Say yes to a custom feature when it resolves a verified customer problem, fits the product you intend to build and has an explicit delivery and maintenance boundary. Say no, or offer a smaller experiment, when the request is mainly an untested condition attached to a possible sale. The decision should account for the future product you will operate, not just the next invoice.
Translate the request into a job
“We need a custom dashboard” describes a proposed solution. Ask who will use it, which decision they cannot make today, how often that decision occurs and what they currently do instead. A buyer’s answer may reveal a missing data export, unclear permissions or a reporting habit that software should not reproduce exactly.
Ask for a concrete example using information the customer is allowed to share. Walk through the task together and write the acceptance condition in ordinary language. If nobody can say what successful use would look like, a development estimate is premature. You need a clearer problem before negotiating a delivery date.
Paul Graham’s 2013 essay on doing things that do not scale argues for hands-on early customer work. That supports learning directly from users; it does not settle which bespoke commitments belong in your product. The decision framework below is my own practical interpretation, not a rule taken from the essay.
Choose between learning work and a product commitment
A temporary manual report can test whether the information changes the customer’s decision. A permanent dashboard adds a continuing product surface. Name which of those you are proposing. The customer should understand whether they are receiving a time-limited experiment, a paid implementation service or a supported feature.
Define the boundary in writing: users, data, frequency, acceptance test, owner and end or review date. Avoid vague assurances such as “fully flexible reporting.” A narrow commitment is easier to assess and price. It also lets the team learn without implying that every future variation is included.
For fintech workflows, include the responsible decision-maker and permitted actions in that boundary. A request to display information is different from permission to change a financial record or initiate an action. Evaluate those consequences with the relevant specialists rather than treating them as extra interface fields.
Estimate the work after the first delivery
List discovery, development, testing, customer setup, documentation and ongoing support. Then ask what happens when the underlying product changes. A feature that needs a separate fix for one account every release creates a different obligation from a reusable capability with a clear configuration.
Use a deliberately hypothetical example. A customer offers 3,000 units for a reporting feature. The team estimates 40 hours of initial work and two hours of monthly maintenance for a year: 64 hours before considering other costs. At an assumed internal cost of 50 units an hour, that is 3,200 units. These are illustrative assumptions, not market prices or AnyShift figures.
The calculation does not automatically mean refusal. The work might buy important learning or support a strategic product direction. It does mean that the founder should identify that purpose explicitly. Do not present a learning investment as a profitable implementation simply because cash arrives first.
Test whether the need repeats
Describe the underlying capability without the customer’s name. Could another account in your intended segment use it without a separate version? Ask existing or prospective users about the same job, while keeping the first customer’s information confidential. Similar wording is less useful than evidence that the workflow and acceptance conditions repeat.
If the need is unique but commercially attractive, treat it as a bounded service decision. Price and staff it accordingly, with a clear effect on the roadmap. A services engagement can be sensible; an invisible services obligation inside a standard subscription is harder to manage.
Give the buyer a concrete answer
A useful answer states what you can deliver, what remains outside scope, the proposed evidence of success and the review point. If the answer is no, explain the product boundary and offer an existing workflow or limited experiment only when it is genuinely suitable. Do not promise an unapproved roadmap item to end an uncomfortable conversation.
Review the result against actual use, delivery effort and repeat demand. The pilot rollout scorecard helps structure that review; the AI action-boundary note addresses consequential automation. For Benelux founders and Pakistani B2B teams serving international markets, the investor question is practical: what did this commitment teach you, and did it strengthen the product? Share the customer evidence and next milestone if that is the business you are building.