Discuss your project
AI Solutions

Generative AI development and custom AI assistants

General-purpose chatbots know a lot about the world and nothing about your organization. We build generative AI assistants and applications that work from your own policies, procedures, product information, contracts and records, show where each answer came from, and only use information the person asking is allowed to see.

Discuss your project How the work is delivered

A good fit if

  • Staff spend hours searching policies, procedures, manuals or past project files for answers.
  • You want an assistant inside your own application, portal or intranet, not a separate chatbot.
  • Answers must come only from approved sources, with citations people can check.
  • Information is spread across SharePoint, file shares, an ERP or CRM and other systems.
  • You need control over where data is processed, which model is used and what is logged.

Another approach may suit you better if

  • Microsoft 365 Copilot already covers the need and your content lives in Microsoft 365. Start with our Copilot service.
  • The answers you need are not written down anywhere yet. That documentation comes first.
  • You need a system that must always be right with no human check, such as final legal or medical advice. We will not build that.

What this service is

We design, build and support generative AI assistants and applications for organizations that need answers from their own information. Typical examples are a staff assistant that answers HR, finance and IT policy questions, a product assistant that helps sales and service teams find specifications and procedures, a contract assistant that finds clauses across agreements, and AI features inside your own customer or partner portal.

These solutions are built on large language models, but the important work is everything around the model: choosing and preparing sources, retrieving the right passages, enforcing permissions, testing accuracy and connecting the result to the places people already work.

How it works

Most of our assistants use retrieval-augmented generation (RAG). When someone asks a question, the system first searches your approved sources for the most relevant passages, then asks the model to answer using only those passages and to cite them. Because answers come from your documents, they reflect your policies and products rather than general internet knowledge, and people can open the source to check. Our guide to how AI assistants answer from your documents explains the approach in plain language.

Where an assistant needs live data, such as an order status in Prometheus ERP or a customer record in your CRM, it calls those systems through their APIs with the user's own permissions. When the assistant needs to take actions rather than answer questions, we design it as an AI agent with approval points.

Built to be trusted

  • Permissions first. The assistant uses the same access rights as your source systems, so it never answers from content the user could not open.
  • Citations on every answer. People can see where an answer came from and judge it for themselves.
  • Tested before launch. We build an evaluation set from real questions and measure accuracy before anyone relies on the assistant.
  • Guarded against misuse. We test for prompt injection, attempts to extract other users' data and off-topic use, and log activity in line with your retention rules.
  • Costs in view. Model usage is metered, so we estimate and monitor running costs per question or per user.

Platforms and privacy

We build on Microsoft Azure AI services, Amazon Bedrock, Google Cloud Vertex AI and open-weight models hosted on servers you control, and connect to Microsoft 365, SharePoint, file shares, databases, ERPs and CRMs. We choose enterprise services whose terms exclude your data from model training. Naming a product does not mean we are a partner of its vendor.

PIPEDA and provincial privacy laws, such as Quebec's Law 25, apply to personal information used in an AI system, including the documents it searches. We minimize the personal information an assistant can reach, document where data is processed and work with your privacy lead. This is general information, not legal advice. See also AI governance and security.

What is included

The exact list is agreed in writing for each project. These are the usual deliverables and the usual boundaries.

Typical deliverables

  • A use case definition with users, questions, sources, permissions and success measures.
  • A source review and ingestion design for documents and data, including how they stay up to date.
  • Retrieval-augmented generation (RAG) that finds the relevant passages and answers with citations.
  • Permission-aware search, so each user only gets answers from content they can access.
  • Integration into your intranet, Microsoft Teams, web portal or line-of-business application.
  • Guardrails for off-topic questions, personal information and prompt injection.
  • An evaluation set of real questions with expected answers, and accuracy results before launch.
  • Logging, usage and cost monitoring, and an administrator guide.
  • Documentation, training and a runbook for the people who operate the solution.

Not included unless agreed separately

  • AI model, platform and cloud usage charges, which you pay to the provider.
  • Writing or restructuring the source documents themselves, unless agreed.
  • Ongoing tuning and support after handover, unless agreed in a support or managed services agreement.
  • Legal review of the answers your assistant gives.

What we will need from you

Most delays in this kind of work come from access and decisions, not from the technical build. Knowing these early keeps the project predictable.

  • A business owner who can decide whether answers are good enough.
  • Access to the source documents and systems, with sensitive information identified.
  • Example questions from real users and the answers they should get.
  • Your privacy, security and data residency requirements.
Delivery

How the work is delivered

Each stage ends with something you can review before the next one starts.

  1. Define

    Agree the users, the questions the assistant must answer, the approved sources, the permissions model and how success will be measured.

    Output: Use case definition and evaluation plan.

  2. Prove

    Build a working prototype on a representative set of your documents and test it against real questions.

    Output: Prototype and accuracy results.

  3. Build

    Connect all approved sources, add permission-aware retrieval, guardrails and logging, and integrate the assistant where people work.

    Output: Production-ready solution.

  4. Pilot

    Release to a pilot group, review answers and feedback, and tune sources, prompts and retrieval.

    Output: Pilot report and improvements.

  5. Launch and monitor

    Roll out, train users and monitor accuracy, usage and running costs against agreed thresholds.

    Output: Live assistant, runbook and monitoring.

Testing and handover

  • The assistant meets the accuracy threshold on the agreed evaluation set before launch.
  • Every answer shows its sources, and unanswerable questions get an honest "I could not find that".
  • Permission tests confirm users cannot retrieve content they are not allowed to see.
  • Prompt injection and data leakage tests are run and documented.
  • Running costs are estimated per user or per question, and monitored after launch.
  • Administrators can add, update and remove sources without our help.

What affects the cost

We do not publish package prices. Each estimate is based on an agreed scope, in Canadian dollars, with taxes shown separately. These are the things that move the number most:

  • The number, size and variety of document sources and systems.
  • How strict the permissions model is and how it is enforced.
  • The accuracy target and the size of the evaluation set.
  • Where the assistant runs and how it is integrated.
  • Model and cloud usage, which grows with the number of users and questions.

Questions buyers usually ask

How is this different from ChatGPT or Microsoft 365 Copilot?

General assistants answer from their training and whatever you give them. A custom assistant answers only from the sources you approve, applies your permissions, shows citations and fits into your own systems. Many organizations use both.

Will the assistant make things up?

Generative AI can produce confident but wrong answers. We reduce that by answering only from retrieved sources, showing citations, testing against real questions and designing the assistant to say when it cannot find an answer. People should still check answers that matter.

Which AI model do you use?

We choose a model based on accuracy on your questions, cost, data handling terms and where it can run. Options include models from Microsoft Azure, Amazon Bedrock and Google Cloud, and open-weight models such as Llama, Qwen, DeepSeek, GLM and Mistral hosted on servers you control. See private AI and self-hosted models.

Can it run in Canada?

Many AI services can be deployed in Canadian cloud regions, though model availability varies by provider and region. Where residency matters we choose services and regions accordingly and document where data is processed.

How do you keep answers up to date?

Sources are re-indexed on a schedule or when documents change, and retired documents are removed. Administrators can see which sources are included and when they were last updated.

Service

AI agents for business processes

AI agents that complete routine steps in a business process across your systems, with approval points and a full audit trail.

Service

Private AI and self-hosted AI models

Open-weight AI models deployed on GPU servers you control, in your own cloud account or data centre, so prompts, documents and outputs stay on infrastructure you govern.

Have documents nobody can find answers in?

Tell us who needs the answers, where the information lives and how sensitive it is. We will reply to arrange a conversation and suggest how to prove the idea on your own content.

Scope an AI assistant