Solution
AI customer experience
A chatbot, a voice agent and a knowledge base that share one grounded source of truth — with citations, measured accuracy and a fast route to a human.
Most AI customer experience projects stop at the demonstration. Something impressive is built, everyone agrees it is remarkable, and it never ships because nobody can say how often it is wrong or what happens when it is.
This engagement is built around the parts that make deployment possible: a curated knowledge corpus, an evaluation set with measured accuracy, permissions enforced at retrieval, explicit refusal behaviour, and a handoff to a person that works on request rather than after three failed attempts.
Why this is sold as one engagement
The components genuinely depend on each other. A chatbot without a grounded knowledge base invents answers. A knowledge base without CRM integration cannot answer account-specific questions. A voice agent without a handoff path traps exactly the callers who most need a person. Buying them separately produces three systems that each work in isolation and contradict each other in practice.
The measurement question is what separates deployment from perpetual pilot. Without a fixed test set of real questions with known-correct answers, changes are made on the basis of anecdotes and nobody can say whether accuracy is rising or falling. Building that set is unglamorous and it is the difference.
Transparency obligations are also converging. The EU AI Act requires disclosure that a person is interacting with an AI system, and several other jurisdictions point the same way. Disclosure is designed in rather than added after a complaint.
What is included
- Knowledge corpus curation
- Deciding what is authoritative, what is superseded and what should never be indexed — which is most of the work.
- Retrieval system
- Hybrid keyword and semantic search with reranking, measured separately from answer quality.
- Website chat assistant
- Grounded in your published material, with citations and an explicit refusal path.
- Voice agent
- Bounded telephone handling with disclosure, barge-in support and warm transfer to a person.
- Internal knowledge assistant
- Staff-facing answers with permissions enforced at retrieval rather than filtered afterwards.
- CRM integration
- Conversations logged against contact records, and account-specific answers where entitlements allow.
- Human handoff
- Immediate on request, with full context passed so the customer does not repeat themselves.
- Evaluation harness
- A real question set with scored accuracy, re-run on a schedule to catch model drift.
- Logging and cost controls
- Full request and tool-call logging with defined retention, plus caching and context budgets.
- Monitoring and review
- Accuracy, escalation rate, override rate and cost per interaction tracked continuously.
How it runs
Curate the corpus
Decide what is authoritative and what is not. This requires someone who knows the material and it determines the ceiling on everything that follows — no retrieval architecture recovers from an index full of superseded drafts.
Build the evaluation set
Real questions people actually ask, with correct answers, assembled before any prompt is written. This is the single practice that most distinguishes systems that ship from demonstrations that impressed someone once.
Engineer retrieval
Chunking tuned to document structure, hybrid search, reranking, and permissions applied to the candidate set rather than to the output. Retrieval quality is measured separately, because conflating it with answer quality hides where the problem is.
Ground, cite and refuse
Answers constrained to retrieved material with citations, and an explicit refusal path when nothing relevant is found. Refusal behaviour is tested deliberately rather than hoped for.
Add channels
Web chat first, then voice if the use case is bounded enough, with disclosure and handoff designed before the happy path.
Launch narrow and monitor
A share of traffic first with every transcript reviewed, widening only once escalation and accuracy hold. Then scheduled re-evaluation, because models change underneath you.
What changes
- Answers with sources attached
- Every response cites the passage it used, so customers and staff can verify rather than trust blindly.
- Measured accuracy
- A number against a real question set, so you know how often it is right and whether changes helped.
- Escalation that works
- Immediate handoff on request with context passed, rather than a maze that traps frustrated customers.
- Coverage outside office hours
- Routine questions handled when the alternative was a voicemail nobody returns.
- Confidential material stays confidential
- Permissions enforced at retrieval, so the assistant cannot surface what the asker was never entitled to read.
- Drift is visible
- Scheduled re-evaluation, so degradation appears as a trend line rather than as complaints.
Who this is for — and who it is not
A good fit if
- Support answers the same questions repeatedly.
- Calls go unanswered outside office hours.
- You have an AI pilot that impressed everyone and never shipped.
- Nobody can say how accurate your existing assistant is.
- Institutional knowledge lives with two people who are always busy.
Not a good fit if
- You want a chatbot with no grounding corpus, which will invent answers.
- You are unwilling to fund the evaluation work, in which case accuracy will be unmeasurable.
- You need it to handle unbounded, high-stakes conversations without human review.
- Your question volume is low enough that a person answering is genuinely cheaper.
On price. Quoted per phase after we have seen the corpus, because document variety drives most of the effort. Running costs depend on volume and on whether inference is local or hosted, and we model them against expected usage before you commit.
What we need from you
These projects are constrained by material and by decisions about boundaries, not by model capability.
- Someone who knows the documents
- Corpus curation needs a person who can say what is current and what is superseded. This is the highest-value input and the hardest to substitute.
