Assistants and search inside your product
Answers from your platform's own records and documents, with sources. You get retrieval over your data, user permissions enforced on every query, and a test set built from real user questions.
Your users tried the AI demo. Now it has to work on your data, inside your permissions, at a cost you can plan. Assistants and search inside your product, generative features in the platform you already run, and agents that act through your own APIs, built on your data and run in production with the rest of your system.

A working AI demo inside a product takes a week. The trouble starts when real users, real data and real costs arrive.
The demo ran on clean sample data
Your production data has gaps, old formats and the edge cases your domain collected over years. A feature that looked sharp on test data gives wrong answers on yours.
It sits beside the product, not inside it
A chat window added on the side cannot see permissions, account state or your business rules. Users try it twice and go back to the screens they know.
Nobody owns it after launch
Model versions change and vendors change prices. Without someone checking answer quality and cost every month, the feature gets worse and users notice first.
Usage costs nobody modelled
Model costs grow with every request. A feature that cost little in the pilot becomes a real line item once every customer uses it every day.
Each one starts from a job your users already do in your platform. The AI part is one component of the system, built and released like the rest of it.
Answers from your platform's own records and documents, with sources. You get retrieval over your data, user permissions enforced on every query, and a test set built from real user questions.
Drafts, summaries and structured output where users already work, such as a proposal from a form. Users review before anything is saved, and running cost is measured before launch.
An agent with one defined job that works only through the APIs your product already has. Every action is validated and logged, and a person approves changes to money, bookings or customer data.
How it runs
It starts with the automation review: one workflow in your product, the users who do it, and what a correct answer looks like. You get a go or a clear no-go, with the reasons.
A working version runs against a copy of your data, in your own environment. It is measured on the questions your users actually ask.
The feature goes into your codebase through the same review and tests as everything else. Permissions, logging and fallbacks are part of the build.
The feature is released like any other change, after it passes the test set built from real user questions. A person approves agent actions that change money, bookings or customer data.
We do not have a client product yet where we built the AI part. The system we can show you is ours: the content platform behind this website, where AI agents draft and update content.
AI agents create and update pages, case studies and FAQ through our command line tool. Every agent write is saved as a draft. There is no path for an agent to publish on its own.
Before a draft is saved, deterministic checks run: required fields, length limits, page structure and banned phrases. A failed check returns a structured error, and the agent must fix the draft. The model never checks itself.
Publishing, discarding and live writes pass only with an explicit approval phrase typed by a person for that specific batch. Earlier approvals do not carry over, and a kill switch stops all live writes.
An agent in your product works the same way: it proposes, code validates, a person approves what matters, and every action is recorded.
These are the risks we plan for on every AI feature. None of them is solved by picking a better model.
| Risk | Likelihood | Mitigation |
|---|---|---|
| Customer data reaches a model vendor | Depends | We build on Azure AI Foundry. Microsoft's terms for Azure AI Foundry state that prompts and documents are not used to train the models. Each request is designed to send only the data it needs. |
| Users see data they should not | Medium | Search and assistants check the user's permissions in your system before retrieval, not after. A user's answer can only use records that user could already open. |
| An agent takes a wrong action | Medium | Agents get a short list of allowed API calls, each validated by code. Anything that changes money, bookings or customer records waits for a person to approve it. |
| EU AI Act obligations | Depends | Each feature is mapped to its risk category early, AI output is labelled where your users must know, and you keep the documentation and logs you need. Legal sign-off stays with your counsel. |
| Quality drops after a model update | Medium | Before a feature uses a new model version, it runs against a test set built from real questions in your domain. |
| Cost grows faster than usage | Medium | Running cost per request is measured on your real data during the prototype and projected at your expected usage. Where a smaller model scores the same on your test set, it is the cheaper choice. |
A project can test a concept, prove a first phase, or support a funding decision. Production systems need a team that keeps the goals, context, and delivery moving.
Useful for a concept test, funding case, MVP, or one defined business need. We scope the useful part, build it, and use the result to decide what comes next.
For systems that matter day to day, a focused team brings smoother delivery, steadier knowledge, and ongoing work around your business goals.
What founders and CTOs ask before they add an assistant, a generative feature or an agent to a platform that already has users.
AI product development is one of four Applied AI lines. All four start with the same automation review on one workflow.
