Vercel Runs 'Our Internal Sierra' for $150K a Year. Write the Invoice Nobody Sends.

Vercel's COO put a price on building your own support agent this week: about $150K a year, against vendors she says cost much more. This is a Ship Story: what Vercel decided, what the number holds, and the one page I would add to the decision.

The short version

• Jason Lemkin's November 6 write-up of Jeanne DeWitt Grosser's SaaStr AI talk says Vercel took lead qualification from 10 people to 1 in six weeks, raised SDR quotas 30% the quarter after, and runs Vertex, a support agent she called "our internal Sierra," on 93% of total case volume for $300 a month in infrastructure and about $12,000 a month in tokens, about $150K a year all-in, with three engineers.
• Twelve months of those two costs is $147,600, so the $150K is tokens and infrastructure and the engineers are outside it.
• At the $1.75 per resolved ticket I charged, that bill equals about 1,600 resolutions a week, and Vercel says thousands of cases a week, so on tokens the build probably wins. What the number leaves out is the engineers and the counterparty.
• Handled is not resolved, the best vendors in Gorgias's benchmark resolve 64% to 75% with zero touch, and when I billed per resolution 2% of cases became more than 200 disputes a week, which was my customers auditing my definition. An in-house agent gets no invoice and no dispute.
• Grosser already asks every quarter whether anyone has built something better, faster, or cheaper. Give the question a page: cases under three definitions, tokens and infrastructure, engineer time, and one vendor quote on the zero-touch count.

What Vercel shipped

The decision first, as Lemkin reports it. Grosser set up a go-to-market engineering team six weeks after she joined, with a mandate to put agents into every GTM function. Every agent gets built by the same three people: a GTM engineer, a data scientist, and the single best performer in the function being automated.

The lead agent came out of a few days of shadowing Vercel's best SDR and logging every tab she opened. It ran in shadow mode for six weeks while she reviewed every output. When she had nothing left to correct, they took her out of the loop. The team of 10 became one person for the US and about 20% of another for Europe and APAC. Cost: under $5,000 a year in infrastructure and tokens and about 20% of one engineer to maintain. Grosser's math is a 32x return against the cost of the 10-person team, and she says those people moved into more valuable roles.

Then 30 SDR workflows got the same treatment, and SDR quotas went up 30% that quarter. The content agent did 96% of major marketing content updates last quarter.

And Vertex, the support agent behind vercel.com/help. Vercel started with an off-the-shelf product, did not get the results it wanted, and built its own in two months. It handles 93% of total case volume, thousands of technical cases a week. Lemkin's line items: $300 a month in infrastructure, about $12,000 a month in tokens, about $150K a year all-in, three engineers on it. Grosser compared that with agentic support vendors that have 150 engineers running equivalent workflows at much higher cost, and set a benchmark for her team: the average builder should be able to out-ship any vendor selling the same outcome.

$150K a year

That is a real result with real numbers, published by the operator. I have no quarrel with the decision.

What the $150K holds

Multiply it out. $300 plus $12,000 is $12,300 a month, and twelve of those is $147,600. So "about $150K a year all-in" is the token bill and the hosting bill. The three engineers in the same paragraph are not in it.

The lead agent has the same shape. Thirty-two times $5,000 is $160,000, and ten people cost more than that, so the denominator in the 32x has more in it than tokens. The write-up does not show the math.

I am not saying the build loses. On tokens it probably wins. I priced support automation per resolved ticket, at $1.75, in the redesign I described in Per-Outcome Pricing: What Gets Clearer and What Gets Terrifying. At that price $147,600 buys about 84,300 resolutions a year, roughly 1,600 a week. At the $0.90 per resolution that Gorgias bills, it buys about 164,000, roughly 3,150 a week. Vercel says thousands a week. If thousands means more than 3,150, Vertex's token bill per case is under the lower of those two prices.

So the comparison turns on the two things the number leaves out. One is three engineers. The other is harder to see.

Handled is not resolved

The write-up says Vertex "handles" 93% of total case volume. It does not say what handled means.

