At SaaStr AI Annual 2026, Amelia Lerutte, SaaStr’s Chief AI Officer, took five months of running 10K, SaaStr’s AI VP of Marketing, distilled what worked into a spec, and built a brand-new one from scratch on stage in about 15 minutes. The whole room built theirs at the same time.
This is the playbook to do it yourself. First the mental model, then the build in ten steps, then each step in full with the data to feed it, the workflows to build first, and the guardrails that keep it from emailing your entire database at 2am.
10K did not start as any of this. Back in January 2026, Amelia was tired of one chore: every Sunday night, copy-pasting marketing, sales, and go-to-market dashboards into Notion so the team could review them Monday morning. So she vibe coded a dashboard to stop the copy-paste. That was the entire original ambition. Five months later it owns the number, builds campaigns, writes email copy from real data, and reminds her about the things she forgets. SaaStr now runs close to 30 agents that have been used almost a million times. None of them started complicated. Neither should yours.
Before You Start: Three Things That Make It Work
#1. One agent, one goal, one brain.
Give each agent a single number to own. That is why SaaStr runs separate agents: 10K owns marketing, QBee owns customer success, another agent owns SaaStr Annual. Do not load one agent with all three. The focus is what makes the outputs good.
#2. The agent itself is the entity.
SaaStr did not build a super-agent or a custom connector layer on top of everything. Each agent is its own thing with its own brain, and it develops its own personality the more you talk to it in the editor. Whoever manages the agent, you, a head of AI, or a forward-deployed engineer you hire, is the person who talks to it the way Amelia talks to 10K every day.
#3. Two layers.
The autonomous layer is the dashboards, the scheduled jobs, and the AI-drafted emails that run in production around the clock. The operator layer is the agent in the editor doing one-off analysis and outbound. The first layer is what your team and customers see. The second layer is the moat, and we will get to why.
The Build, At a Glance
- Pick one number, then write the spec. The more detailed the spec, the better the agent.
- Dump in every spreadsheet you already have. Real history is your ground truth, and most of it is not in any API.
- Build v1 in a vibe coding platform. Spec plus data into Replit, roughly 15 minutes to a working version.
- Connect Salesforce first. Read pipeline and revenue, then write back.
- Hook up your other APIs, one at a time. CRM, marketing automation, Slack, Google Calendar. Stairstep.
- Build workflows one at a time. Dashboard, then daily ideas, then campaigns, then a newsletter builder.
- Set what runs on its own, and what asks first. Pulling data is autonomous. Emailing the database is not.
- Build the hallucination guard before the first send. Swap in real numbers, block any send that does not match.
- Keep a memory file the agent reads every session. Voice rules, contacts, and every correction in one file.
- Verify against real data, then deploy. Check the first several outputs by hand before you trust it.
Step 1: Pick One Number, Then Write the Spec
Everything starts with one number at the top. For 10K it was paid attendees and net event revenue against a hard date. Pick yours before you write a line of code. An event is tickets plus sponsor revenue against target. A launch is signups plus activations plus paid conversions. Revenue ops is new ARR plus expansion minus churn against the quarter. Write it on a sticky note and do not start without it. Every integration, chart, and prompt downstream serves that one number.
Then write the spec. Amelia’s rule is simple: the more detailed you are, the better the inputs the agent asks you for and the better the outputs you get back. A generic AI VP of Marketing spec gives you a generic agent. SaaStr published the exact 20-page spec it used, plus sample historical data, so attendees could build their own version on the spot. If you are not sure what belongs in your spec, ask Claude. Tell it the goal and have it help you draft the spec before you hand it to your build tool.
One thing that matters as much as the spec: how you treat the agent. Amelia treated 10K as a dashboard on day one and told it exactly that. Now she treats it as a co-pilot and a coworker. That shift is part of the build.
You can grab our spec here.
Step 2: Dump In Every Spreadsheet You Already Have
Before you wire a single integration, collect every spreadsheet, CSV, and report you currently use to run this part of the business and drop them into one folder. The agent should treat these as ground truth and build the first dashboard around them.
