“How fast can we be live?” came up on four of my intro calls in August, usually as the second question, right after “is our data safe.” The answer we give every prospect: two to three weeks from kickoff to a live agent doing real work in your team's Slack. Not a pilot. A working agent with a defined role, producing deliverables you would otherwise have to hire for or absorb yourself.
That number earns skepticism, and I understand why. Most owners I talk to are picturing a software rollout: a discovery phase, an implementation partner, a steering committee, and a login nobody uses by spring. That picture comes from ERP and CRM projects. An agent build is closer to onboarding a new hire, and a good hire is productive in weeks.
This guide walks through the actual schedule we run, week by week: what we need from you, what happens on our side, what can stretch the timeline, and what the month after launch looks like.
Before the clock starts: the planning session
The engagement opens with a 30-minute planning session. You bring three things.
First, the systems your team already works in. Your CMS, CRM, analytics, inbox, project tracker. Whatever runs the business today.
Second, the process knowledge for the work you want off your plate. A written SOP is ideal. A recorded walkthrough, or twenty minutes of you explaining how the work gets done, is enough.
Third, a goal we can measure. “Publish two blog posts a week” or “every invoice matched within a day” beats “help us with AI.”
We map the highest-impact agent roles for your business against those three inputs and tell you what shipping them takes.
Two things you do not need: new software and a technical hire. Agents show up in Slack, Microsoft Teams, or WhatsApp. If your team runs on email and SharePoint, we work with that. I talk to a lot of teams that got burned by a past tool rollout and flinch at the word “platform,” so it usually takes a second pass to land that there is nothing for your staff to install, license, or learn.
And sometimes the planning session ends with us saying you are not ready yet. If the work you want handled has no process behind it and no measurable goal, an agent will automate the confusion. We would rather you write the SOP first and come back than launch something that erodes your trust by week two.
Week one: design the role
Week one is paperwork, in the best sense.
Every agent is the sum of five parts: a job description, a personality, a set of tools it may touch, a library of skills, and the AI model it runs on. Week one is where those get decided, in writing, with you in the room.
The job description does the heavy lifting. It sets what the agent owns, what good output looks like, and which decisions it can make alone versus which need a human signoff. Roles are narrow on purpose. Narrow scope is what makes the work reliable. A content agent drafts your weekly posts and queues them for review; it does not also negotiate with vendors and reconcile your books. When a prospect asks for one agent that does everything, we split the request into two or three narrow roles instead, because that is the version still working in month six.
Tools get scoped the same week, and they get scoped read-only first. An agent can read your analytics long before it can edit a page on your site. Write access is earned after you have watched the work for a while, the same way you would not hand a first-week hire the keys to your billing system. Data safety is usually the next question, and the answer is short. Your data is not used to train models, and the default configuration stores nothing.
Voice work rounds out the week. Any agent that writes anything a customer might read gets tuned on your real material (emails, posts, past proposals) so the drafts sound like your business instead of a template.
By Friday of week one you have signed off on each role's job description, its tool scope, and its approval boundaries. That signoff costs you about an hour, and it is the hour that decides how month three goes.
Week two: build, connect, rehearse
Week two is the build. We configure each agent's skills, meaning the repeatable how-tos it applies: keyword research, on-page SEO, weekly reporting, call summaries, invoice matching. We connect the tools scoped in week one and wire the approval gates so drafts route to the right person.
Then we rehearse. The agents run real tasks against your real systems while the output still comes to us. This is where the rough edges show up: a report pulls the wrong date range, a draft sounds close to your voice but misses it, a permission turns out to be missing. Better that we catch those than you.
Midway through the week, the agents appear in your Slack and your team starts talking to them directly. The interaction works the way it would with any teammate. Type “summarize last week's calls and draft three follow-ups” and the agent runs with it, with the follow-ups landing as drafts for your salesperson to review and send. Teams learn to delegate to agents the same way they learned to delegate to a new hire: cautiously at first, then routinely.
Week three: launch
Go-live is quieter than people expect. There is no cutover weekend and no migration. The work simply starts counting, and the first real deliverables land: a blog draft in your review queue, a weekly report in your inbox Monday morning, follow-up drafts sitting behind your approval before anything leaves the building.
In week three, every piece of external-facing work routes through a human. That is deliberate, and it stays that way until you decide otherwise. You are checking quality, and the agents are building the track record that earns looser gates later.
The month after: the governor stays on
Nobody asks about week four on a sales call, which is a shame, because week four decides more than weeks one through three do.
At the end of month one we run a full review with you: every deliverable shipped, every miss, every place a human had to step in. Only after that review do approval thresholds loosen, and only where the record supports it. A reporting agent that has nailed twelve straight reports can start sending them directly. The agent drafting your customer emails keeps a human approval in the loop for as long as you want one there.
Ongoing management is the unglamorous part and the reason the agents keep working. Models update, tools change their interfaces, your business shifts priorities, and somebody has to retune for all of it, weekly. A set-and-forget agent decays in weeks. That decay is the pattern behind most of the DIY attempts we get called in after.
The compounding shows up across the first 60 to 90 days: content depth, organic traffic, cleaner pipeline data, hours back on your team's calendar. Two to three weeks gets you a working agent. Two to three months is when it starts to feel like staff.
Get the next piece when it lands.
Short notes for owner-operators putting AI to work. No hype, no listicles.
What stretches a three-week build to five
Three things slow builds down, and none of them are technical.
Access is the big one. Waiting on a login, an API key, or an outside IT contractor to grant a permission adds more days than any other cause. If you can name the person who owns access to your systems before kickoff, you have protected the timeline.
Undocumented process is the second. When the answer to “how do you do this today” is three people disagreeing in a meeting, design week turns into refereeing. The fix is cheap: have the person who does the task today record themselves doing it once.
Approval design is the third, and it is the one worth slowing down for. Deciding who reviews what, and what ships without review, is a real management decision. Teams that rubber-stamp it in week one end up revisiting it in week four.
Common questions about the timeline
How long does it take to launch an AI agent for a small business?
Two to three weeks from kickoff to a live agent producing real deliverables, with the agents rehearsing inside your Slack by week two. Compounding results (organic traffic, content depth, pipeline contribution) typically show up across the first 60 to 90 days.
What do we need to have ready?
Access to the systems the agent will work in, some record of how the work is done today (an SOP or a recorded walkthrough), and a measurable goal. Plan on roughly two hours of your own time across the planning session and the week-one signoffs.
Do we need to buy new software?
No. Agents work inside Slack, Microsoft Teams, or WhatsApp and connect to the tools you already run. If your team does not use a chat tool, we set one up as part of the build.
What if the agent gets something wrong?
In the first month, every external-facing piece of work routes through a human before it ships. At the month-one review we walk through everything together, and approval thresholds only loosen where the track record supports it.