Snorkel Stopped Selling The Tool. Revenue Grew 18x.

On September 22, Snorkel AI announced a $350 million Series E at a $3.5 billion valuation, led by Insight Partners and S32. Seventeen months earlier it was valued at $1.3 billion.

The funding number is not the interesting part. The interesting part is what Snorkel changed about what it sells.

For most of its life, Snorkel sold software that helped companies label their own training data. Last year, according to TechCrunch, it switched to delivering completed datasets and simulated training environments to customers. It calls this data-as-a-service. Its annualized revenue run rate is now $375 million, which the company says is an eighteenfold increase over the past 12 months.

A seven-year-old company stopped handing customers a tool and started handing them the result. Growth followed. If you run a startup that sells a tool for doing hard work, this is worth thinking through carefully, because the same move can either unlock your business or quietly wreck your margins.

Snorkel came out of research at a Stanford AI lab. The founders spent four years on it before launching commercially in 2019.

In 2021, when it raised a $35 million Series B, CEO Alex Ratner described the product to TechCrunch as a way for subject matter experts to apply labels programmatically. He said roughly 80 to 90 percent of the cost and time of an average AI project went into labeling and relabeling data. His software was built to shrink that from months to hours or days.

That was a tool pitch. It made the customer faster, but the customer still owned the work. They needed their own experts, their own process and their own time.

The 2025 shift flipped that. Now Snorkel produces the dataset itself, using a mix of its own software, models that generate data synthetically, and human subject matter experts. The customer receives a finished asset they can train on. TechCrunch notes the demand is being driven by AI labs that want high quality training data and will pay for it.

A tool asks the customer to do three things before they get value: learn it, staff it and run it. Each step is a place where deals stall and renewals die.

A finished output removes all three. The buyer compares your price against what the result is worth to them, not against the price of other software. That is usually a much bigger number.

Look at how Ratner framed the problem in 2021. If labeling was 80 to 90 percent of the cost of an AI project, a tool that made labeling cheaper captured a slice of that budget. A company that simply does the labeling can compete for the whole budget.

There is also a timing reason. When a market is new, buyers often lack the people to operate a tool well. AI labs moving fast want the dataset next month, not a new internal team. When the buyer's constraint is talent or time rather than money, selling the finished output fits how they actually buy.

Snorkel is not alone. The day after its announcement, TechCrunch reported that Ema raised a $77 million Series B led by Creaegis, bringing its total funding to $140 million.

Ema sells what it calls AI employees: systems that coordinate multiple agents to complete multi-step processes in HR, IT and finance. Three details from the coverage matter for founders:

1. Pricing. Ema does not charge by seat or by tokens. CEO Surojit Chatterjee said pricing is tied to completed tasks and business outcomes.
2. Margins. Despite doing work that used to belong to software vendors and IT services firms, Chatterjee said gross margins are close to 80 percent, and they improve as the systems learn from deployments.
3. Expansion. More than 90 percent of customers have expanded beyond their first use case, and net dollar retention is around 180 percent.

That last point is the quiet benefit of selling results. When one department sees a finished outcome, the next one wants it too.

Here is where founders get into trouble. Selling the finished output often means paying people to produce it. That changes what your revenue number means.

TechCrunch spelled this out in the Snorkel story. Several AI data companies report large gross annualized revenue: Mercor at $2 billion, Handshake at $1 billion and Micro1 at $500 million. But these companies pay roughly 60 to 70 percent of their top line directly to the domain experts doing the work. Their net revenue is much lower than the headline.

Snorkel says it handles this differently. Because it sells datasets and training environments rather than hours of human labor, payments to its experts are counted in cost of goods sold. The revenue is reported in full, and the cost of producing it shows up as a cost.

For a first-time founder, the lesson is not about accounting rules. It is about being clear with yourself and your investors:

• Know your contribution margin per delivery. If you charge $50,000 for a finished dataset and pay $35,000 to contractors to make it, you have a $15,000 business, not a $50,000 business.
• Report both numbers. Sophisticated investors will ask for gross and net. Offering both first builds trust.
• Watch the trend. The point of combining software with human experts is that the software share should rise over time. If your cost per delivery is flat after a year, you are running a staffing agency with a nicer website.

