Chatbots

Answers that can be checked against your information.

A useful business chatbot needs reliable source material and a clear job. We work on the information it can use, the questions it should answer and the point at which a conversation needs clarification, a business action or a person.

QUESTION → UNDERSTANDING → ANSWER

Where this capability helps.

01

Employee knowledge

Help staff navigate approved procedures and documentation, with supporting material they can inspect.

02

Customer guidance

Explain a service or process using maintained business information, then guide the visitor to an appropriate next step.

03

Structured intake

Gather the details a request needs through conversation, confirm them and pass a useful brief into the agreed workflow.

An answer, its source and its boundary.

The sample equipment guide below supports a question about replacement requests. It says nothing about temporary equipment. A grounded conversation should preserve that distinction.

ILLUSTRATIVE DESIGN / SAMPLE INFORMATION

EMPLOYEE

“How do I request replacement equipment?”

EXAMPLE ANSWER

Submit an equipment request with the asset reference and a description of the issue. Your team lead will review it.

Source: sample equipment guide, §3
A follow-up the source does not answer

“Can I borrow a laptop while I wait?”

The sample guide does not cover temporary equipment. A useful response makes that gap clear and refers the question to the equipment team.

Showing a source link is useful only if that source supports the answer. Retrieval quality and answer accuracy need to be checked together.

See this in a employee knowledge workflow. When an answer needs a record check or an approved action, connect the conversation to an agent workflow.

An assistant your team can keep informed.

  • A conversation built around maintained sources and the questions it should answer.
  • A way to inspect supporting information, clarify gaps and hand over to a person.
  • Representative question tests and a process for reviewing answers as sources change.

Decisions that shape the system.

  1. 01

    Establish which information is authoritative

    We identify the documents that should govern an answer, who maintains them and how updates reach the assistant. Conflicting versions and obsolete guidance need handling before they become confident replies.

  2. 02

    Design retrieval around real questions

    Document structure, terminology and access permissions affect what is found. We test representative questions against the retrieved passages before assessing the final answer.

  3. 03

    Give the conversation a useful next step

    An assistant may answer, ask one necessary question or hand over with context. We define these paths around the service, including where a conversation becomes an agent workflow with permitted tools.

  4. 04

    Fit the interface to its users

    We agree the channel, response length, source presentation and handoff experience. An internal procedure assistant and a customer-facing guide have different information and interaction needs.

How we would evaluate it.

Criteria are agreed for the intended task and tested on representative material. The items below describe the evaluation approach.

Retrieval
Check whether the relevant, permitted source material is found for each question.
Supported answers
Review factual correctness, whether cited passages support the response and how conflicting information is handled.
Conversation handling
Include follow-ups, missing details, unanswered questions and successful handoffs.
Maintenance
Change or remove source material and verify that subsequent answers reflect the update.

Practical questions.

Can it work with private company information?

We assess the sources, user permissions and processing arrangements together. Retrieval must respect the access rules agreed for the application. Hosting, administrative access, conversation logging and retention are defined for the project.

Does it need to be fine-tuned on our documents?

Often the first step is retrieval from maintained sources. Fine-tuning serves different needs, such as a particular response format or specialised behaviour. We choose the approach after testing what the task requires.

Can it update records as well as answer questions?

It can be designed to request a permitted tool action. We treat that as an additional workflow with explicit inputs, confirmation and permissions. An answer grounded in a document does not itself authorise a system change.

What would make this useful for you?

Start with a maintained set of documents and the questions your users ask. We can assess the information, conversation and handoff together.

We build client solutions using established models. Integrations, data access and operating arrangements are agreed for the project. Our proprietary model programme explores future deployment options.

Discuss your projectExplore this in a walkthrough