I wrote One Support Agent, Three Resolution Rates about exactly this word. Gorgias tested 13 vendors on 212 live stores, and under a zero-human-touch rule the top five resolved 64% to 75% and the median 48%, per Lemkin's own write-up of that benchmark. Under a looser rule, where a conversation counts as resolved once 72 hours pass without a human agent, a customer who gave up counts. Same agent, different number.

When you buy from a vendor priced on the outcome, the definition is on the invoice. Sierra, as I laid out in The Sierra Playbook: What They Actually Do Differently, charges a pre-negotiated rate per resolved case with escalations free, so its revenue depends on the definition holding. And the customer gets to push back. I billed that way. My 48-hour definition held for 98% of cases, and the other 2% became more than 200 contested cases a week.

Those disputes were work. They were also my customers auditing my number every week, with their own money on the line. It was a review I never had to schedule.

Build the agent yourself and nobody sends the invoice. Nobody contests the 93% either. The SaaS to AI Business Models thesis is that pricing moves to the outcome and the margin structure moves with it. The part people skip is that the outcome price came with a counterparty, and the in-house build quietly removes one.

The invoice nobody sends

Grosser already has the habit that fixes this. Every quarter she asks her GTM engineering team whether anyone has built something better, faster, or cheaper, and if so Vercel switches. That is the right question. It needs a page to answer with.

Four lines, once a quarter, for any agent you built instead of bought.

Cases, under three definitions. Zero human touch. Your own working definition, the one behind the 93%. And no related follow-up from the same customer in seven days. Thirty sampled conversations scored three ways is enough to see the spread.

Tokens and infrastructure. The $147,600.

Engineer time. All three, at whatever share they spend, plus the best performer's review hours while the agent was in shadow mode.

One vendor quote, priced on the zero-touch count from line one. Not the vendor's own resolution claim. Yours.

That is the invoice a vendor would have sent, written by the only party who can write it now. Ramp Built Its Harness in Six Weeks. Kill the Build-Versus-Buy Memo. has my three lines for before the build: the first internal customer by name, the person who reads the traces every week, and the on-call roster. Vercel's three-person build answers the first two by construction, because the best performer in the function sits in it. This page is for after, when the build is running and the only number in circulation is the token bill.

The reviewer who left

One more date. The lead agent's exit rule was that the best SDR had nothing left to correct. That is a good bar for leaving shadow mode. It says nothing about week twelve.

A classifier of mine scored 89% on held-out data and was at 61% in production six weeks later, which I wrote up in Rules 97.5%, Trained 100%, Model 87.5%. The Bake-Off Table Needs a Slice and a Date.. Nobody was reviewing by then. Vercel's own inbound agent, as I wrote in Agents Read the Design System. Read Is Not Checked., stopped following some of the rules in a 1,000-line prompt, and 14 of them moved into code. So Vercel knows agents drift. Put the reviewer's return on the calendar: a sample of outputs on a set day each quarter, scored by the person who trained it. Line one of the invoice can come out of the same sitting.

This week, if you run an agent you built instead of bought: write down what is inside your "all-in" number, then score 30 of its closed cases three ways. If the three rates are close, you have a result worth publishing. If they are not, you just did the audit a vendor's customers would have done for you.

Related answer: Is building your own support agent cheaper than buying one?

Sources: Jason Lemkin, "Top 5 Learnings from Vercel's COO: 10 SDRs Down to 1, a $5K-a-Year Lead Agent, and 93% of Support Cases Handled by AI," SaaStr, November 6, 2026, read in full; every Vercel figure and every statement by Jeanne DeWitt Grosser is as Lemkin reports it from her SaaStr AI talk. The $147,600, the weekly resolution equivalents, and the $160,000 are my arithmetic on those figures. The Gorgias benchmark figures and the $0.90 per resolution are as reported in "One Support Agent, Three Resolution Rates." The $1.75 per resolved ticket, the 98% and the 200 contested cases a week, and the 89% and 61% are from my own work, in the posts linked above.