Your Agents Should Beat Your Best Reps, Not Match Them. And 14 Other Hard Truths About B2B + AI Today, From Our SaaStr AI Annual AMA

We closed out SaaStr AI 2026 with our deep dive AMA, and the theme this year: software is the least static it has ever been.

The products we run change more in a month now than they used to in years. The old playbooks work but they are just … too slow. Planning once a year, winning the land grab, hiring ahead of growth, letting sales earn trust slowly. None of it survives a market that resets every few months. And for the first time, agents aren’t just cheaper than people, they can be better.

The top takeaways:

1. Stop building agents to hit 80% of your best rep. Aim for 120%.

For a year we built every agent to reproduce 80% of our best human. That’s the wrong target. Two weeks before Annual, 10K, our AI VP of Marketing on Replit, wrote outbound emails that assembled the top 40 people worth meeting at the event and reasoned through why. The copy was good. The analysis underneath is what got me. Nobody on my team could hold that much in their head. It was the first time I watched an agent do something clearly better than a person, not cheaper, better.

After that, I started spotting it everywhere. Our inbound agent booked 682 real, qualified meetings, better than any BDR I’ve ever worked with, and it never lowers the bar at the end of the month to hit a quota. Social selling was always slop. Social support is a 120% move: an agent that reads my OAuth gripe, checks the actual account, and hands me the fix inside the comment does something no human ever could.

→ What to do:Audit every agent against one question, “where could this beat my best human instead of just matching them?” Then fully automate inbound before anything else.

2. Planning is mostly dead. The best startups plan weekly now.

Planning has great ROI when the world is stable. In 2021 any 20 year old product grew, so you’d ship one big release a year, plan hard, and staff up capacity. Ask the fastest growing companies at Annual how often they plan today and they’ll tell you weekly. When the world is unstable, every hour you spend planning is an hour you’re not shipping better product. We’re not in product-led growth anymore. We’re in a product-driven world, where products that were impossible a year ago have massive market pull.

→ What to do:Cut your planning cadence to weekly and pour the saved hours into shipping product faster. If you run short on reps, hire them the next day.

3. Don’t hire a “GTM engineer.” Find your internal tool nerd.

A company with about 1,000 employees asked how to hire their first GTM engineer. My answer: don’t post the job. Those people mostly don’t exist yet, and the few who do don’t want to work for you. Amelia runs all our systems and she couldn’t have done it a year ago, because the tooling didn’t exist. Look inside instead. Find the one person in marketing who hacked something together through Zapier, built a dashboard nobody supports, and is curious enough to keep pushing. Put them in charge of the agent stack.

→ What to do:Before you open the req, hand the agent stack to the curious tool nerd already sitting in your marketing team. Only hire outside if literally nobody inside can do it.

4. Vertical AI wins because the agent becomes a better domain expert.

VCs make too big a deal of vertical software. It has been around since software shipped on floppies. What’s new is that a vertical agent can be a far better domain expert than a horizontal one, because you can load it up. Owner’s AI CMO doesn’t need to know everything a CMO knows. It needs to know how restaurant menus, delivery, prep, and upsell work, which images make people buy, and what order to present things in. It even needs to know that guacamole is one of the highest margin add-ons at a Mexican restaurant. That focus is why a specialized agent beats a general one.

→ What to do:Load your vertical agent with deep, specific domain knowledge (your industry’s version of the guacamole margin) instead of building it wide and shallow.

5. Your sales team has to know the product cold now.

The most valuable person in a B2B sale is the one who knows the product cold. I got to know Cody, a forward deployed engineer at Replit, because I hit a bug. The second time we talked, he explained exactly why I was having the issue and how to get around it. Compare that to a legacy vendor that put Amelia on the phone for four hours, blamed her, and wouldn’t help unless we booked another four. When products change every week, a rep who answers like Cody is irreplaceable and a rep who can’t gets bypassed.

Golf and schmoozing are done, and you can’t build a real relationship over Zoom. Last summer I asked Marc Benioff on the 20VC podcast what he wanted in the agent era, and in an unguarded moment he said he wished every customer had a forward deployed engineer and didn’t have to pay until they were deployed.

→ What to do:Make product mastery the hiring bar for sales, and get customers to real value fast, ideally before the contract is signed.

6. Track the slope of improvement, not today’s slop.

A CFO tool founder pushed me on hallucinations and slop. She’s right that a lot of it doesn’t work yet. But the number to watch isn’t today’s slop, it’s the slope. Everything we’ve built would be a Squarespace site without the messy stuff underneath it. The value to slop ratio improves every single week. LinkedIn comments are a small example. Three months ago they were “go team” garbage. Now they’re good enough that you almost want your AI to talk to their AI. For real products the value is 50 times the slop, and that ratio keeps moving.

→ What to do:Judge a tool on how much it improved over the last 90 days, not on the bugs you can find in it today.

7. AI didn’t kill engineering hiring. It started an arms race.

Engineering demand peaked in 2020 and 2021 when budgets were infinite and nobody had to be profitable. It dipped, and now it’s on a tear again, up 20 to 30% in a lot of places. You can build basic software with far fewer engineers now, so the bar for what’s impressive shot way up. Replit and Lovable are both near $500M in 18 months, and neither is phoning it in. Everyone will hire every 10x and 100x engineer they can find because the productivity is insane.

What’s collapsed is demand for the B-tier CS grad. My son at Penn State is off the charts in math, and labs recruit him for what he’s written. His CS classmates have zero offers. My daughter’s at Stanford and nobody’s going into CS. We don’t need fewer engineers. We need fewer 2021 engineers and more great ones.

→ What to do:Keep hiring every great engineer you can, and stop hiring the B-tier engineer who needs three months to build an admin page.

8. It’s not a land grab. It’s an inertia grab.

There’s no land grab in AI, because customers know to re-evaluate every vendor in 8 to 10 months. The budget is there, and so is the willingness to switch. What there is instead is an inertia grab. We’re Replit super fans, and the odds we move our agents off it approach zero. Not because of data lock. Because we’ve gotten so good at it that a competitor would have to be twice as good to be worth the switch, or Replit would have to stop innovating.

Being great at your tool matters more than picking the one that’s 11% better. And customers don’t just buy your product as it is today, they buy your roadmap for the next 24 months. That’s the real moat now.

→ What to do:Get in market during this window, and get deep on your core tools. Mastery and roadmap trust are what keep customers, not lock-in.

9. In venture right now, nothing matters but growth.

All of venture capital right now comes down to growth. Not gross margins. Not whether the revenue recurs. Not multi-year contracts. Great founders and growth, and that’s it. Rippling went from $2M to $500M in two years. People are wiring money into companies with 48 hours to decide and no chance to talk to anyone.

The hard part: growth that would have gotten funded 18 mo