This matters more than it sounds. Most of your useful history is not in any API. Sponsor pricing from five years ago, who attended your VIP dinner, the workbook your CFO maintains, none of that lives in your CRM or your ticketing tool. If you do not load it on day zero, the agent will only ever know what the live integrations expose. It also lets you show real charts on day two instead of day fourteen, and it anchors every AI output in real numbers instead of guesses.
Do not be embarrassed by the mess. SaaStr started with data scattered across CSVs, Salesforce, and Marketo in a hundred different places. Amelia just gave the agent the CSVs. Drop the raw exports in unchanged, let the agent write the parsing, and re-upload whenever the source updates. If security is a concern, hook up natively through your CRM and marketing automation APIs instead.
Feed it in this rough order: marketing-sourced or marketing-touched revenue first, then campaign data (what worked, what did not, every ad and agency spend), then email data (opens, clicks, who reads and who does not), then a progress tracker of what you have done so far in 2026. The agent will tell you, sometimes bluntly, that the campaign you thought was gangbusters did not actually work.
Step 3: Build v1 in a Vibe Coding Platform
Drop the spec and the data into Replit, or whatever build tool you use, and let it generate the agent. On stage, the first working version took about 15 minutes. Amelia’s only manual touch on her rebuild was telling it to make the dashboard purple and rename it.
If all you leave with is a dashboard that pulls from your CRM, that is a real win. The dashboard was 10K’s entire origin. Everything else stairsteps on top of it.
Step 4: Connect Salesforce First
The first real integration Amelia built was a Salesforce connected app so 10K could read, and eventually write, pipeline and revenue.
She is not Trailblazer certified. She owns an Agentblazer hoodie and is not sure she is even Agentblazer certified. She asked Claude how to build the connected app, and it walked her through it. For most teams this is the same first step. Reading closed-won revenue and pipeline through the Salesforce API also gives you historical comparisons and projections, which is most of what you want the agent doing on Mondays.
Step 5: Hook Up Your Other APIs, One at a Time
After Salesforce, add the rest as you go: your marketing automation platform, social if you want the agent to post (SaaStr still writes its social by hand), Slack for daily summaries, and Google Calendar.
That last one is one of the best examples in the session. Sending hundreds of speakers their personalized calendar invites used to be a full person’s job and took a week. Each invite had the speaker’s session time, the correct venue address rather than the obvious one, green room logistics, and the right press, marketing, and executive teams cc’d. 10K did all of it through the Google Calendar integration in about 20 minutes.
The task was menial, not glamorous, and that is exactly why you should hand it to the agent. Your time is better spent building the agent and running campaigns only you can run. The agent is great at the menial work, and it does it perfectly and fast.
One rule for every integration: cache the data to your own database with a short refresh window. Never let a dashboard page hit a third-party API on every load, or it will be slow, expensive, and rate-limited inside a week.
Step 6: Build Workflows One at a Time
This is where the agent turns from a dashboard into a co-pilot. Build these one at a time. Do not try to hook up everything at once, and watch for the doom loop, the spiral of “what about this, and what about that,” where you plan ten workflows and ship none. Write the future states down, then build them in order.
Workflows SaaStr AI runs live:
- Daily ideas. Every weekday morning, the agent emails three to five specific moves for the day, each tied to a real number and doable in under two hours. Grounded in your data and your one goal, the ideas get sharper as you give feedback. The first batch was fine. After Amelia told it which ideas were too expensive or just bad, the next batches got good.
- Win-back campaigns. Pull everyone who came to SaaStr Annual last year but has not bought a ticket this year. The agent runs the list, finds who lapsed, enriches contact data through a tool like Clay, and drafts the campaign.
- Light competitive campaigns. When Replit sponsored SaaStr, 10K pulled the list of similar companies attending and drafted outreach to a competitor making the case they should be there too. Finding competitors, finding contacts, and pulling the right data to highlight all ran without a human stitching the steps together.
- Website action emails. When attendees played the games on the SaaStr site to unlock a ticket discount code, 10K collected their emails and sent the reminders automatically. Every time it sent a daily reminder, SaaStr saw ticket spikes.
