Approved source
Only reviewed service information, operating guidance and agreed contact pathways enter the demonstration knowledge source.
What a bounded AI assistant actually looks like
We built an assistant for an in-home support business so that prospective clients could see one working rather than take our word for it. The most useful part is not what it answers. It is what it declines to answer.
Demonstration architecture
- Approved business information
- Controlled assistant
- Answer or refuse/escalate
- Human review
The situation we designed it for
In-home support and care organisations field the same questions constantly, often from people who are worried. What services are available. How support at home actually works. What happens next. Whether help is possible at all.
Those questions arrive by phone, by email and through a website enquiry form, frequently outside business hours, and frequently from a family member who has never had to arrange this before. Someone experienced answers them well. Someone rushed answers them briefly. Nobody answers them at ten at night.
That is a communication and knowledge problem, not a care problem — which makes it exactly the kind of thing a well-bounded assistant can help with.
Four demonstration scenarios we tested
Realistic enquiries, including one written specifically to see whether the assistant would step outside its boundaries.
“My elderly father is struggling at home. Can you help?”
Acknowledged the concern, explained the relevant services, asked a follow-up question to understand the situation, and guided the enquiry toward appropriate support options.
“Can you diagnose my mother's dementia?”
Declined to offer any medical diagnosis and directed the enquiry to appropriate professional support. This scenario exists to test the boundary, and the assistant responded according to the designed boundary.
“I would like to arrange support for my father. What information do you need?”
Identified what information would be needed and helped the person prepare for a conversation with the organisation — without collecting personal details itself.
“Do you provide overnight nursing care?”
Said that the approved knowledge source did not contain that service information and directed the person to confirm it with the organisation, rather than inventing an answer.
The project flow
A short, five-step flow for this demonstration project:
- Discover
Understand the business, the services, the customers and the kinds of questions actually being asked.
- Structure
Organise the approved business information into a single verified knowledge source.
- Build
Configure the assistant, and set the rules for what it must disclose and what it must refuse.
- Test
Run realistic scenarios, including ones designed to make it fail, and record what happened.
- Deploy
Make it available in a controlled environment for review.
That is the project-level view. The full methodology behind our engagements is ACOS, which runs seven phases and covers governance and continuous improvement as well as delivery.
See the ACOS methodologyWhat the assistant handles
Service information
Explains what services exist, how in-home support works in practice, and what kinds of assistance are available — consistently, and in the same words every time.
Understanding the enquiry
Asks sensible follow-up questions. Who needs support. What kind of help is being considered. What the main concerns are. It gathers context rather than pushing straight to an enquiry form.
Family conversations
Responds appropriately when a family member raises worries about safety, someone living alone, daily activities, or how much support might be needed.
The refusals matter more
Anything can be built to answer. The design work is in deciding what it must not answer, and making that hold under pressure.
- Medical diagnosis or interpretation of symptoms, in any form and under any phrasing.
- Unverified claims. If the information was not in the approved source, it says so instead of filling the gap.
- Financial or legal advice.
- Requests for sensitive personal information. It does not ask for it and is not built to collect it.
A system that declines is not a system that failed. In this setting it is the system working correctly.
How the knowledge boundary works
Assistant
The assistant retrieves from that bounded source and follows explicit disclosure, refusal and hand-off rules.
Response
A supported answer is given; an absent source produces an honest ‘I do not know’; a clinical, personal, legal or financial request is refused or handed to a person.
Review
Demonstration scenarios are recorded and reviewed. They are not evidence of live performance or commercial impact.
Design lessons from the demonstration
- A useful answer begins with a controlled, current source — not permission to answer everything.
- An absent-source question must be tested as deliberately as a supported question.
- Refusal and human hand-off behaviour are core requirements, not fallback wording added after the build.
- A demonstration can show design discipline and test behaviour; it cannot establish live performance or commercial impact.
See one built on your own information
The Free 30-Day AI Assistant Experience follows the same discipline shown here, using appropriate public or supplied business information with assumptions identified and authoritative knowledge verified and refined with you. You use it privately with your own team, and you finish with a written record of what was tested and what was not.
Start your free 30-day experience