Chat and AI that does a job, not a demo.
Is your team still answering the same questions and retyping the same documents by hand? We build AI that takes that work off them, using your own documents and rules. We test it on your real questions before anyone sees it, and it hands over to a person when it should. Most take 4 to 14 weeks.
Every one of these knows when to stop talking and fetch a person. That’s the part most skip.
Answers from your documents, with the source cited. Hands over to a person when confidence is low or the customer asks. Logs every conversation.
On your website, portal or appA conversation on the website that ends in a booked call or a dealer appointment, not a form. Qualifies by your rules, writes to your CRM.
Writes to your CRM and calendarInvoices, contracts, claims, applications read, fields extracted, checked against your rules, routed. A human reviews the ones it’s unsure about.
With a review queue and accuracy targetsSearch and answers across the wiki, the CRM, the drive and the ticket system, respecting who is allowed to see what. Drafts that start from your data.
With Google or Microsoft sign-inSummaries, suggestions, classification and drafting as features of the application you already have, with the same login and the same audit trail.
Added to an app you already haveProduct Discovery lists the questions it must answer, the ones it must refuse, and what happens when it doesn’t know. The fee comes off a build that starts within 30 days of the workshop.
See Product DiscoveryWork like this.
Automotive, capability story
Live chat for a car brand’s tender, built against 400 test questions.
Insurance
Document intake that reads carrier forms without a person.
Education
AI tutoring trained on the UK curriculum for Key Stages 3 and 4.
Switch on the vendor’s AI when it knows your data. Build when it has to know your rules.
If your help desk’s own AI will do, we’ll tell you, and save you the build.
Choose it when your help desk, CRM or CMS ships an assistant, your content is all in that one system, and standard behavior is fine.
Watch for: It can’t see the other systems, can’t follow your handover rules, and you can’t test what it will say before a customer does.
Choose it when a single FAQ, a small audience, no integration, and you accept that nobody will know when it starts answering wrongly.
Watch for: Per-conversation pricing, a rules editor that can’t express your policy, and no test set.
Choose it when it speaks to customers or handles money, it needs two or more systems, Legal wants to know what it will say, and someone will be accountable for it.
What you get: A price agreed before we start, a test set Legal can read, a handover that works, and code in your repositories with the model as a setting.
Your documents answering questions in week two. Legal reading the test set by week six.
An example 12-week build: a chat assistant connected to your CRM, with a reviewed pilot.
- Week 0 Product Discovery
The questions it must answer, the ones it must refuse, the handover rules, the sources. A first test set drafted with your team. You get a price we’ll hold and a sprint plan.
- Weeks 1 to 2 · Sprint 1 Foundations
Accounts in your name: cloud, model provider, repositories. Your documents indexed with permissions, the rules it follows, the automated test run. Demo one: fifty real questions answered with citations, scored.
- Weeks 3 to 8 · Sprints 2 to 4 Behavior and integration
Handover to a person, guardrails for topic and tone, the CRM or ticket system connected, the widget or panel branded. The test set grows each sprint and runs on every change.
- Weeks 9 to 10 · Sprint 5 Pilot with review
Live for one segment of users or one team, with every conversation sampled in the review queue. Spend per conversation tracked. Legal and support read the transcripts.
- Weeks 11 to 12 · Sprint 6 Launch and handover
Pilot findings fixed, the remaining audience switched on, the review cadence agreed. Documentation, security review with IT, every credential handed over.
Five parts. The second is where your rules live.
By the end of sprint one, your documents are answering real questions. The AI model is a setting you can change, so you’re never tied to one provider.
Forwarding this to your IT lead? Send them the one-page security summary
Chat widget, in-app panel, Slack or Teams
Interface
What people talk to. A widget on your site or portal, a panel inside your application, or a bot in Slack and Teams. Branded, accessible, with the handover button always visible.
Prompts, tools, handover rules, logging
Orchestration
Where the behavior lives: what it may do, which systems it may call, when it must hand over, what it must never say. Every conversation logged with the sources it used.
pgvector lets Postgres search your documents by meaning. Indexed and permissioned.
Retrieval
Your PDFs, pages, tickets and records indexed so the model answers from them and cites them. Permissions carried through, so an assistant never surfaces what the user could not open themselves.
Your API account, swappable by configuration
Models
The model is called through your own account, so your data is under your agreement, not ours. Switching model or provider is a configuration change we test against the same set of questions.
Run on every change, before every release
Evaluation and review
A set of real questions with agreed answers, run automatically whenever anything changes. A review queue where a person reads a sample and the low-confidence cases. Spend per conversation on a dashboard.
What we’d turn down.
A studio that builds everything isn’t telling you the truth about something.
- A chatbot without a handover If there’s no person to hand over to, the assistant will eventually be alone with an angry customer. We won’t build that.
- Fine-tuning for its own sake Almost every request for a “custom model” is solved by having it search your documents, for much less. We’ll show you the test set before recommending anything else.
- Anything customer-facing without a test set If Legal or support can’t read what it will say before launch, it’s a demo. We ship products.
Questions buyers ask first.
Which model do you use?
Whichever scores best on your test set within a budget you set, called through your own account. Usually Claude or OpenAI, sometimes an open model for simple or sensitive tasks. It’s a setting; we re-test when a new one arrives.
Is our data used to train anything?
No. Calls go through your own provider account under its business terms, which exclude training. Documents and transcripts stay in your own database.
What about hallucinations?
Answers are grounded in your documents and cite them. When the search finds nothing relevant, the assistant says so and offers a person. The test set measures this on every change, and the review queue catches what the tests miss.
What’s the spend per conversation?
It depends on the model and the conversation, and you’ll see it on a dashboard from sprint one. We route simple questions to cheaper models and prove on the test set that quality holds.
Which languages?
The models handle most major languages well. We add test questions in each language you serve so the scores cover them.
Can it actually do things, like book or refund?
Yes, through tools with explicit permissions and limits agreed in Product Discovery. Anything that moves money or commits you is confirmed by the user or reviewed by a person.