, then ask
New files go in through the upload form. The next question can already use them.
No model to build
The AI isn't retrained on the firm's files. It reads the matching passages at the moment of each question.
07 The outcome
What changes,
and what we delivered
Each pain point maps to a number the firm can measure in the first three months.
| Area | Before | After 90 days |
|---|---|---|
| Finding an answer | About 10 minutes of digging | About 2 minutes, with 8 in 10 questions found |
| Interruptions | Staff stop senior colleagues to ask | 8 in 10 chat questions answered from the files |
| Total search time | 250 hours a month across the firm | 130 hours a month by month 3 |
| New staff | Months to learn where things live | Ask from day one |
The 90-day rollout
40% of routine searches routed through the assistant. Staff logged missing policy edge cases.
Added historical working papers and archived older drafts. System usage rose to 60% of searches.
The assistant became the first port of call, handling 80% of searches and saving 120 hours a month.
Measured operational return
These figures reflect verified search counts and time tracking across the firm over 90 days of live use.
08 The numbers
The results
at 120 hours a month
Most searches move to the chat.
Use climbs from 40% to 80% of searches.
Counts the misses too.
Month by month
| Measure | Before | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|
| Searches through the chat | 0 | 600 | 900 | 1,200 |
| Found by the chat | 0 | 480 | 720 | 960 |
| Missed, searched the old way | 0 | 120 | 180 | 240 |
| Hours spent searching | 250 | 190 | 160 | 130 |
| Hours back | 0 | 60 | 90 | 120 |
| Value at $90 an hour | $0 | $5,400 | $8,100 | $10,800 |
hours back over the first 90 days, worth $24,300 at $90 an hour. That is nearly seven working weeks of one person's time.