Home / About / How we decide
How we decide · the frameworks behind the shipping

Most product decisions look like luck from the outside. Here are ours, before the outcome was known.

A page about the actual frameworks we use — build vs buy, hire vs contract, ship now vs wait, customer-driven vs founder-driven — and three real decisions we made in the last six months, with the reasoning we wrote down at the time. Two of them turned out right. One is still being argued about. That's the point.

Written: May 2026 · Updated whenever we make a hard call: next due Q3 2026 · 9 min read ~ 1,950 words
01 · the principle

Slow on entry. Fast on reversal.

Every decision we make is rated on two axes that have nothing to do with cost or upside: how reversible is it, and how much information do we have right now. If a decision is hard to reverse, we slow down before making it. If it turns out wrong, we move fast to undo it. The first half is easy to write down. The second half is the part most teams get wrong.

The reason most companies don't reverse decisions even when they should is that the decision becomes someone's identity. The person who pushed for it has to lose face to undo it. So we made a small structural choice early: nobody owns a decision after the meeting ends. Whoever championed it doesn't carry the social cost of unwinding it. We unwind it because the evidence changed, not because someone was wrong.

If the new evidence would have changed our call back then, the call changes now. We don't owe a past version of ourselves loyalty.

This is also why we write decision notes — not for documentation, but for ourselves. When we look at a six-month-old note and the reasoning still holds, we leave the decision. When the reasoning doesn't hold, we change it. The note is the evidence trail; the meeting is just where the trail started.

One implementation detail: we don't have a formal RFC process. The "decision note" is usually a Slack thread saved to a Notion page, or a paragraph in the weekly memo. Lightweight on purpose — heavyweight processes are how four-person teams pretend to be twenty-person teams.
— · —
02 · framework

Build vs buy.

The default in most startups is "build it ourselves, it'll be faster." This is almost always wrong in the first eighteen months and almost always right after. The question isn't speed; the question is whether the thing you're building is on your critical path to a moat.

We buy when…
vs
We build when…
Buy signal
  • It's commodity infrastructure. Email send, OAuth, observability — table stakes, not differentiators.
  • A vendor exists with three years of incident history and we'd recreate the same incidents.
  • Building it would consume an engineer for > 6 weeks and the result is at best parity.
  • The cost is a known monthly number we can predict against revenue.
Build signal
  • It touches our actual moat. Anything that handles raw signal data, model orchestration, or research generation — we own end-to-end.
  • The vendor has roadmap risk against us (will they enter our category?).
  • We need iteration speed measured in days, not quarters.
  • The behavior we want is so specific that vendor customization will cost more than the build.
The test
If you can buy it and the buy doesn't change how a customer experiences the product, buy it. If the buy means you can't fix a bug a customer reported on a Thursday and ship by Friday, build it.

We get this wrong about a third of the time. Usually we build something we should have bought, because building feels like progress and writing a check feels like admitting we couldn't. The check is almost always correct.

— · —
03 · framework

Hire vs contract.

We've kept the team deliberately small. Eleven people as of May 2026, growing to roughly fifteen by year-end. Every additional headcount is the most expensive ongoing commitment a startup makes — not because of salary, but because of the meetings, the onboarding, the reporting structure, the slow erosion of high-bandwidth communication.

We contract when…
vs
We hire when…
Contract signal
  • The work has a defined endpoint — a launch, a redesign, a migration.
  • We need a specialist for < 6 months of full-time equivalent.
  • The work is adjacent to our core but not on the daily critical path.
  • We can scope it in writing and tell when it's done by looking at the artifact.
Hire signal
  • The role creates more roles — it scales the team's output, not just adds one unit.
  • The work is open-ended and the right answer changes with context only an insider has.
  • We'd want this person in the room when we're rewriting the strategy a year from now.
  • The cost of being wrong about this hire is recoverable in < 90 days.
The test
Can you describe the deliverable in one sentence? Contract. Does the answer to "what should this person work on next?" require a conversation with the founders every Monday? Hire.

Our brand work, our website illustrations, our annual security audit, and most of our marketing video work is contracted. Engineering, product, design system, customer-facing roles, and anyone who handles the model layer are W-2.

— · —
04 · framework

Ship now vs wait.

The bias inside the building is always to ship. The bias outside the building should be a customer's right to ignore us. So the question we ask before any release isn't "is this ready?" — it's "if a customer never hears about this, are we proud of what they encounter when they bump into it?"

We ship when…
vs
We wait when…
Ship signal
  • The worst-case experience is "this feature isn't useful yet" — not "this gave me wrong data."
  • We can roll it back in < 30 minutes without affecting other customers.
  • The feedback loop from real users is faster than the feedback loop from internal testing.
  • Three customers have asked for the thing in three different words.
Wait signal
  • It writes to a customer's CRM, sends on their behalf, or surfaces a claim with a confidence number.
  • Rollback requires a migration or a customer-side change.
  • The data shown could be wrong in a way the customer can't detect — and a wrong claim costs them a deal.
  • We caught a bug in staging that we don't fully understand the root cause of yet.
The test
If the feature being wrong would result in a customer sending the wrong email to a real prospect, wait. If the feature being wrong would result in a customer saying "huh, that doesn't quite work" and trying again, ship.
— · —
05 · framework

Customer-driven vs founder-driven.

There's a recent fashion to build only what customers explicitly ask for. There's an older fashion to build only what the founders believe in regardless of feedback. Both, taken alone, kill companies. The skill is knowing which mode you're in for which question.

