Enterprise chatbots that actually know your business

A language model asked a question with no supplied context will produce the most plausible-sounding answer it can. It has no mechanism for distinguishing "I recall this" from "this pattern fits" — which is why an ungrounded assistant will confidently invent a refund window, a compatibility matrix or a policy clause that has never existed.

Grounding changes the task. The system retrieves the relevant passages from your content first, hands them to the model, and instructs it to answer only from that material and to state clearly when the material does not cover the question. The model stops being a source of facts and becomes a writer working from documents in front of it. This is not a prompt trick; it is an architecture, and its accuracy depends almost entirely on whether retrieval found the right passages.

The remaining failure mode is subtle and worth naming: retrieval returns something related but wrong — the policy for a different region, the previous version of the pricing sheet — and the model answers faithfully from the wrong source. We defend against it with metadata filtering (region, product, effective date, audience), reranking that scores passages for actual relevance rather than embedding proximity, and citation display so a human can see instantly which document an answer came from.

Across our rescue engagements, moving answer accuracy from around 60% to above 90% has never required a better model. It has always come from five changes to retrieval.

Structure-aware chunking. Splitting documents every 500 tokens severs tables from their headers and clauses from their conditions. We chunk on document structure — sections, list groups, whole table rows with headers preserved — so a retrieved passage is self-contained and means the same thing out of context as in it.

Hybrid search. Pure vector search is weak on exact identifiers: SKUs, error codes, policy numbers, version strings. Combining keyword search with semantic search catches both the paraphrased question and the one that pastes an error code.

Reranking. First-stage retrieval optimises for recall; a cross-encoder reranker then scores the top candidates for genuine relevance to the question. This single step typically produces the largest accuracy jump of anything on this list.

Metadata and freshness. Every chunk carries its source, effective date, region, product and audience. Queries filter on them, superseded content is excluded rather than merely outranked, and reindexing runs on a schedule tied to how your content actually changes.

Query understanding. Real questions are compressed, misspelled, and often follow-ups that depend on the previous turn. We rewrite the query with conversation context before retrieval — an unglamorous step that fixes a large share of "it didn’t understand me" complaints.

The moment that decides whether customers trust your assistant is the moment it cannot help. Done badly — "I’m sorry, I didn’t understand that" repeated three times before a hidden link to a contact form — it converts a mild question into a complaint. Done well, it is faster than not having used the bot at all.

Our escalation carries the full conversation, the customer record, what the assistant tried, and why it stopped, straight into your helpdesk as a ticket or into live chat as a routed conversation. The agent picking it up starts from a summary, not from "hi, how can I help?". Handling time on escalated contacts typically falls even as volume shifts to them.

We also make escalation explicit and easy at every turn. A visible "talk to a person" affordance reduces frustration and, counter-intuitively, reduces its own usage: customers who know the exit exists are more willing to let the assistant try. And every escalation is logged as a signal — clustered weekly, it is the most honest backlog of what your assistant and your documentation still need.

Containment rate alone is a misleading metric: an assistant that stonewalls customers contains conversations beautifully. We track four numbers together and treat them as a set.

Resolution rate — the share of conversations where the customer’s problem was actually solved, measured by follow-up contact within seven days rather than by whether the session ended. Deflection that produces a phone call two hours later is not deflection.

Answer correctness against the evaluation set — several hundred real questions with known good answers, scored automatically on every change. This catches regressions from content edits, model updates and prompt changes before customers do.

Correct-refusal rate — how often the assistant declines when the answer genuinely is not in the knowledge base. A falling refusal rate alongside a rising resolution rate usually means it has started guessing, and it is the earliest warning signal available.

CSAT on bot-resolved conversations, tracked separately from human-resolved. Our deployments generally land in the 4.4–4.7 range once tuned; anything materially below that indicates a tone or escalation problem rather than a knowledge one.

One assistant should serve every surface, with per-channel behaviour rather than per-channel rebuilds. The knowledge base, policy and evaluation set are shared; what changes is formatting (WhatsApp gets short messages, the web widget can render tables), authentication, and which tools are available. Rebuilding a separate bot per channel is how organisations end up with three assistants giving three answers.

Multilingual support is largely free at the model level and expensive at the content level: the assistant can answer in twenty languages, but your knowledge base exists in one or two. We retrieve in the source language and answer in the customer’s, which works well for factual content and needs review for anything legally binding. Where regulation requires approved translations, we route those answers to reviewed content only.

Identity matters for both trust and regulation. Our assistants identify themselves as AI on first contact, and where a signed-in customer is talking to the assistant, answers are scoped to their own entitlements — an authenticated assistant that can read the customer’s orders is dramatically more useful and needs correspondingly careful access control.