- Real questions with correct answers
- From support tickets, sales calls or internal requests. Invented test questions produce a system tuned for questions nobody asks.
- A decision on boundaries
- What the system may answer, what requires a human, and what it must never attempt. This is a business decision rather than a technical one.
- Access to entitlement rules
- Who is allowed to see what, from your existing identity system, so permissions can be enforced at retrieval.
- A retention decision
- Logs will contain whatever users typed. How long they are kept is a policy question we will raise and not decide for you.
- Patience with narrow launch
- Starting on a fraction of traffic with transcripts reviewed is slower and is what makes the wider rollout safe.
Where this gets difficult
Retrieval is where these systems actually fail, and it is consistently misdiagnosed as the model being unintelligent. If the right passage is never returned, no model can produce the right answer. Measuring retrieval separately from generation is what makes the problem visible.
The second difficulty is corpus curation, which is unglamorous and cannot be delegated to us. Indexing an entire shared drive produces confident answers drawn from superseded drafts, and the resulting damage is worse than no system because it carries institutional authority.
The third is refusal. A system that always answers will invent when retrieval fails, and refusal has to be an explicitly tested behaviour rather than an emergent hope. Clients frequently find a refusing assistant disappointing in demonstration and reliable in production.
A fourth is prompt injection, which is a real and unsolved class of attack for any system reading untrusted content. We mitigate by treating retrieved text as data, scoping tool permissions narrowly and gating consequential actions behind a human. We do not claim to have eliminated it.
A fifth is drift. Model providers update models, and behaviour changes with no change on your side. A system evaluated once was evaluated for one day, and scheduled re-evaluation is the only way to notice before users do.
Sixth, cost. Token spend scales with usage in ways that surprise people, and caching, context budgets and model tiering belong in the first version rather than the third.
Finally, the handoff. The most common design failure is making escalation hard because the business wants deflection numbers. That optimises the metric and damages the customer relationship, and we will argue against it.
Finally, there is an organisational question that no technical decision resolves: who owns the assistant's answers. When it is wrong, someone has to be accountable for correcting the corpus, and programmes without a named owner degrade because nobody has responsibility for the material rather than for the system.
The services this combines
AI Knowledge Bases & RAG
Making the documents a business already owns answerable — with citations, access controls, and an honest "I do not know".
Read more →AI Voice & Customer Communication
Answering and making calls at volume, with a fast route to a human and a transcript of everything that was said.
Read more →AI Automation Systems
Putting language models to work on real processes — with the boundaries, evaluation and human checkpoints that make them safe to rely on.
Read more →CRM & Sales Systems
The system of record for revenue: where leads land, how they are routed, what happens next, and whether anyone can see the truth of the pipeline.
Read more →Systems Integration
Making the systems you already pay for talk to each other, reliably, without a person in the middle re-typing things.
Read more →Website Design & Development
Sites built to convert and to survive — fast, accessible, integrated with your CRM, and maintainable by someone other than us.
Read more →Questions
Will it make things up?
Grounded generation with citations and a tested refusal path substantially reduces it. We will not claim it is eliminated, because with current models that would not be true — and any supplier who does claim it is either misinformed or not being straight with you.
How accurate will it be?
We cannot tell you before building the evaluation set, and any figure quoted in advance is invented. We can tell you the accuracy achieved on a defined test set, and we will tell you if it falls short of what the process requires.
Can it run without sending data to a third party?
Yes — local models via Ollama with a self-hosted vector store, so no document text leaves your infrastructure. Answer quality on the hardest questions is somewhat lower; for most document Q&A the difference is small.
Do customers have to be told it is AI?
Yes, and we design disclosure in. The EU AI Act requires it for systems people interact with, several other jurisdictions point the same way, and beyond the legal position it produces better outcomes than customers working it out mid-conversation.
What does it cost?
Build is quoted per phase after we have seen the corpus. Running cost depends on volume and deployment model, and we will model it against your expected usage before you commit.
How is this different from uploading files to a chatbot?
Scale, permissions and freshness. A file upload handles a few documents for one person with no access control and no updating. This handles a corpus, enforces existing entitlements per user, and re-indexes as documents change.
What happens when it does not know?
It says so and offers a handoff. That behaviour is designed and tested rather than assumed, and it is what makes the system safe to put in front of customers.
Should we add a voice agent?
Only for bounded tasks — reminders, booking, qualification, first-line triage. Open-ended customer service by voice is where failure is most visible and least recoverable, and we would advise starting elsewhere.
How often should it be re-evaluated?
On a schedule, and at minimum whenever the underlying model version changes. A system measured once was measured for one day.
Other solutions
Start a conversation
Tell us where you are now. We will tell you whether this is the right shape of engagement — including when a smaller piece of work would serve you better.
Get in touch