Journal · Business workflows

Pre-seed judgment before the metrics arrive.

Before operating metrics exist, investment judgment can still be disciplined: separate founder insight, feasibility and buying behavior, then choose tests that could change the view.

Conclusion

Before operating metrics exist, investment judgment can still be disciplined: separate founder insight, feasibility and buying behavior, then choose tests that could change the view.

A falsifiable pre-seed decision loop

  1. State the unusual insightIdentify a user's painful task and explain why available alternatives leave it unresolved.
  2. Separate the uncertaintiesKeep technical feasibility, buyer demand and repeatable delivery as distinct questions.
  3. Choose a discriminating testUse a realistic task, a simpler baseline and an explicit result that could weaken the thesis.
  4. Inspect evidence and failureReview the output, buyer behavior and implementation effort rather than a pitch summary.
  5. Revise the research stanceRecord Investigate, Watch or Pass with the reason the evidence supports it.
  6. Set the next milestoneName the remaining uncertainty; feed the new result back into the original insight.
Original proposed research process, illustrated in the article through public Qontext materials. No founder interview, experiment or investment has been conducted for this draft.

The absence of revenue changes the evidence required

Pre-seed judgment should assess whether a team has discovered an important problem and can resolve its critical uncertainties efficiently. My conclusion is that missing operating metrics call for more explicit reasoning about evidence, rather than either automatic rejection or unrestricted faith. A credible view identifies what is known, why it matters and which observation would overturn it.

Seedcamp's FAQ says its entry point can include companies before product and revenue. It emphasizes founders' problem understanding and tendency to act. This is one fund's stated approach, not a universal investment rule. It nevertheless illustrates why requiring established retention or a standard revenue threshold at every stage can answer the wrong question.

Keep insight, feasibility and adoption distinct

A founder may understand a painful workflow without having solved it technically. A technically effective prototype may address a task nobody wants to buy separately. A buyer may fund an experiment before anyone knows whether delivery can repeat. I would keep those uncertainties separate because evidence that supports one can leave the others untouched.

Founder-market fit should also be specific. Instead of awarding points for a familiar employer or confident vocabulary, ask what the founder knows about the user's work that a plausible competitor misses. Request an example of a rejected design, the observation behind that decision and what the team changed afterward. Those are proposed diligence questions, not claims that I have interviewed any named founder.

A public case shows the boundary of desk research

Qontext's 5 February 2026 company announcement describes a $2.7 million pre-seed round and a product intended to make company context reusable across AI applications. It names Lorenz Hieber and Nikita Kowalski as founders. The announcement also makes production and performance claims. Those claims belong to the company; this article does not independently validate them or establish its current financing stage.

The interesting hypothesis is that separately maintained context becomes a bottleneck as organizations deploy more applications. The unresolved questions are practical: who experiences the duplication, who controls the source systems and who buys a shared solution? A financing announcement supports the existence of a disclosed round. It does not settle those questions, prove founder insight or establish product-market fit.

Choose a test that distinguishes competing explanations

For a company pursuing this problem, I would propose one constrained workflow with frequently changing information, such as support staff answering questions about negotiated contract terms. Compare existing application context and basic retrieval with the proposed shared layer. Include an amendment and a permission change. Ask authorized reviewers to judge the answer and the evidence supporting it. This is a proposed test, not a reported Qontext evaluation.

If all three setups work equally well, the extra infrastructure needs another justification. If the new layer improves answers only after extensive manual curation, the question shifts to delivery cost and repeatability. If the gain survives changed facts and transfers to a second application, the cross-application thesis becomes more credible. Each result narrows an uncertainty without pretending to establish the whole company.

Learning speed requires a record of changed beliefs

Shipping quickly is observable activity, but learning requires a decision to change. I would ask for the team's prior hypothesis, the evidence collected and the resulting product choice. A team discovering that buyers prefer an embedded workflow over a horizontal platform may be learning well even as its original pitch weakens. The relevant evidence is the quality of the revision and the next test it enables.

Anthropic's evaluation guide recommends starting early with concrete tasks and checking whether graders reflect intended behavior. I take that as a practical tool for learning, not a substitute for buyer discovery. An evaluation can reveal whether amended terms are handled correctly; it cannot establish whether a department will allocate budget to the solution.

The danger is making a neat story look like evidence

The strongest objection to this approach is that almost any plausible founder can produce a compelling insight and a small favorable test. Selection bias, friendly design partners and an accommodating evaluator can make a weak business look unusually promising. Public research is especially vulnerable because successes are easier to find than abandoned implementations.

My response would be to seek a disconfirming observation deliberately: a buyer who declined, a failed task, a simpler substitute that performed well or a customer who would use the feature only if bundled. Then ask whether the explanation predicts another observable outcome. If the founder attributes every rejection to an unsophisticated customer and every failure to a temporary model limitation, the thesis has become difficult to falsify.

A useful judgment ends in a decision and a learning milestone

For this public context-infrastructure case, my provisional research stance would be Investigate: the coordination problem is coherent, while incremental performance and purchasing authority remain unresolved in the evidence reviewed here. That is a research priority, not an investment recommendation or a claim of completed diligence. The next steps would be a buyer conversation, an inspectable task comparison and a delivery-effort review.

The investment implication is to match conviction to the uncertainty a financing round can resolve. A milestone might establish that an identified buyer will deploy the product for a second application under real access controls. If simpler tools satisfy that buyer, revise the view. Pre-seed judgment is most useful when another person can see both the reason to continue and the result that would make you stop.

Source register

  1. Seedcamp: FAQs · investor-stated-criteria
  2. Qontext raises $2.7M to build the context layer for AI · company-announcement
  3. Anthropic: Demystifying evals for AI agents · primary-technical-research

← Back to journal