There’s a growing wave of AI agent skepticism on LinkedIn right now. And honestly? Some of it is earned. A lot of founders bought an AI SDR, didn’t train it, and got garbage results. Then they posted about how “AI agents don’t work.”
But here’s what I know after 8 months of running 20+ agents across our entire go-to-market at SaaStr — with just 3 humans and a dog: $4.8 million in additional pipeline sourced by agents. $2.4 million in closed-won revenue. Deal volume more than doubled. Win rates nearly doubled. And none of it cannibalized our existing inbound.
It works. But not the way most people think it does.
Let me break down what we’ve actually learned — the real stuff you won’t see in the LinkedIn posts.
The Results Are Real, But So Is the Work
Let me give you the honest numbers first.
Eight months in, our AI agents have generated $4.8M in additional pipeline and $2.4M in closed-won revenue that was first-touch sourced from an agent. Our deal volume has more than doubled. Our win rates have nearly doubled. And we’ve sent over 60,000 high-quality AI-generated emails just on the sales side — not even counting the nearly 1 million interactions through our vibe-coded apps.
Here’s what matters most about those numbers: this was all additive. It did not cannibalize our other inbound revenue sources. We didn’t drop anything when we deployed these agents. We still send marketing emails. We still do outbound ourselves. We still send gifts. We still invite people to the SaaStr house. All the things we used to do before — we still do them. The agents augmented everything.
But here’s the honest truth you won’t see on LinkedIn or X: we maintain these agents every single day. Literally every morning before anything else, we’re checking our agents. Amelia and I each spend 15-20 hours per week — that’s each, not combined — actively managing, iterating, checking responses, making sure nothing hallucinates, making sure the agents are talking to people the way we want them to.
The time we used to spend managing humans on our team? We now spend that same amount of time — if not more — managing agents. The difference is there’s no people drama, and the agents work at a much higher capacity and scale than a human ever could.
At some point, you realize you simply cannot keep up with your agents. They’re faster than you. They work 24/7/365. They can always answer a question, always book a meeting, always reach back out. The humans become the bottleneck.
The Secret Nobody Tells You: Agents That Require Deep Training Cannot Be Self-Trained
I was meeting recently with the CEO of a next-generation AI go-to-market company — they already have millions in revenue and are publicly launching soon. I asked what their secret sauce was.
The answer: they do everything. The onboarding, the tagging, the first campaigns — all of it. They do it almost to a fault. Some customers think it’s too easy and don’t even realize how much human energy is going into deployment behind the scenes.
That’s the learning. If you haven’t deployed many agents — or any for real — you need to have an honest conversation. Not with someone in sales who doesn’t know how the product works. Talk to a forward-deployed engineer. Talk to a leader. Find out what it’s actually going to take in the first 14 days, the first 30 days, and every single day after that.
Then you have to actually do it. Otherwise, it’s like going to the doctor, getting a prescription, and never taking the medicine. It literally will not work.
A lot of the agents we use are pushing downmarket to be more self-service. So far, that doesn’t work. Agents that require deep training cannot be self-trained yet. It will come — agents are getting dramatically better every quarter. But for now, be skeptical. If you buy a cheap tool that claims it’s self-trained, make sure it actually works. If you buy a more complex tool, talk to someone senior enough on deployment who actually knows the product.
The 90/10 Rule: Buy 90%, Build Only 10%
Here’s our rule of thumb: buy 90% of your AI stack. Only build the 10% where no vendor can do it well and it’s a P1 priority.
We’ve followed this ourselves. The vast majority of our agents are third-party tools that we’ve trained and customized heavily. We only built custom agents where we had a very specific use case that no vendor could handle — like our AI VP of Marketing (more on that below).
Kyle, the CRO at Owner (one of our portfolio companies), has followed roughly the same approach. He bought a bunch of third-party agents, made them work, and then hired a former founder/engineer — someone who was literally a CEO of an LLM company — to build a proprietary in-house tool for the 10% that needed to be custom.
That’s an extreme case. For most of you, building custom agents probably won’t make sense yet. Focus on making the bought tools work first.
How to Evaluate AI Agent Vendors (Don’t Skip the Basics)
I don’t know why people throw away basic evaluation practices just because something has “AI” in the name. Here’s what we do — and what you should do:
Ask for a dedicated resource. I asked every single AI tool we now deploy for help. I told them: one, I’m going to need an FDE (forward-deployed engineer), and two, let me talk to people who have actually used this.
