Ship It Live — $1,999 flat
Design + build + deploy · kickoff in 24h · limited slots
Ship It Live · AI assistant sprint
An AI assistant that knows your business, live in 21 days.
Answers drawn from your actual content, with sources shown and a human one click away. Not a wrapper around a generic model pretending to know you.
Quick answer
A custom AI chatbot trained on your own documents and site content can be built and live in 21 days from a fixed $1,999. That covers content ingestion, retrieval with citations, a branded chat widget, human handoff and analytics. Model fine-tuning and regulated advice use cases are excluded.
The plan
Day one to day 21, written down.
You get this schedule before you commit, not after. If a phase slips, that is mine to absorb — the price and the date were agreed before anything started.
- 1
Days 1–3
Scope and content
What it must answer, what it must refuse, and which content is trustworthy enough to answer from. Most chatbot failures start as content problems.
- 2
Days 4–9
Design and ingest
The chat experience designed properly — suggested starting questions rather than a blank box — while your content is ingested and indexed.
- 3
Days 10–16
Build and tune
Retrieval, citations, refusal behaviour and handoff. Tuned against a real question set from your actual support history rather than invented examples.
- 4
Days 17–19
Adversarial testing
Deliberately trying to make it wrong — off-topic questions, prompt injection, edge cases. What it does when confused matters more than what it does when it knows.
- 5
Days 20–21
Launch
Widget live on your site, analytics on, and a short handover on updating content and reading what people are asking.
What from $1,999 gets you
Everything needed to launch — nothing padding the invoice.
Who books this
Built for
Support teams answering the same forty questions every week
Companies with documentation nobody can find and everybody needs
Founders who want an AI feature that is genuinely useful, not a demo
Real situations
Five ways this shows up, and what actually fixes it.
Not hypothetical — the shape of the problem before someone books this sprint, and the specific part of the build that resolves it.
Situation 1
Answering the same forty questions every single week
A support team spends a meaningful share of every week typing near-identical answers to the same recurring questions, pulled from documentation that already has the answer written down.
A chatbot trained on that real documentation answers those forty questions directly, with citations, freeing the team for what actually needs a human.
Situation 2
Documentation nobody can find, that everybody needs
A company has thorough internal or customer-facing documentation, but finding the right page still means asking a colleague who happens to remember where it is.
Retrieval-based answers grounded in that real content — with citations — surface the right document in seconds instead of a search-and-ask cycle.
Situation 3
Pasted the docs into ChatGPT and it still makes things up
A team tried the fast version — pasting documentation into a general model — and found it confidently answers questions the documentation never actually addressed.
Answers are retrieved only from ingested content and cited, with an honest 'I don't know' fallback instead of a confident invention.
Situation 4
An AI feature that's technically live but nobody trusts
A previous chatbot attempt shipped but gets bypassed by users who've learned it gives unreliable answers, making the feature dead weight rather than useful.
Adversarial testing in days 17-19 deliberately tries to break it — off-topic questions, edge cases — before real users find the same gaps.
Situation 5
No way to escalate when the bot is out of its depth
A support chatbot with no human handoff leaves frustrated users stuck in a loop when their question falls outside what it can actually answer.
Handoff to email, a form, or an existing support tool is built in — both for the user asking and for the bot recognising its own limits.
The honest part
When 21 days is the wrong answer.
A fixed deadline only works when the scope genuinely fits inside it. These are the cases where I will tell you so rather than take the booking.
Regulated advice — medical, legal or financial guidance to consumers. The liability is real and no amount of disclaimer text fixes it.
Fine-tuning a model on your data. It is almost never the right answer for this use case, and I will explain why rather than sell it.
Voice assistants or phone systems, which are a different stack with different failure modes.
Inside the 21 days
- Ingestion of your real content — site, docs, PDFs, help centre
- Retrieval-based answers grounded in that content, with citations shown
- A chat widget matched to your brand, embeddable in one line
- Honest fallbacks: it says it does not know rather than inventing
- Human handoff to email, a form, or your existing support tool
- An analytics view of what people ask and what it failed to answer
- Content refresh so answers stay current as your docs change
- Source code, prompts and configuration yours at handover
Quoted as one number before day one. A week running long is mine to absorb.
Outside the line
- Regulated advice — medical, legal or financial guidance to consumers. The liability is real and no amount of disclaimer text fixes it.
- Fine-tuning a model on your data. It is almost never the right answer for this use case, and I will explain why rather than sell it.
- Voice assistants or phone systems, which are a different stack with different failure modes.
Real work, sized and priced separately. You hear it before you book, not at handover.
Questions
Before you book
It answers only from your content, shows which document each answer came from, refuses when it does not know, and hands off to a human when it should. A general model with your docs pasted in will confidently answer things you never wrote, which is the exact failure that makes support bots untrusted.
Grounded answers versus generic chat
A 21-day chatbot sprint ships retrieval over your content, citations, refuse, and human handoff. It is from $1,999, not a promise of perfect answers.
| In scope | Out of scope |
|---|---|
| Ingest a corpus you control | Pasted-docs chatbot cosplay |
| Retrieve and cite source content | Whatever the model remembers |
| Refuse when the corpus is silent | Confident invented paragraphs |
| Handoff to a human path | A widget with no escape hatch |
Design and handoff bar
- Empty, low-confidence, refuse, and handoff states exist
- Sources are visible when answers rely on them
- Off-topic prompts get a boundary and redirect
- Corpus updates have a named owner
- Adversarial QA happens before real traffic
Soft planning rule (not an SLA): lock the primary job early, review edge cases before launch week, and keep accounts and source in your name at handoff.
Extra questions
How do we keep content fresh?
Can it give regulated advice?
Is day 21 a guarantee it never hallucinates?
Other sprints
21 days from now, this could be live.
A 30-minute call decides whether the scope fits. If it doesn't, I'll tell you what would.