Most "AI agent" projects die in scope, not code. The model works fine. The two-week version stalls because nobody decided what the agent is allowed to answer and where it gets the answers. So let me name the work.
We build Slack agents for a living. A support agent we shipped now answers in 12 minutes, down from 4 hours. That number came from cutting scope, not from a smarter model.
Week one: the boring infrastructure
The first week is plumbing, and it is where most of the risk lives. You register a Slack app, set the right OAuth scopes, and decide between Socket Mode and an HTTP endpoint. Socket Mode is faster to stand up. An HTTP events endpoint scales better and is easier to monitor. Pick based on traffic, not vibes.
Then you wire the event handler. Slack expects a response within 3 seconds. Your agent will take longer than that. So you acknowledge the event immediately, do the real work async, and post the answer back to the thread. Skip this and Slack retries the same message, and your agent answers twice. That bug is the first one everyone hits.
Next is the knowledge boundary. An agent with no boundary will confidently answer anything, including things it should not. We point it at a defined corpus - docs, a wiki, a ticket history - and we make "I don't know, here is who to ask" a first-class response. A RAG setup over that corpus is most of the build. We price that pattern as a RAG Chatbot for a reason; it is its own discipline.
Week two: making it trustworthy
The second week is about trust, because a wrong answer in Slack spreads fast. You add the retrieval layer, tune what counts as a confident match, and set a threshold below which the agent escalates to a human instead of guessing.
You also decide what the agent can do versus only say. Reading and answering is low-risk. Taking actions - creating tickets, posting to channels, hitting internal APIs - needs guardrails. We gate write actions behind explicit confirmation early on. You can loosen that once the team trusts it. You cannot un-send a bad automated message.
Then you log everything. Every question, every retrieved chunk, every answer. Week two is when you watch real traffic and find the questions the corpus does not cover. Those gaps are the actual product work, and you only see them with logging in place.
Why two weeks is the right box
The deadline forces decisions. Without it, teams keep adding "just one more" data source and the agent never ships. A two-week box makes you answer the only questions that matter. What does this agent do? Where does it get answers? What does it refuse?
We can hold that timeline because we have done it. Ten production products in seven months, built by a two-person team, one of us an ex-AWS Federal Solutions Architect. We have shipped enough agents to know the failure modes before they happen.
The most common failure is treating the agent as a search box with a personality. It is not. A useful Slack agent is a narrow tool with a clear job and a clear "I don't know." Narrow ships in two weeks. Broad ships never.
What we hand you
You get the deployed agent, the code, and the infrastructure. You own all of it. Fixed price, agreed up front, and we exit when it works. No retainer you cannot leave.
If you already run on agent tooling, our OpenClaw Installation sets up the agent runtime in your environment. If you are starting from a knowledge base and want answers in Slack, the RAG Chatbot build is the right shape. Either way the scope is fixed before we start, so two weeks means two weeks.
One honest caveat. If your "knowledge" is scattered across five tools with no clean source, week one gets longer because we have to find the corpus first. That is a data problem, not an AI problem, and we will tell you that on the first call instead of after the invoice.
A Slack agent that answers in minutes is not exotic. It is a defined corpus, an async handler, a confidence threshold, and the discipline to refuse. Two weeks of focused work, not a moonshot.
Want to scope yours? Book a free 30-minute call and we will tell you if two weeks is realistic for your setup.