Automation

Custom GPT & AI Assistant Development

A custom AI assistant is the fastest way to stop the same internal question being asked five times a week. It answers from your own SOPs, price lists, proposals and past projects, shows the document it took the answer from, and says so when the answer is not there.

  • Answers cite the source document
  • Access limited by team and role
  • Your own accounts and API keys

Who this is for

Teams where the knowledge exists but nobody can find it: the pricing rule that lives in one manager’s head, the SOP saved in a folder three levels deep, the proposal from last year that answers exactly this client’s question. Agencies, consultancies, clinics, manufacturers and any company with a long onboarding period. It is one of the builds under marketing and business automation services, and it uses the retrieval and guardrail approach set out on AI automation services.

It is customer-facing only by exception. An assistant your staff use can be a little wrong and still save time, because the person reading it knows the business. The same tolerance does not apply to a visitor on your website, which is why that is built differently under AI chatbot development.

What the assistant answers

  • Process questions. How a refund is handled, who approves a discount, what the onboarding checklist is, what to do when a client asks for something outside scope.
  • Commercial questions. Current prices, package inclusions, discount limits and payment terms, taken from the live document rather than from someone’s memory of it.
  • Past work. What was proposed to a similar client, which approach was used, what went wrong and what was learned, pulled from proposals, reports and retrospectives.
  • Technical and product detail. Specifications, compatibility, warranty terms and the answers your sales team currently gets by messaging an engineer.
  • Drafting from your material. A first-draft proposal, scope note, job description or client update in your own format, filled from the record rather than invented.
  • New-joiner questions, which are the same twenty questions every time and are the clearest early win.

How it is built, and what that rules out

The assistant retrieves from your documents and answers only from what it retrieves, with the source shown next to every answer so a person can check it in one click. That design choice is deliberate and it has consequences:

  • It will not answer from general knowledge when your documents are silent. It says the material does not cover it, which is the behaviour that makes it trustworthy.
  • It is only as current as the source. If the price list is out of date, so is the assistant. The build includes deciding who owns each document and how updates flow through.
  • Fine-tuning is usually unnecessary. Retrieval plus a good instruction is cheaper, easier to correct and does not need retraining when a policy changes. I propose fine-tuning only where a genuinely fixed style or format is required.
  • Access follows your permissions. Salary documents, client contracts and personal data are either excluded from the index or restricted to the roles that already see them.
  • It does not take actions. Answering is one job; acting in your systems is a different and more consequential one, covered on AI agents for business.

How a build runs

  1. Collect the questions first, not the documents. Two weeks of the questions your team asks each other tells us which twenty documents matter and which four hundred do not.
  2. Audit and clean the source set. Duplicates, superseded versions and contradictions are the main cause of wrong answers. Where two documents disagree, someone has to decide which is right, and that decision is part of the project.
  3. Decide the platform. A custom GPT or equivalent in the assistant tool your team already pays for, or a self-hosted retrieval setup where the documents cannot leave your systems. Cost, data location and who maintains it decide this, not preference.
  4. Build with citations on by default and a scored test set of real questions with known correct answers.
  5. Pilot with one team for two weeks, logging every question it could not answer. That log is the improvement plan.
  6. Roll out with an update routine: who refreshes which document, how often, and how the index is rebuilt.

What you receive

  • The assistant in your own workspace, on your own model account, with the document index in storage you control.
  • A cleaned, versioned source document set with a named owner for each document.
  • The instruction and configuration as an editable document, and the scored test set for checking future changes.
  • Role-based access rules and a written note of what was deliberately excluded from the index.
  • A question log and a monthly report of what it could not answer, which doubles as a list of the documentation you are missing.
  • Documentation, training and an optional monthly refresh of the index and instructions.

Pricing pointer

Setup is quoted on request; a single-team assistant on a clean document set sits in the lower to middle part of the range. A company-wide assistant with role-based access, several source systems and a self-hosted index sits at the top, and document cleanup is usually the largest line item. Maintenance is quoted on request; model and platform subscriptions are billed to your own accounts. If the tooling decision is the open question, start with AI consulting and tool setup.

Next step

Send me the five questions your team asks each other most often and where the answers currently live. I will come back within four business hours with whether an assistant helps, which documents would have to be fixed first, and a fixed price.

Frequently asked questions

Do you fine-tune a model on our data?

Usually not. Retrieval from your documents plus a well-written instruction is cheaper, updates instantly when a document changes, and lets every answer cite a source. Fine-tuning is worth considering only for a fixed output style or format, and it does not make a model know your prices any more reliably than reading the price list does.

Where does our data go?

That is decided during the build, in writing. The options are a hosted assistant on a business plan whose terms exclude training on your content, or a self-hosted index where documents never leave your infrastructure. Cost and convenience favour the first; regulated or confidential material often requires the second. Either way, accounts and keys are yours.

How accurate is it?

It is measured against a test set of real questions with known correct answers before launch, and the score is reported to you rather than described as high. Accuracy depends far more on the quality of the source documents than on the model, which is why document cleanup is part of the project rather than a prerequisite you handle alone.

What happens when our prices or processes change?

You update the source document and the assistant follows, which is the main advantage of this design over training answers into a model. The build assigns a named owner to each document and sets a refresh routine, so a stale answer traces back to a stale document and a person who is responsible for it.

Can different teams see different things?

Yes. The index is segmented and access is granted by role, so sales sees the pricing and proposal material, operations sees the SOPs, and documents such as contracts or salary records are excluded or restricted. What is excluded is written down at handover so there is no ambiguity about what the assistant can reach.

Can it work in Nepali?

It handles Nepali and mixed Nepali and English questions, and answers in the language asked. Retrieval works better when the source documents are consistent in one language, so where your SOPs are a mixture, part of the cleanup is deciding which language each document lives in.

Ready to talk about your project?

A free 30-minute call, a straight answer about what would move the numbers, and a written proposal within 48 hours if we are a fit.