When a founder reaches out, the easy answer is yes. They have budget, they have urgency, and our calendar has slots. The problem is that some of those projects don't work, no matter who builds them. We have learned, slowly and at our own expense, to spot a few of those upfront.
Here are three projects we turned down this year, and what we recommended instead.
The seed-stage founder who wanted a custom ML model
A founder of an early-stage productivity tool came in wanting "an ML model that learns each user's writing style." They had 200 weekly active users and roughly 6,000 documents in total, most of which were one-line meeting notes.
We looked at the data and told them: there isn't enough signal here to train anything useful. A model fit on 30 documents per user would either memorize or do nothing.
What we suggested: ship a keyword-based template system, instrument which templates each user picks, and revisit the ML question at 10,000 active users. Six months later they came back to discuss a different problem. The template system was doing fine. They thanked us for not letting them spend $80,000 on a model that would have embarrassed them at launch.
The Series A wanting to build their own auth
A B2B SaaS company at Series A asked us to build "a serious enterprise auth system" because their CISO had concerns about Clerk's data residency. We spent half an hour with their CISO. The concern was real but solvable: Clerk supports EU data residency on their Business plan.
We told them: do not build auth in-house. We have built auth in-house for two clients in the past five years. Both times, two years later, they paid us to rip it out and migrate to a hosted provider. The maintenance cost is higher than people think, especially around SSO, SCIM, and the occasional CVE in a JWT library.
They upgraded their Clerk plan and added a contract clause about data residency. Total time: one afternoon, including legal review. Engineering bandwidth saved: a full quarter of one senior engineer.
The enterprise CIO wanting a Salesforce replacement
A regional insurance company's CIO wanted a custom CRM to replace Salesforce. Their reasoning was reasonable: Salesforce was costing $400,000 a year in seats, and they used about 12% of the features.
We scoped the rebuild. It would have taken 14 to 18 months, cost north of $1.2M, and at the end they would own a smaller, less-supported version of what they already had. They needed something live in 4 months, before their next fiscal year.
What we recommended: keep Salesforce for the existing seats, build a small Attio-based workspace for the new use case (claims triage), and revisit consolidation in 18 months when the migration window opens. They went with that. The Attio workspace shipped in 9 weeks. The Salesforce bill is up for renegotiation next quarter, this time with a credible threat behind it.
Why this matters
Three pieces of work we didn't take, that we couldn't have taken honestly. The aggregate revenue we declined was somewhere around $1.5M. We are not bragging about that number. We are pointing at it because the alternative is worse: shipping projects that don't deliver, watching a client miss a runway window, becoming the reason the next round was harder to raise.
If your project is shaped wrong, we will tell you. If we can't help, we will usually know somebody who can. That is what we mean when we say we work with funded startups: we work in the interest of the company, not the contract.