From 10 minutes of digging to a 2-minute answer
How an IT services firm turns scattered policies, FAQs and how-to guides into one assistant that follows the links between them and names its sources. Set up in three weeks.
01 At a glance
The short version
An IT services firm kept its answers in three places: policy documents, FAQs and step-by-step guides. We set up one assistant that searches all three, follows the links between them and names the records it used.
An IT services firm
A support team and new starters who ask the same policy and how-to questions all week. The answers exist, spread across documents, FAQs and setup guides.
Answers in three places
One question meant opening a policy, an FAQ and a guide. About 150 questions a week, at around 10 minutes each.
One assistant, linked records
A database with three linked tables, a loading flow that sorts new material, and a chat that finds records and follows their links.
56 hours back a month
By month 3 the assistant answered 7 in 10 lookups in under 2 minutes. That delivered $8,190 net value over the first 90 days.
The shift
Before: staff dig through three places and piece the answer together. After: one question, one answer, with the sources named.
02 The pain
The answers existed.
Finding them took too long.
The firm had the knowledge. It was split across documents, FAQs and procedures, and nothing joined them up.
Three places for one answer
A new starter asks how to set up a laptop. The steps sit in one guide, the security rules in a policy and the fix for the usual hiccup in an FAQ. Someone has to open all three.
Search finds words, not connections
Keyword search returns the page with the matching words. It does not know the guide, the policy and the FAQ belong together, so people miss the related one.
The same questions every week
About 150 lookups a week at around 10 minutes each. That is roughly 100 hours a month spent finding answers that are already written down.
No clear source
Two people answer the same question from different pages and give different steps. Nobody can say which page an answer came from.
03 The solution
One assistant that
follows the links
Staff ask a question in plain English. The assistant searches the policies, FAQs and guides, then follows the links between them to pull in the related records. It answers and names its sources. This kind of setup is often called Graph RAG (AI search that follows links between records).
New material goes in through one flow. An AI step sorts it as a document, an FAQ or a procedure, checks it fits that table, and files it.
Each record can point to related ones. A policy points to its FAQs and guides. A guide points back to the policy it came from.
The assistant searches all three tables, takes the IDs of the best matches, looks up the linked records and writes one answer with sources.
Human review stayed in
Staff confirm each answer before it goes to a customer or a new starter. The assistant uses only what is in the database and names the records it used, so a person can check any source in seconds.
04 Implementation
What we set up,
and what it gave the team
Built on Supabase (the database), n8n (the workflow tool) and OpenAI (the AI models). Three weeks, in order: database first, then loading, then the assistant.
| Part | What it gave the team |
|---|---|
Database | A Supabase database with search by meaning switched on, so a question finds records by what they say, not just the words used. |
Three tables | Documents, FAQs and procedures, each with the columns that suit it, such as title, steps and department. |
Link columns | Spaces on each record to point to related ones, such as a policy's FAQs and guides. |
Loading flow | New material is sorted, checked against the table's layout, made searchable and filed in the right table. |
Search tools | One search per table, each returning the 10 closest matches. |
Look-up tools | Three tools that fetch a record by its ID. This is how the assistant follows a link. |
Chat window | Where staff ask questions. It remembers recent messages, so follow-up questions make sense. |
Source rule | Answers use only the database and always name their sources. |
Create the database and the three tables. Agree which columns link to which.
Build the loading flow. Load the policies, FAQs and guides behind the 20 most common questions.
Switch on the assistant. Test it on real questions, fix gaps and fill in the links.
05 The workflow
The two flows
behind the assistant
The whole build on one screen. The top half loads new material into the database. The bottom half is the assistant staff talk to.
A new document is sorted by an AI step, then sent down one of three paths: documents, FAQs or procedures.
Each path shapes the record to its table's columns and files it as a new row. The FAQ path can split one document into several FAQs.
A chat message reaches the assistant. It has memory, three search tools (one per table) and three look-up tools for following links.
06 Under the hood
The structure behind
the linked answers
The first block is the table that holds the firm's policies, with two spaces for links. The second is part of the instructions the assistant follows to use those links.
The policy table. Line 6 makes it searchable by meaning. Lines 9 and 10 hold the links to related FAQs and guides.
After a search, the assistant looks records up by ID in any of the three tables. That is how it follows a link.
Its routine: search, pick out the IDs, fetch the linked records, write one answer. It cites sources and uses only the database.
07 System walkthrough
One question,
three linked records
| Stage | Records in hand | What happens |
|---|---|---|
| Ask | 0 | A new starter types: "How do I set up my laptop?" |
| Search | 1 | It finds "New Hire Laptop Setup" (IT, easy, 9 steps). |
| Follow the links | 3 | It uses that record's ID to fetch the linked FAQ and policy. |
| Answer | 3 | It writes the 9 steps, adds the linked rules and fixes, and names all three. |
An illustration of what happens when the guide, policy and FAQ are linked.
Related records come too
A search for "laptop" returns the guide. The links add the policy and FAQ that go with it.
Less room to guess
It answers only from the database and cites what it used, so there is less to make up.
Every answer has a source
Staff can open the named records and check the answer fast.
Add a table, not a rebuild
A new record type, such as runbooks, is one more table and one more search tool.
08 Delivered outcome
Delivered operational
impact at 90 days
Each pain point mapped to a verified operational metric measured from the first week of deployment.
| Area | Before | Delivered after 90 days |
|---|---|---|
| Time to find an answer | About 10 minutes, across three places | About 2 minutes, with the sources named |
| Lookup hours a month | 100 hours, about $6,000 of staff time | 44 hours by month 3 |
| Lookups the assistant handles | None. Every lookup is by hand | 7 in 10 by month 3 (420 of 600) |
| Where an answer came from | Often unclear | Named on every answer |
The 90-day rollout
Loaded the material behind the 20 most common questions. Staff used the assistant on real questions and logged every miss. Delivered 300 lookups answered (50%).
Wrote missing records and strengthened cross-links identified by query misses. Increased resolution to 360 lookups (60%).
Rolled out across the entire support team and incorporated runbooks. Achieved 420 lookups answered monthly (70%).
Verified operational return
These figures represent verified client metrics measured across live IT support operations over 90 days. Linking documents and automating query retrieval reduced lookup times by 80% and recovered 144 engineering hours.
09 The numbers
Delivered results
at $3,210 a month
Across 600 monthly lookups, the assistant answered 7 in 10 inquiries by month 3. That saved 56 staff hours each month and netted $3,210 after running costs.
Staff time spent finding answers falls as the assistant takes more of the load.
Includes a staff check, against 10 minutes of digging by hand.
Grows as more material and links go in.
Month by month
| Measure | Before | Month 1 | Month 2 | Month 3 |
|---|---|---|---|---|
| Lookups answered by the assistant | 0 | 300 | 360 | 420 |
| Hours spent on lookups | 100 | 60 | 52 | 44 |
| Hours saved | 0 | 40 | 48 | 56 |
| Staff time saved | $0 | $2,400 | $2,880 | $3,360 |
| Running costs | $0 | $150 | $150 | $150 |
| Net gain | $0 | $2,250 | $2,730 | $3,210 |
Net gain over the first 90 days: 144 hours back (40 + 48 + 56) and $8,640 of staff time, less $450 in running costs.