When does a first customer require manual service rather than more product?

Manual service is often the fastest way to discover the product. It is also the easiest way to accidentally build an agency and call it software.

The right question is not “Can this step be automated?” It is “What do we still need to learn before automating it is safe or useful?”

The YC advice is more specific than “do things that don't scale”

YC's essential startup advice says early founders should get the first customer by any means necessary, including manual work that would not survive at scale. The point is not romantic founder suffering. The point is proximity.

Airbnb's early team manually photographed listings. That work improved the listing, but it also put the founders in direct contact with the supply side of the marketplace. They learned what made a listing useful while they were doing the work.

Max Kolysh makes the sequencing clearer in How to Get Your First 10 Customers. Customers one to three are a personal, high-touch phase. Customers four to ten are still a manual learning phase. The next step is not “buy more software.” It is to wait until the pitch, case studies, and value proposition are repeatable enough to justify a system.

That distinction matters for agent products. A human may be filling the gap because the agent is missing a capability. Or the human may be supplying judgment the product should never pretend to automate.

Three ways to handle the missing step

Choice	Fits when	The trap
Do it yourself	You still do not understand the workflow, the exception cases, or what the customer values	You become the invisible operations team and learn nothing because the work is not recorded
Hire an operator	The work is understood, repeats across customers, and a human handoff is part of the intended service	You hide product gaps behind labor before knowing whether the economics work
Build the step	The same boundary appears repeatedly and the output can be checked	You encode one customer's strange process and mistake it for a product requirement

The first option is usually the right one when the question is still unclear. Manual work gives you the missing context. But manual work only compounds if you keep notes about the decisions, exceptions, inputs, corrections, and time spent.

For an agent extracting fields from messy documents, that might mean the founder reviews every output for the first few customers. The useful artifact is not only the corrected document. It is the record of why the correction happened: missing source, bad mapping, customer preference, or a workflow the product should not support yet.

That record turns service into product learning.

When manual service is the wrong answer

There are three warning signs.

First, the customer is paying for a result but the founder cannot explain which part of the work creates the result. That is not concierge onboarding. It is outsourced mystery.

Second, every customer requires a different process and none of the differences are being classified. You may have demand, but you do not yet have a repeatable product boundary.

Third, the manual step is expensive and invisible in the price. The company can look like software on the revenue line while behaving like a services business in the cost base.

Patrick Collison's YC conversation about the lean startup playbook gives the useful counterweight: launch quickly, but adapt when failure is expensive. Payments are not a domain where you casually ship a broken first impression and hope the market teaches you. You keep a live feedback loop, but you control the blast radius.

That is the fit rule I would use.

Use manual service when it reduces uncertainty. Use a human operator when the service is part of the intended offer and the economics are visible. Build the step when the same boundary keeps appearing and you can test whether the automation preserves the customer outcome.

Short version: do the work manually until you know what the work is. Then decide whether the product, a person, or a hybrid should own it.