Custom AI assistant development: Answers from your own documents. Every one with a source.
I build custom AI assistants that answer your staff and customers from your company knowledge, show the source, and respect who may see what. Live in about 3 weeks, and the code is yours.
- 126 systems in production
- Live in about 3 weeks
- The code is yours
3 sources you may open
Source
Refund policy v4, page 2
Google Drive · visible to the support team
From a documentation assistant I shipped for a developer tools platform
Watch it answer
It cites the source. It knows what you may see.
A support question, slowed down so you can follow it. Two sources cited, one contract withheld, and a draft a person sends.
knowledge-assistant · #support
Example
Sam K. · support team, in Slack
Can a customer on an annual plan get a refund after 45 days? And what does the Northwind contract say?
Searching the sources Sam is allowed to open
Answer
Yes. Annual plans can be refunded within 60 days of renewal, prorated for the unused months1. Two past tickets were handled exactly this way2.
- 1 Refund policy v4, page 2 · Google Drive
- 2 Tickets #7731 and #7904 · Help desk
The Northwind contract is restricted to Finance, so I cannot show it to you. Ask Priya in Finance.
Draft reply · for you to send
Hi Dana, you are within the 60 day window, so I can refund the unused months of your plan...
The line I draw
It answers and drafts. It never acts on its own.
What it does
- Answers from your documents, with a source on every answer
- Shows each person only what they could already open
- Says I do not know, and hands the question to a person
- Drafts the reply, the summary or the brief for you to send
What it never does
- Changes a record in your systems
- Sends anything on its own
- Quotes a document you are not allowed to see
- Guesses when your docs are silent
Lives where people already ask
- Slack and Teams
- Your website
- Your intranet
- Email drafts
- ChatGPT and Claude
The short answer
What is custom AI assistant development?
Custom AI assistant development is building an assistant that answers your staff or customers from your own company knowledge, with a source on every answer. It reads your documents, follows the access rules they already have, says I do not know when they do not cover a question, and drafts replies or summaries for a person to send. It changes nothing in your systems. When it needs to act, that is an agent.
Scripted chatbot
Follows a decision tree of canned replies.
Fine for a few questions that never change.
ChatGPT or a custom GPT
Answers from the open internet or a few uploaded files, with no access rules.
Enough for a small team and a handful of shared files.
Custom AI assistant
Answers from all your sources with a citation, and shows each person only what they may see.
Right when knowledge is scattered, restricted or customer facing.
Where assistants earn their keep
Six places people ask it most
Customer support answers
Answers product and policy questions on your site or help desk with a link to the doc, and hands the hard ones to a person.
45%Fewer Tickets (Month 1)See the build
Policies and SOPs
Staff ask in Slack or Teams how leave, expenses or a process works, and get the answer with the policy attached.
New hire onboarding
New people ask the questions they would rather not ask twice, and every answer points to the document behind it.
Sales and proposal answers
Finds what you already said in past proposals, security answers and product specs, and drafts the reply for review.
Contract and document lookup
Finds the clause, quotes it with the page, and drafts a summary for the lawyer or manager who decides.
15minAvg Review TimeSee the build
Website product questions
Answers visitors about plans, features and orders from your published docs, then hands off to sales or support.
Permissions and data
Your access rules. Your cloud.
Six controls ship with every assistant, so it can read your real documents without leaking them to the wrong person.
- Access rules from each source
- A source on every answer
- Eval set from your real questions
- Docs treated as data, not orders
- Spend cap per user
- Every question logged
Your repo. Your index. Your keys.
How a build runs
From scattered docs to a live assistant
Delays rarely come from the model. They come from messy documents, unclear access rules and waiting for system access. I sort out all three in week one.
Week 1
Audit the knowledge and collect real questions
I map where answers live, who may see what, and which documents are out of date. Then I gather real questions from your team and your tickets, each with the answer a person would give.
You get: A knowledge health report and an eval set built from your real questions
Weeks 1 to 2
Connect the sources, keep the permissions
Each source is indexed with the access rules it already has. Search combines meaning with exact terms, so product codes and error messages are found as reliably as plain questions.
You get: A searchable index in your cloud that mirrors your access rules
Weeks 2 to 3
Build the assistant and pass the eval set
Answers with a citation on every claim, a confidence threshold, a clear I do not know with a handoff to a person, and drafting for replies and summaries. The eval set runs on every change.
You get: A staging assistant passing the eval set, with a score you can read
Week 3
Soft rollout to one team
One team, or a slice of site visitors, asks real questions. I tune the threshold on what they actually ask and watch the gaps it finds in your docs.
You get: A rollout report and a written go or no go call
Launch and 30 days
Full launch, tuning, and the keys
Live for everyone, with 30 days of tuning after launch. Your team gets a runbook for adding sources and a recorded walkthrough for whoever will own it next.
You get: Repository, prompts, index pipeline, eval set and runbook, all in your accounts
What is included
What you walk away with when the assistant goes live
The assistant, in your cloud
Running in your account and calling models through your own API keys, so the usage bill goes to you at cost.
Connectors that keep permissions
Drive, SharePoint, Confluence, Notion, your help desk or a folder of PDFs, each document carrying its own access rules.
Answers with sources
A clickable source on every answer, a confidence threshold, and a handoff to a person when the docs do not cover it.
An eval set and report
Built from your real questions with agreed pass criteria. Rerun before every model swap or prompt change.
A gaps dashboard
What people asked, what it could not answer, and which documents to write or fix next.
