B2B AI Agent定价法则：必须翻倍ACV否则不投

This week one of our portfolio companies launched its first real agent. It’s a vertical software company at about $30M ARR, growing well, with a loyal base and a sales-led motion.

The agent doubles ACV. There’s no free version. In the first week, 3% of the customer base bought it.

I think that’s what a great agent in B2B looks like: customers pay twice as much in the first week because the value is obvious.

I now invest by this rule. I won’t back an agentic B2B company unless it unlocks budget beyond the software budget, and the proof is deal sizes at least 2x what the pre-AI competitors get. Most companies shipping AI right now don’t clear that bar.

Top takeaways:

• A real agent does a job a human used to do. Customers pay for it out of labor budget, which is why it can double ACV.
• AI features that make the existing workflow a bit better don’t move ACV, so vendors meter them with credits on top of seats. Customers experience that as a price increase.
• “Launch free, monetize later” was a reasonable hedge in 2024. In 2026 it usually means the product isn’t ready to charge for.
• If you can’t double ACV, adoption stalls, or customers resent the bill, fix the agent before you touch the pricing.

1. An Agent That Doubles ACV Gets Paid From the Labor Budget

Software budgets are capped. Every CIO I talk to is consolidating vendors and cutting existing lines to fund AI. An agent that competes for the software budget is fighting for a shrinking pool, often against its own seat line.

An agent that does real work gets paid from a different budget: the person who used to do that work, the agency, the contractor, or the task that never got done because nobody had time.

That’s how this company can double ACV without pushback. Customers compare the agent to the hours it replaces, and against that number, 2x the subscription is cheap.

It’s also why 3% adoption in week one matters. With a sales-led motion and some onboarding, the whole base was never going to switch in seven days. But 3% of a loyal base paying double in the first week, with no discount and no trial, shows the value was clear before anyone had used it for a month.

2. 10K Co-Sold a Six-Figure Deal, and Our Sponsor Renewals Went From the 20s to the 70s

I believe this partly because we’re seeing it at SaaStr.

10K, our AI VP of Marketing, has been running for 11 months. On a six-figure Base44 / Wix deal, it built the decks, priced the custom work, flagged the procurement terms, moved the deal to Commit in Salesforce, and filled out the vendor paperwork. It co-sold the deal alongside me and Amelia.

Renewal rates for our $100K+ sponsors went from the 20s to the 70s this year. 10K says it only deserves credit for “operational support.” Either way, that’s what agent-level value looks like from the buyer’s side. If someone sold me 10K as a product, I’d pay several times what we pay for any single tool in our stack, because I’d be comparing it to the humans we didn’t hire.

Very few vendors selling us AI today are in that category. Most are selling a slightly better version of what we already had.

3. HubSpot Tripled a Small Team’s Bill With Credits

When the AI doesn’t do a job on its own, vendors can’t charge 2x, so they meter it.

HubSpot’s new model is this fall’s best-known case. Amelia priced it for a three-person marketing team with 10,000 contacts, a weekly newsletter and about 20,000 workflow actions a month. The old bill was about $1,300 a month. The new bill is about $3,700, mostly credits for workflow actions that used to be unlimited. A newsletter send costs the same credit as an outbound sequence. One HubSpot customer we know called to say the newsletter was moving elsewhere, and got a discount.

The bill tripled, and the customer got no new capability.

Compare the CRM-focused agents in our network: Monaco, Lightfield and Attio. They’re closing much larger deals than in the Pipedrive era because they take over real work. If an agent puts four meetings on your calendar the first week you turn it on, nobody argues about the price.

When customers complain about the bill, the agent usually isn’t doing enough. When they thank you for it, you can charge more.

4. Keep the Seat as the Seat. Sell the Agent as Its Own Line.

Products sold per seat should mostly stay per seat. Customers like predictable seat pricing and dislike surprise charges on top of it. Products that were always transactional, close to revenue or data usage like Stripe or Databricks, can be usage-based, and customers accept that.

Trouble starts when vendors want the economics of variable pricing without agent-level value, so they add per-action fees on top of an $80 or $400 seat. The base sees that as a price increase, because it is one.

The companies doing this well keep the seat price where it is and sell the agent as a separate SKU, priced by the job it does. That’s the path to 2x.

5. Salesforce Charges a 60-80% Premium, Box 20-40%, and Podium Built $100M in Agent ARR

The incumbents show the range.

Salesforce charges a 60-80% premium for its premium AI editions. As of last quarter, only about 5% of sales and service knowledge workers were on them. Benioff said it directly: “Most of them have not started their AI transformations yet.” The premium exists, and most of the base hasn’t decided it’s worth paying.

Box repackaged around AI with Enterprise Advanced at a 20-40% price premium. It reached about 10% of revenue in under two years, and billings growth picked up before revenue did. That’s a real upgrade path. It’s 20-40% because the product is a better version of the same job.

Podium built $100M in AI agent ARR in under 24 months selling to home services businesses. The agent answers the phone, books the job and follows up, work customers used to pay people to do.

Improving a workflow got Box to 20-40%. Podium’s agent does the work itself, and that’s what supports 2x.

6. In 2026, Launching Free Usually Means You Can’t Charge for It Yet

The standard advice is to launch AI free, get adoption, learn how customers use it, then add pricing. Figma roughly did that. At over $1B in revenue and about 92% gross margins, it launched AI free, added credits without charging, then started charging. Ulrik Lehrskov-Schmidt of Willingness to Pay, who joined us on The Agents this week, makes a good case that this captured value the seat model had been missing.

I’m more skeptical. Figma Make was uncompetitive with Replit, Lovable and v0 for a long time, and I think launching free was partly cover for a product the team didn’t think it could charge for. Canva was worse off. It was so worried about token costs eating an 80% gross margin that it never shipped anything competitive enough to charge for. Then the models got better, Gamma and Higgsfield took the use cases, and Canva restructured 22% of the company this month.

I’d make the team charge for it on day one, at the same price as the AI-native leaders or higher. If your own loyal base won’t buy it, you find out in one week instead of four quarters.

There’s one fair exception. Ulrik had a UK client at about $30M that launched its first agent with credits from day one, and adoption stalled. Now they can’t tell whether the product or the pricing failed. If you don’t yet know what the product should do, launch without a paywall to find out, and be honest internally that you’re still in product discovery.

7. Five Things a 2x Agent Does

From building our own agents and selling into a base of 400,000 B2B execs, the agents that earn 2x have five traits in common:

• They finish the job. The output is a closed deal, a booked meeting, a resolved ticket or a filed form, with no human needed to finish it.
• They use your data. 10K is valuable because it can see 13 years of invoices, every call note and every sponsor’s results. A horizontal LLM can’t.