Ask for customer references. I see too many people skip this because “it’s an AI tool.” Ask for a reference. Ask for one in your vertical if possible. If they push back, maybe don’t use that vendor. Most of these companies have at least one customer that’s somewhat like you.
Demand FDE time at deployment. Salesforce put an FDE on our success team. Qualified has an FDE on our success team. Replit has an FDE on our success team. With Artisan, anytime I have an issue or idea, I go straight to the CEO or head of product. You should ask for — and expect — some level of dedicated support to get started. You may not need it weekly later, but you absolutely need it upfront.
Trust your gut. If it doesn’t feel right, don’t buy it. If your Spidey sense says this agent isn’t going to work, it won’t work. Buy another one. Even if the brand is less well-known, even if it’s scrappier — if it feels right and the team is proud to show you their product, go with that one.
One more thing on evaluation: the best agents should take you as far down the journey as possible before you have to pay. A lot of AI go-to-market tools can’t practically offer free deployments due to the human onboarding costs. But the best ones get you as close to production as they can. Marc Benioff said on 20VC that he wished Salesforce had enough FDEs to get every customer into production on Agentforce before they had to pay. It’s not practical, but that’s the right instinct.
Multi-Agent Management Is Messy (And That’s Okay for Now)
Here’s what people always ask me: “You have 20 agents — what are you using as your MCP?”
The honest answer: we don’t have one. Not a real one, anyway.
What we have is what I’d call “MCP light” — a combination of Zapier webhooks, Salesforce as our system of record, and a lot of manual context-sharing between agents. We have so many webhooks firing into our Zapier account I can’t even count them. All of our third-party tools push data back to Salesforce, either through native integrations, APIs, or Zaps.
Sometimes I just copy-paste context from one agent and put it into another. It’s not elegant. It’s not clean. But it works.
This is the reality of multi-agent management in early 2026. It’s a lot of hodgepodging things together. But I think this is a “right now” problem, not a “forever” problem. By the second half of 2026, I believe native integrations will solve most of this.
My advice: Pick one source of truth for your data (Salesforce, HubSpot, whatever) and push everything back to it. Get used to your agents talking to each other — it happens, and it’s fine. Get used to talking to your agents yourself. And for now, get comfortable with some manual context-sharing between systems.
If you use specialized tools like we do (versus an all-in-one agent builder), you’ll deal with more of this messiness. We prefer specialized tools because the output quality is better. But you might reasonably trade some quality for quality-of-life by using an all-in-one platform. Both approaches are valid.
A Practical Go-To-Market Flow You Can Copy
Here’s an actual Zapier flow we run that you could adapt:
- Catch a webhook from a website form submission
- Push to Google Sheets (yes, I keep backups of everything in Sheets — don’t judge)
- Create/update a contact in Salesforce
- Add the contact to a Salesforce campaign (which can trigger Agentforce if you want)
- Find account-level records in Salesforce to see what this company has done with you before
- Enrich with Clay — pull LinkedIn activity, summarize context
- Send a Slack notification with all the context: contact info, account history, Clay enrichment
- Optionally generate a Gamma presentation or landing page customized for this prospect
- Draft a Gmail or push to your AI SDR platform
That’s nine steps, touching six or seven different tools. It’s not simple. But it gives you a fully enriched, context-rich go-to-market motion that would be impossible to do manually at scale.
The AI SDR Playbook: Hyper-Segmentation Is Everything
If you’re rolling out your first AI SDR, here’s the single most important thing I’ve learned: hyper-segment everything.
I see people running one campaign for 10,000 leads. That’s insane. I max each campaign at 100-500 contacts. Every campaign, every sub-agent gets highly customized training for the exact segment it’s targeting.
Don’t segment the old-school way — by geography, title, or role. None of those exist in my outbound funnel. Instead, segment by context and intent:
- Website visitors (deanonymized)
- Inbound leads who filled out a form
- Abandoned trials or carts
- Event leads from webinars or conferences
- Former customers who changed jobs
- Current customers for expansion
- Recent marketing leads from gated content
- Leads you never followed up with (we famously gave these to Agentforce)
- Lapsed contacts you forgot to talk to
- Low-scoring leads that still show intent but nobody wants to call back
- Alumni from previous events
- Warm referrals and community members
I haven’t exhausted this list after 8 months. Start here, not with cold outbound. Your AI agent has zero context for why it should reach out to a random cold lead. But it has rich context for every segment above.
