Top 10 Takeaways from The Agents #006: The Numbers Behind Our Full Go-To-Market Agent Stack

On the latest episode of The Agents, we walked through the actual back ends of our top go-to-market agents: commit counts, API stacks, monthly costs, and live demos. We run SaaStr with 3 humans and 20+ AI agents. Here are the 10 takeaways, anchored in the numbers.

1. 10K replaced an entire BI workflow for $257 a month

10K, our AI VP of Marketing, runs at $257 a month. That is roughly $3,084 a year, or under 3% of one loaded marketing-analyst headcount. It started in January as a dashboard hooked to Marketo and Salesforce, and now owns the number, runs daily forecasting, tracks every campaign in real time, and pushes our top 3 marketing ideas every morning.

-> Price every agent against a loaded salary, not against a free tier.

2. One agent handled 402,000 interactions a team of 3 could never staff

Amelia AI, our inbound agent running on Qualified, fielded 402,000 chat interactions across 2.25 million sessions on annual.com for a single event cycle. Three people cannot physically field 402,000 conversations. To do it with humans you’d need a rotating BDR floor, and even a strong BDR handles a few dozen meaningful conversations a day at best. Efficiency is the wrong frame here. At three people, 402,000 conversations is a volume we could not reach with humans at any speed.

-> Deploy agents where the work exceeds human capacity, not only where it’s cheaper.

3. 614 meetings at ~$85K each is ~$52M of theoretical pipeline

Amelia AI booked 614 qualified meetings for this Annual. At our ~$85K average sponsorship, that’s roughly $52M of theoretical pipeline created by one agent. Obviously not all of it closed, or we’d have a very different sponsor number. Set it against headcount: a strong BDR books 10 to 15 qualified meetings a month, so 614 in one event cycle is on the order of 3 to 5 BDR-years of booking output, compressed, with almost zero complaints on the booking experience.

-> One agent can produce BDR-years of booked pipeline in a single cycle.

4. ~1,000 commits in ~120 days, and lines of code told us nothing

10K is about 4 months old with roughly 1,000 commits, 7 to 8 a day. Annie, our event site turned agent, went from 18,000 to 45,000 lines of code in two weeks, about 1,900 new lines a day. Any engineer will tell you 7 commits a day and a code base tripling in a fortnight is slop, and some of it is. It doesn’t change the output. We’re improving one internal application daily, not shipping to a million users on a fragile base.

-> Judge agents by output, not by commit count or lines of code.

5. A 20% markup plus a 20% discount, and the slide that costs you $1,400 a deal

We hate discounting, and the data says we have no choice, so we mark up and discount in a controlled band. The slide is where the money goes. A rep who smells a deal slipping goes from a planned 20% off to 25, then 30, then 34, and the data says it doesn’t move the close rate. On a $10K ticket, sliding from 20% to 34% off is $1,400 of pure margin handed over per deal for zero measurable lift. Across 100 deals that’s $140K gone. Amelia AI gives the right discount inside hard rules and never panics its way down the curve. Think real-time CPQ.

-> Guardrail the discount and you stop paying for the panic.

6. Artisan recovered $500K from B-leads humans never touch

Your A-leads don’t need an agent. A million-dollar inbound gets a reply from your laziest rep in 60 seconds. The gold is in the B-leads: real signal, real score, but never worth a human’s time, so they rot in the database. We pointed Artisan at exactly those and it returned an extra $500K. Against a cost in the low thousands a month, that’s $500K out of leads that were worth zero to humans, because no human would ever work them.

-> Point sales agents at the B-leads, not the deals humans already chase.

7. QB manages 150 accounts at <90 days old, with zero Salesforce data wired in

QB, our AI VP of Customer Success, is under 90 days old and manages roughly 150 sponsors with personalized outreach. No human CSM wants 150 accounts; they want 5, plus a renewal commission and an incentive to hide problems. Asked live, with no prep, which sponsors were most at risk of not renewing, QB flagged the accounts that never logged in, went dark, or complained most. We’d put it in the top 15% of CSMs we’ve worked with. It produced that risk analysis with no Salesforce data connected yet. The economics of one CSM per 150 accounts simply don’t exist with humans.

-> An agent can carry 150 accounts at the attention a human gives 5.

8. Same data, same specs, two builds, two different strategies

We rebuilt 10K twice in one week on the same inputs. The Replit version came back biased toward email marketing. The Lovable version came back biased toward ads and outbound spend. Identical specs, identical data, divergent recommendations. As teams start running parallel agents against the same data set, treat output as non-deterministic by default. You cannot assume two agents on the same brief converge, and that has real implications for how you review and reconcile what they tell you.

-> Two agents on the same brief won’t always agree, so reconcile rather than assume.

9. A 1,500-person blast from a banned address, and the cost of speed

Under time pressure, 10K pulled a clean VC and founder list from 10,000 records in minutes, even caught and corrected its own error (it had confused a CRM company with a similarly named fund). Then it sent the invite to 1,500 people from an email address banned in its rules for years, and the send was irreversible. A new marketing hire makes the same mistake. An agent can fire a 1,500-person blast in the time it takes to approve it.

-> Scale your review window to the blast radius, not the task.

10. Highest open rate came from the most context, not the best copy

Agent Force (we call it King Boo, it works our ghosted leads) posts our highest open rate of any agent. Context is why. Running on Salesforce, it holds all our Salesforce, Qualified, and Momentum data in one place, since Salesforce now owns both Qualified and Momentum. Open rate tracked context volume more than creative. If you’re on Salesforce, the path of least resistance to a high-context agent is already sitting in your stack.

-> Open rate follows context, so put the agent where the data already lives.

Five more things worth pulling out

**Almost none of these started as agents.**10K was a dashboard, QB was a sponsor asset tracker, Annie was a website on Squarespace. Build the dumb tool that kills one annoying task, keep shipping, and the agent shows up on its own.**Headless Salesforce is the highest-leverage hour you can spend.**Connect Replit, Lovable, or V0 to the Salesforce API and build the dashboard the UI won’t give you. Jason doesn’t have a Salesforce login anymore; 10K hits the API directly.**The more you engage an agent, the better it gets.**Daily reps make the context window richer and the output sharper. The set-and-forget pitch is backwards, and it’s how people end up with an autonomous agent quietly doing the wrong thing at scale.**One source of truth ends the sales-versus-marketing number fight.**When the dashboard pulls straight from the API, nobody argues about whose number is right or which date range got used. The number is the number.**Most of this is clever, not complicated.**Behind the agents are forms that look like other forms, plus smart wrappers and guardrails. A year from now you’ll realize you could have built it yourself, so build it yourself.

The unit economics in one paragraph

A $257-a-month agent owning the number. One inbound agent fielding 402,000 interactions and booking 614 meetings worth ~$52M in theoretical pipeline. $500K pulled from dead B-leads. A 150-account CSM that’s 3 months old. Run those against headcount, and waiting starts to look expensive. Watch the full episode for the back ends and live demos on The Agent