How to deploy a useful AI assistant in two weeks
Most chatbot projects fail because the team scopes for six months and burns out at month two. Here is the two-week playbook we use at Stellar AI, end to end.
- playbook
- automation
Most chatbot projects we have inherited had the same shape. Six months of scoping. A 40-page requirements document. A demo that looked great. And then nothing in production, because the team had run out of energy before they ever shipped.
We do it differently. Every customer assistant we have launched in the past two years went live in two to four weeks. Here is the cadence.
Week one: kill the scope
Day one is a single working session, usually two hours. We do three things:
- Pick a single workflow. Not “all of customer support”. One specific intent, like “answer questions about shipping status” or “qualify website visitors for the sales team”. Smaller is better.
- Find the knowledge source. What does the bot need to know? Usually it is one Google Doc, one Notion page, one help center, or a database export. If we cannot point at it on day one, the project is not ready.
- Decide where it lives. Website widget? WhatsApp? In Slack for the internal team? Pick one channel for v1.
Days two through five are setup. Access to the knowledge source, a fresh deployment environment, a basic prompt structure. By Friday the bot can answer one question in one channel.
Week two: make it real
Days six through eight: edge cases. We sit with the customer’s team and try to break the bot. Every weird question, every typo, every “what if I ask it about pricing instead”. The output of this is a list of refusals, redirects, and escalations.
Days nine and ten: integration. CRM, helpdesk, calendar, whatever the bot needs to actually act in the world. This is where most of the real engineering happens.
Day eleven: soft launch. We point a small share of traffic at the bot — sometimes a single agent’s queue, sometimes a single landing page — and watch what happens.
Days twelve to fourteen: tune and hand over. Whatever the soft launch surfaced gets fixed. The customer’s team gets the admin walkthrough. We agree on what “production” means and we flip it on.
What we cut
Plenty.
- A polished UI. v1 looks like a chat. Custom branding waits.
- More than three integrations. v1 connects to three systems max. The rest are roadmap items.
- Edge-case completionism. If a question shows up in fewer than 1% of conversations, the bot says “let me get a human” and we move on.
- Analytics dashboards. We use the logs. Dashboards come after we know what to measure.
Why two weeks
Because attention is the rarest resource on these projects. The customer’s team can focus on one thing for two weeks. They cannot focus on it for six months. If the bot is not measurably useful by the end of week two, something is wrong with the scope, not with AI.
The funny thing about two-week deployments is that the bot is rarely “done” at the end of them. It is in production, doing real work, and we keep improving it for months afterward. But it is shipping, it is paying back, and the team is not exhausted.
When two weeks is wrong
Sometimes it is. If the knowledge source has to be built from scratch (no docs, no FAQs, just tribal knowledge), the first two weeks are documentation, not bot-building. If the integration is to a 1990s ERP with no API, that is its own project before AI gets a look in.
But for most teams asking “should we have an AI assistant”, the right answer is “yes, and you should ship the first version of it before the end of the month”. Anything longer and you will not ship at all.