The more context you give your agent, the better the output. This is no different than how you use Claude or ChatGPT every day — you talk to it, give it context, tell it about your business. Same rules apply with AI SDR agents.
And tell your agent what you can’t do. This is a nuance I only learned after months of deployment. AI agents are self-gratifying — they try to beat their own metrics. Over time, they start making promises you can’t keep. Ours started offering people speaking slots at SaaStr Annual, which isn’t how our speaker selection works at all. Once I explicitly told the agent what we don’t do, the quality jumped significantly.
Bad context equals bad emails. Period. Train on the best of everything — your best email copy, your best follow-up sequences, your best case studies — and then also define the boundaries clearly.
Why We Built Our Own AI VP of Marketing
We looked at every third-party AI marketing tool on the market. None of them could do more than content. The real problem we needed solved was orchestration — coordinating data across agents, prioritizing initiatives, and creating executable daily plans grounded in actual performance data.
We also knew from experience that every time we tried to onboard a human with all our data, they got overwhelmed. So we built our own AI VP of Marketing. We nicknamed it “10K” — because the goals were 10,000 attendees at SaaStr Annual and $10M in revenue for the year.
Here’s how we built it:
- Collected data from our agents, third-party tools, Salesforce, Zapier workflows, and internal historical data
- Fed it into Claude Opus with very clear goals and context (this required upgrading to Max — Pro ran out of memory over a weekend of heavy analysis)
- Pushed the output to Replit so the whole team could access it as a web app
- Told it to generate both high-level strategy and granular daily executable tasks
The key instruction: “Don’t give me generic strategy ideas. Give me executable tasks grounded in our data that three humans and a dog can actually do.”
What came back was remarkable. A week-by-week game plan with specific campaigns, specific timing, specific channel allocation. It told us how much to spend on LinkedIn ads and what the ads should say. It told us which campaigns to run through Artisan vs. Qualified vs. Agentforce. It literally told me what to post on social media.
Some stuff it said to blow up or abandon. Some stuff it said to bring back. And a bunch was net-new. Because it was all rooted in data, I trust it.
I talk to 10K every day. “Where are we? What should we be doing today? Where are we falling behind?” It keeps me honest and focused. Sometimes I push back — it once suggested a campaign I didn’t think was compelling enough. We debated it, it looked at the data, and agreed to change course.
It’s not always right. But it’s not always wrong either. And it keeps me more organized in this one vector than I’ve ever been.
The Maturity Curve Is Real
Here’s the honest assessment of where AI tools stand today:
Coding and support tools → Most mature. These work well today.
Sales tools (AI SDRs, conversational agents) → Getting there. Work well with significant training and maintenance.
Marketing tools → Not nearly as mature as vendors claim. That’s why we had to build our own AI VP of Marketing. Most marketing AI tools only handle content. Orchestration, campaign planning, cross-channel coordination — none of that exists in a turnkey product yet.
By the second half of 2026, I believe all of this will connect natively. Your AI VP of Marketing will integrate directly with your AI SDRs, your Agentforce, your Clay tables, your LinkedIn ads — all of it. There won’t be any excuses for shooting from the hip in B2B marketing anymore.
But we’re not there yet. For now, it’s messy, it’s manual in places, and it takes real human effort every single day.
Start With Something That Just Isn’t Getting Done
If I had to distill everything we’ve learned into one piece of advice: find something in your go-to-market motion that just isn’t getting done, or is getting done at a mediocre level, and put an agent on it.
Don’t try to replace what’s already working well. Do that as your 10th or 20th agent. Start with the low-hanging fruit — the customers that are too small for your team to call back, the leads that take too long to respond, the contacts with lower scores that still have intent but nobody wants to prioritize.
Because even modest yield from those segments is magical. It’s entirely additive revenue that you never would have captured otherwise.
The emails our agents send aren’t the best ever written. I’d say they’re “pretty good.” But we’re getting scale — 60,000 high-quality, personalized emails that we could never have done manually. More high-quality, pretty-good interactions with more people, more often. That’s the formula.
And that’s worth $4.8 million in pipeline and counting.
