Customer support assistants
First-line resolution for the questions your team answers repeatedly, with ticket creation and handover when needed.
The chatbots people hate are the ones that guess. A useful assistant is narrow, grounded in content you control, honest when it does not know, and quick to hand over to a person. That is a design decision long before it is a model decision.
An assistant that tries to answer anything will be wrong often enough to lose trust permanently. One scoped to a defined body of knowledge, with a confident path to a human, earns its place quickly.
Responses drawn from your documentation with the source shown, so a user can verify what they are told.
When the answer is not in scope, the assistant says so and routes onward instead of improvising.
Users only get answers from content they are entitled to see — critical for internal assistants.
An evaluation set built from real questions, re-run whenever the prompt, model or content changes.
First-line resolution for the questions your team answers repeatedly, with ticket creation and handover when needed.
HR, IT and policy questions answered from your own handbooks, with access respected per employee.
Search across contracts, manuals and historical tickets that returns an answer rather than ten documents.
Beyond answering: check an order, book a slot, raise a request — with confirmation before anything is changed.
One assistant reachable from your site, app, help centre and messaging platforms without rebuilding it each time.
What people ask, what went unanswered, and which documentation is missing — the loop that makes it better over time.
A production assistant is a retrieval system with a conversation on top. Most quality problems trace back to retrieval, not to the model, so that is where the engineering effort goes.
Chunking, embedding, hybrid keyword and vector search, and reranking — tuned against your actual content.
Follow-up questions resolved against history so a user is not forced to restate context each turn.
Read-only lookups by default; anything that writes requires explicit confirmation.
Every conversation logged with the sources retrieved, so a bad answer can be diagnosed rather than guessed at.
Capability in isolation changes nothing. These are the systems this normally has to reach in order to be useful.
Create, update and escalate tickets with the conversation attached.
Help centre, wiki, document store and past resolved tickets indexed together.
Single sign-on so answers respect the permissions the user already has.
Live order, account and entitlement lookups during the conversation.
A support question from arrival to resolution, including the case where the assistant should not answer.
The same six stages apply to every AI build we run. The evaluation set is created before development starts — without it there is no way to tell an improvement from a regression.
Is a model the right instrument here? A deterministic rule is cheaper, faster and easier to audit when the problem allows one. We say so when it does.
Sources identified, access agreed, content parsed and indexed. Data quality sets the ceiling on everything after this point.
Real examples with known-good outcomes, collected before any development. This is what makes later decisions measurable.
A working slice scored against the evaluation set — accuracy, latency and cost per request all measured rather than estimated.
Wired into the real system with permissions, rate limits, spend caps, fallbacks and human approval wherever the action is irreversible.
Quality, cost and drift tracked after launch, with the evaluation set re-run whenever a prompt, model or data source changes.
These are not add-ons quoted separately. They are the conditions under which a system like this is allowed near production data.
The assistant can only surface content the signed-in user is already allowed to read.
Retrieved content is treated as data, never as instructions, with tool access constrained regardless of what a document says.
Provider configurations that exclude your data from model training, confirmed in writing.
Retained transcripts with retrieved sources, so any answer can be reconstructed and explained.
Sizing, stock, delivery and returns answered instantly, with live order lookup where the customer is signed in.
Employees get an answer from the current handbook instead of forwarding an email to a shared inbox.
Answers restricted to approved wording, with regulated topics routed straight to qualified staff.
Appointments, preparation instructions and practical guidance — deliberately excluding anything clinical.
Availability, viewing arrangements and property detail answered around the clock, qualified before it reaches an agent.
Site teams query drawings, specifications and method statements by asking rather than searching.
By constraining what it can answer from and measuring the result. Responses are grounded in retrieved documents with citations shown, low-confidence cases are escalated rather than guessed, and an evaluation set of real questions is re-run on every change. Perfect accuracy is not achievable; knowing your accuracy and having a safe failure path is.
A scoped assistant over an existing, reasonably organised knowledge base is typically six to ten weeks including evaluation and integration. The variable is content quality — if the documentation is scattered or contradictory, fixing that is the project, and we will say so early rather than after launch.
Cost is driven by conversation volume, model choice and how much context each request carries. We measure cost per conversation during the prototype and design to a target — retrieval strategy and caching move that number substantially. You will have a per-conversation figure before committing to a full build.
Yes. Checking an order, booking an appointment or raising a request are all reasonable. We keep read operations open and require explicit user confirmation before anything is written or changed, so a misunderstanding cannot quietly cause damage.
Send us your top fifty support questions. We will tell you honestly how many an assistant could handle well.