Customer-driven on…
vs
Founder-driven on…
Listen mode
  • Workflow ergonomics. How the product fits inside an existing day — keyboard shortcuts, CRM field mapping, sequencing rules.
  • Integration surface — who we integrate with next.
  • Error messages, empty states, the language used in the UI.
  • Pricing structure and packaging.
Conviction mode
  • What the product is for. The thesis — that signal-first outbound beats volume-first — was set before our first customer.
  • What we refuse to do (see /anti-patterns).
  • The tone and visual identity. We're not going to A/B test a font.
  • The metrics we publish on /open.
The test
If a customer asked for something and we said no, can we still explain the no in plain English a year from now? If yes, it was a founder-driven call. If we can't, we got conviction confused with stubbornness.
— · —
06 · receipts

Three real decisions. Reasoning before the outcome.

Frameworks are easy to write. Here are three actual decisions we made and what we wrote at the time. Two have outcomes now. One is too early to call — we'll update this page when we know.

01 · Outsource the data warehouse layer to a vendor instead of building on Postgres
Feb 2026 · revisited May 2026
The question
We needed a place to land hundreds of millions of signal events per day with sub-second OLAP queries. The cheapest path was extending our existing Postgres. The fast path was a managed columnar warehouse.
Framework
Build vs buy. Critical path? Yes, but for query performance, not the data itself. The signal extraction is the moat. The warehouse is plumbing.
What we picked
Buy. Picked a managed warehouse with usage-based pricing. The engineer we would have spent on warehouse tuning shipped two model improvements in the same window.
Outcome right call
Three months in, the warehouse cost is 11% of revenue and stable. The two model improvements moved our brief precision from 71% to 84%. The buy paid for itself in the engineer's freed time, before counting any infra savings.
02 · Ship the HubSpot integration before the Salesforce integration was fully bidirectional
March 2026 · revisited May 2026
The question
Salesforce one-way sync was working. Bidirectional was 60% there. Meanwhile our HubSpot waitlist had grown to 412 teams, and they were the loudest about it on every call.
Framework
Ship now vs wait + Customer-driven on integration surface. The HubSpot wait was lost revenue. The Salesforce bidirectional was a polish item — existing customers had a workaround.
What we picked
Pause Salesforce bidirectional. Ship HubSpot. We told the four Salesforce customers what we were doing and why. Three were fine with it. One was annoyed but kept their seat.
Outcome mixed
HubSpot integration converted 38 waitlist teams to paid in the first six weeks. But Salesforce bidirectional took an additional ten weeks to ship because the team had to reload context, and one of those four customers churned in May citing "the integration story." Net positive, but the cost was higher than we modeled.
03 · Hire a founding designer instead of contracting through a studio for the rebrand
April 2026 · too early to evaluate
The question
We needed a real brand system — logo, type, color, voice, the whole thing. Two paths: contract a known brand studio for a 12-week engagement, or hire someone full-time who'd own it forever.
Framework
Hire vs contract. One sentence test: "can you describe the deliverable in one sentence?" Sort of — but the answer to "what should this person work on next?" needed a conversation with the founders every week.
What we picked
Hired full-time. Founding designer started in May. The bet is that brand work compounds — every page, every email, every deck benefits from one person holding the whole system in their head.
Outcome too early
Two weeks in. They've shipped a logo refresh and started the type system. We'll know whether this was the right call by Q4 2026 when we see how much of the brand system has been internalized vs. how much we're still re-deciding. Check this page in November.
— · —
07 · the slow lane

What we won't decide quickly.

Most of this piece is about how to move faster on the right things. Just as important is the list of things we won't move fast on, period. These are the decisions where the cost of being wrong is asymmetric — the upside of speed is small, the downside of a mistake is large or unrecoverable.

Anything that changes the data customers trust. Schema migrations on signal output. Confidence-score recalibration. Model swaps that change how briefs read. These ship behind opt-in flags first, then go default after at least one full sales cycle (90 days) of customer feedback.
Pricing changes. We've changed pricing once and won't again for at least four quarters. The reasoning is in /changelog. Pricing trust compounds; pricing distrust compounds harder.
Anyone we put in front of customers. A bad CS hire is felt within a week. A great one takes months to fully ramp. Both directions argue for slowness — we'd rather under-staff support for a quarter than over-hire and rebuild.
The list on /anti-patterns. Adding or removing something from that list is a multi-week conversation involving every founder and at least one customer. The list is a public commitment, not a marketing surface.
Raising another round. We raised once. The next raise — if there is one — will be telegraphed at least a quarter in advance on /open. We won't take opportunistic capital because someone showed up with a term sheet on a Tuesday.
Selling the company. Not happening. If it ever does, the reasoning will be public before the close.

Everything not on the slow list is on the fast list. Default to motion, default to writing the reasoning down, default to changing course when the evidence changes. That's the whole game.

If you've read this far

You're either deciding whether to work with us, deciding whether to work at us, or wondering how we operate.

All three are reasonable. The fastest way to find out whether our decision-making actually shows up in the product is to use it for two weeks. The fastest way to find out whether it shows up in the people is to talk to us.

If you disagree with any of the four frameworks — or have a candidate for a fourth real decision we should be writing up — write back. We update this page whenever the call deserves it.

AM Asif M. · Co-founder · Last full revision May 2026. The decision logs that fed this page are not public, but we'll share specific reasoning over email if you're considering a similar call at your own company.