A customer assistant needs to know what it may answer, what it must decline and how a person takes over. We plan AI chatbot work for London businesses around an approved knowledge source and a testable support journey. Softglimmer works remotely from Ranchi, with scope and review windows agreed with your team.
Recognise the visitor’s question.
Use approved information.
Answer within scope or hand over.
An example journey, shaped around your requirements.
Imagine a visitor asking about an exception to a published policy. An assistant should not invent a promise just to finish the conversation. The workflow below is illustrative: the business approves the information, defines the limits and remains responsible for decisions that need a person.
Select the pages or documents the assistant may use and name an owner for updates. Test questions whose answer has changed or is absent. The system needs a clear uncertainty response, not a confident guess dressed as customer service.
Answering a product question is different from changing an order or accessing an account. Scope any connected action explicitly, with authentication, confirmation and a record of its result. Begin with an informational assistant if those responsibilities are not ready.
Recognise the visitor’s question.
Use approved information.
Answer within scope or hand over.
Illustrative planning example. The final scope follows your requirements.
Discuss these decisions before choosing the final implementation.
Decide whether the assistant handles only approved public answers or needs account data. Private records require an explicit access model and appropriate provider arrangements.
Define where an unresolved conversation goes and who responds. Include the question, relevant context and the reason for handover without collecting unnecessary personal information.
Use a reviewed question set with correct, incomplete, conflicting and adversarial examples. Define what failure means and which outcomes stop release. A fluent demonstration is not a reliability test.
The proposal names the records, journeys and checks included in your release.
Approved sources, update owners, excluded topics and the boundary between automated information and human decisions. Your team approves policy and privacy requirements.
A usable chat interface with uncertainty, unavailable service and human-contact states. Define whether a handover creates a ticket, sends an agreed notification or offers a visitor-controlled contact path.
A representative test set, observed failures, cost limits and monitoring responsibilities. Document the provider accounts, retention settings and conditions for pausing the assistant.
Use these scenarios to agree the expected behaviour and the person responsible for resolving an exception.
Confirm that the assistant explains the limit and offers the agreed next step. It should not invent a price, availability, policy exception or guarantee.
Test how the knowledge source is refreshed and how obsolete answers are detected. Name the person responsible for correcting source material and approving the new result.
A conversation must not bypass account permissions or accept instructions hidden in retrieved content. Review the scope and controls before enabling connected actions.
Useful answers about scope, responsibilities and remote delivery.
No. The scope includes evaluation, uncertainty behaviour and monitoring. We test known failure cases and keep a human route available rather than treating fluent wording as proof of correctness.
That requires a separate review of access, provider arrangements, permitted data and retention. Do not upload confidential material to a consumer tool by default. Your business approves the relevant privacy and contractual requirements.
Yes. Softglimmer is based in Ranchi and works with London teams remotely. We agree review windows and account ownership, including seasonal time changes, before delivery.