Runbook and handover
Prompts under version control, the refresh schedule, a guide for adding sources and a recorded walkthrough.
The stack
Model agnostic, with the index in your own database
I pick the model on your eval results, not on a logo. Retrieval quality matters more than the model, so most of the build goes into search, permissions and the eval set.
Claude models
Answering and drafting, picked on eval results
OpenAI and open models
Where the eval says they win on cost or accuracy
Postgres and pgvector
The index, in your own database
Hybrid search and reranking
Meaning plus exact terms like error codes and SKUs
MCP
So people can ask it from ChatGPT or Claude
Slack, Teams or a web widget
Where people already ask questions
AWS or your cloud
Hosting, queues and secrets in your account
Tracing and evals
Langfuse, Logfire or your existing stack
How engagements work
A fixed scope, quoted after one call
The price depends on how many sources it reads, how complex your access rules are and where it has to live, so I quote it once I have seen your documents.
- Most teams start here
One knowledge base, one channel
About 3 weeks
One team’s documents, answered where that team already works.
- Knowledge audit and eval set
- Answers with sources and a handoff
- Soft rollout and 30 days of tuning
Fixed scope after a call
Many sources, many teams
4 to 6 weeks
Several sources, access rules per team, or staff and customer channels together.
- Everything in the first shape
- Connectors that keep access rules
- A gaps dashboard per team
Fixed scope after a call
Ongoing partner
Month to month
For an assistant that keeps growing, or one that is ready to become an agent.
- New sources and channels
- Model swaps checked against the eval set
- Upkeep and monitoring
Fixed scope after a call
Your options
Custom GPT, off the shelf tool or a custom build?
A custom GPT or a Claude Project is often enough, and when it is, I will tell you. A custom build earns its cost when permissions, many sources or customers enter the picture.
Best when
- Custom build with me
- Many sources, access rules per team, or answers going to customers.
- A custom GPT or Project
- A small team, a handful of files, and everyone may see everything.
- Off the shelf tool
- Your knowledge already lives in one suite and its own assistant covers it.
- In house, or nothing
- You have engineers free to own it for years, or the questions are rare enough to leave alone.
Permissions
- Custom build with me
- Mirrors the access rules of each source, so people only see what they could already open.
- A custom GPT or Project
- Anyone who can use the GPT can reach what was uploaded to it.
- Off the shelf tool
- Usually follows the suite’s own permissions, inside that suite only.
- In house, or nothing
- As good as the engineer who builds it.
Sources and freshness
- Custom build with me
- Drive, SharePoint, Confluence, Notion, your help desk and PDFs, refreshed on a schedule or when a doc changes.
- A custom GPT or Project
- Files you upload by hand, and upload again when they change.
- Off the shelf tool
- Strong inside its own suite, weaker for anything outside it.
- In house, or nothing
- Whatever your team finds time to connect.
Accuracy you can check
- Custom build with me
- An eval set of your real questions, a score before launch, and a source on every answer.
- A custom GPT or Project
- Spot checks by whoever set it up.
- Off the shelf tool
- Vendor defaults. You rarely see a score.
- In house, or nothing
- Depends on whether anyone builds an eval set.
Where it lives
- Custom build with me
- Slack, Teams, your website, your intranet, email drafts, or inside ChatGPT and Claude.
- A custom GPT or Project
- Inside ChatGPT, for people with seats.
- Off the shelf tool
- Inside the vendor’s app.
- In house, or nothing
- Wherever you build it, or nowhere.
Cost and ownership
- Custom build with me
- A fixed build quote, model usage billed to your account at cost, and the code is yours.
- A custom GPT or Project
- A seat per user and an afternoon of setup. Nothing to own.
- Off the shelf tool
- A subscription per seat, and the index stays with the vendor.
- In house, or nothing
- A salary, or the hours your team keeps losing to searching.
Fit check
Is a custom assistant the right call?
A good fit when
- Your answers are spread across several tools, and people keep asking the same questions.
- Some documents are restricted, and the assistant has to respect that.
- Answers go to customers, so every reply needs a source and a safe handoff.
- You want the code and the index in your own accounts.
Not the right call when
- A few files and a small team. A custom GPT or Claude Project is enough, and cheaper.
- There is little written down yet. The docs come first.
- It has to update records or send on its own. That is an agent, and I build those too.
- You are still deciding where AI fits. Start with the AI Audit instead.
Questions
What people ask before a build starts
A custom AI assistant answers from your own company knowledge (policies, product docs, past tickets, contracts) and shows the source for every answer. ChatGPT on its own does not know your business and cannot tell you which document an answer came from. A custom assistant also follows your access rules and says I do not know when your documents do not cover the question.
Go deeper
Guides and tools on knowledge assistants
- Knowledge agent packageThe fixed scope version: one knowledge base, one channel.
- AI agent developmentWhen the assistant needs to update records or send on its own.
- What is RAG?How retrieval works, in plain business terms.
- ChatGPT for business vs custom agentsWhen an off the shelf tool is enough.
- Custom vs off the shelf chatbot costWhat each option costs over time.
- Agentic RAG in productionThe engineering behind retrieval that holds up.
- AI agent cost calculatorA first read on what it costs to run.
- Build vs buy calculatorCost out building against buying over time.
- RAG, definedRetrieval augmented generation in one page.
Tell me where your answers live. I will tell you if it needs a build.
Three quick questions or a short call. You get a straight answer on whether a custom GPT, an off the shelf tool or a custom assistant fits, and a fixed scope if it is worth building.