- Newsletters. The attendee newsletter was built by the agent. Instead of guessing what to include while tired, Amelia asked the agent what mattered, which stripped out her own bias. SaaStr eventually vibe coded a full newsletter builder into 10K so the templates and the countdown timer stopped breaking.
- Ads. Strong ads need fresh creative constantly, more than any human keeps up with. The agent generates endless variations of copy and images, proposes a plan, and you test what works. Amelia regularly hands 10K several images and asks which should be the ad, then runs the test. Every idea is grounded in your data and your one goal, which is all it thinks about.
Step 7: Set What Runs on Its Own, and What Asks First
Be explicit about what the agent can do autonomously and what it must check on first. Pulling data, building dashboards, and proposing campaign ideas all run on their own. Sending email runs semi-autonomously: the agent sends Amelia a test, asks if she likes it, and only sends after she says yes.
You do not want your AI VP of Marketing to instantly email your entire marketing database. That should scare you a little, and the fix is to draw the line clearly in the spec.
Step 8: Build the Hallucination Guard Before the First Send
10K’s own top takeaway, when Amelia asked it to summarize the session: guardrails beat prompt engineering. Her story makes the case. She asked 10K for the VCs who came to SaaStr Annual last year and have not returned this year. It said there were about 400 and offered to draft outreach. She said great, give me the names. It paused and said: oh, hold on, I made that up. So she told it to go pull the real data, and then it did.
The manual rule is simple: talk to your data, verify the output, then send. Agents are fast, which makes it tempting to glance at a result and hit send.
The engineering fix is what actually protects you, and the spec is specific about it. List the real numbers in the prompt every time. Then, on the server, before anything sends, replace every number in the agent’s output with the ground-truth value from your database. If a number in the draft does not match a real value within a small tolerance, the system flags it and refuses to send. Build that guard before the first AI email goes out, not after. One wrong number to your team or your list erodes trust faster than you can rebuild it.
Step 9: Keep a Memory File the Agent Reads Every Session
The institutional memory lives in a single file. SaaStr keeps a project file at the root with the one goal, the voice rules, the team contacts, the send-domain rules, and every correction the operator has ever made. The agent reads it at the start of every session.
The voice rules are the part you will lean on constantly. Real examples that have mattered: never say SaaS, always say B2B. Always net revenue, never gross. Send from the one verified domain, never a personal address. Encode each correction the moment you make it, and the file works like onboarding documentation that never goes stale.
Step 10: Verify Against Real Data, Then Deploy
Check the first several outputs by hand before you trust any workflow. Amelia was nervous the first time 10K sent an email directly: what domain would it use, what reply-to, would it dump too many images and land in spam. She tested it heavily before letting it run. Do the same, then let it go.
Why This Compounds: The Operator Layer Is the Moat
The autonomous layer is what everyone sees. The operator layer, the agent in the editor, is the advantage. Every time you ask it a one-off question, pull the top 200 VIPs to invite, find a specific deal in Salesforce, rerun last year’s deep dive against this year’s list, the agent writes a small reusable script and leaves it behind. Ask once, and you have the answer for good. The script library grows, every correction gets encoded permanently, and the system gets sharper week over week instead of starting cold every time.
That is why you keep one editor session open for weeks rather than closing it. The accumulated working memory is the point. The spec’s own framing: after three months, the system knows more about how you run marketing than a new hire would after a year.
Is 10K Actually a VP of Marketing?
By its own assessment, not entirely. 10K does not think it has replaced a VP of Marketing. It says it has replaced roughly 60% of the basic functionality. It does not own people. On nearly everything else, it holds its own in many ways. If not as a VP, at least as a senior team member and IC.
You are deploying an agent, giving it one goal, feeding it real data, and building. The ideas get better, the campaigns get sharper, and the menial work that used to eat a person’s week disappears in 20-minute increments.
The agent does not need to be SaaStr’s full 10K on day one. Ours took five months of iterations (although we could get there much faster today). Yours can start with a single dashboard this afternoon.
You can grab the exact 20-page spec SaaStr used to build 10K live, plus sample historical data, at saastrannual.com/resources (direct link to the spec).