Not every tool should make this switch. Run through these five questions before you commit.

1. Do customers struggle to get value from your tool without heavy help? If onboarding calls, services hours and custom setup are already a big part of every deal, you are partly selling the outcome anyway. Making it explicit may be simpler.

2. Is there a clear, measurable result? A dataset, a closed support ticket, a reconciled ledger, a finished report. If you cannot define "done" in a sentence, outcome pricing will create endless arguments.

3. Does your software make you much faster than a human team alone? This is where margin lives. If your tool lets one expert do the work of five, you can charge near the value of five and pay for one.

4. Is the buyer short on people or time more than on budget? That is when they will pay a premium to skip hiring.

5. Can you survive a lumpier revenue pattern for a year? Project-based or outcome-based revenue can swing more than subscriptions early on.

If you answer yes to at least four, it is worth testing. Writing out the old model and the new model side by side, with unit economics for each, is a good use of your Foundra business plan before you change a single price.

Snorkel did not stop being a software company. It used its software to become better at delivering the result. You can test the same move in stages.

Step 1: Pick one segment. Choose the customer group that already asks you to "just do it for us." Leave everyone else on the tool.

Step 2: Define the deliverable and the acceptance test. Write down exactly what the customer receives and how both sides agree it is complete.

Step 3: Price on value, then check cost. Ask what the result is worth to the buyer. Then calculate your fully loaded cost to produce it, including contractor pay, compute and your own team's time. Aim for a gross margin you would be proud to show an investor.

Step 4: Run three to five paid pilots. Track hours per delivery, rework rate and time to completion. These are your real unit economics.

Step 5: Automate the most repeated step first. Every pilot should make the next one cheaper. If it does not, stop and figure out why.

Step 6: Decide with data. After the pilots, compare margin, sales cycle and expansion against your tool customers. Keep whichever wins, or run both if they serve different buyers.

Quality becomes your problem. When you sell a tool, a bad result is partly the customer's fault. When you sell the output, it is all yours. Build review and quality checks before you scale.

Concentration. Demand for AI training data is being driven by a small number of very large labs. If three customers make up most of your revenue, one budget change can hurt badly.

Headcount creep. Human experts are part of the product in this model. Without constant automation, your team and costs grow in step with revenue.

Valuation expectations. Investors may value outcome revenue differently from subscription revenue, especially if a large share goes to contractors. Be ready to explain why your margins will improve.

What is data-as-a-service? It is Snorkel's name for delivering completed training datasets and simulated environments to customers, instead of selling software that customers use to build those datasets themselves.

How fast did Snorkel grow after the change? Snorkel says its annualized revenue run rate reached $375 million, an eighteenfold increase over the previous 12 months. It raised a $350 million Series E at a $3.5 billion valuation in September 2026.

Is selling outcomes the same as running a services business? Not if software does most of the work. The difference shows up in gross margin. Ema, for example, reports margins close to 80 percent while charging for completed tasks.

What is the difference between gross and net revenue here? Gross revenue is everything the customer pays. Net revenue subtracts what you pass through to contractors. Some AI data companies pay 60 to 70 percent of gross revenue to experts, so their net revenue is far smaller than their headline number.

Can an early-stage startup make this switch? Yes, and it is often easier early because you have fewer customers to migrate. Start with a small pilot group and measure contribution margin on every delivery.

1. Snorkel AI triples valuation to $3.5B as demand for AI training data booms, TechCrunch, September 22, 2026
2. Snorkel AI scores $35M Series B to automate data labeling in machine learning, TechCrunch, April 7, 2021
3. Ema raises $77M as AI starts eating into enterprise software and services, TechCrunch, September 23, 2026
4. 17 Must-Know AI Startup News Stories (Sept 27, 2026), GTStudios