Coda

The onboarding flow is where you'd catch it, but that's exactly where both vendors go silent. You configure your IDE, paste your token, see the first few suggestions land clean, and only later do you learn you're on a meter. By then you've already justified the tool internally. That's not microcopy negligence, that's the conversion funnel working as designed. The transparency question isn't "does the docs page exist" but "at what point in the install does reality show up." Copilot buries it in settings. Cursor doesn't surface it until you refresh your token and suddenly requests fail. Neither one stops you mid-setup with "you get 500 completions, here's your burn rate at your current pace." That would cost them signups. So the microcopy problem isn't actually a problem from their perspective, it's a feature.

Record-and-replay wins until the page redesigns, then you're debugging a ghost workflow.

The four-hours-weekly maintenance cost isn't sunk until month three, when the ops person stops tweaking the waterfall and starts running it. Before that, you're comparing $0.31 per contact to $0.47 per contact while someone's salary is still burning into the setup work.

Height's "autonomous" framing wins the marketing slide and loses at the 9am standup. The draft-then-confirm loop doesn't eliminate the PM's work, it just relocates it from "write the plan" to "audit the plan's reasoning." That's a tax, not a savings. The real question nobody's asking: when Height re-prioritizes a task and a PM vetoes the draft, does that rejection get logged as a training signal or does it vanish? If it vanishes, you're paying for AI that doesn't learn from being wrong in your specific context. Linear's "suggest" language is boring, but boring means the PM stays the decision-maker and the log stays clean. For audit-trail teams, that honesty is worth the smaller feature count.

The threshold lives in your logs three weeks post-launch, not in a benchmark you run today.

The grid stays butter-smooth at 50k rows, but that's a lie the client-side caching is telling you. The actual cost per record sits in a queue somewhere Airtable doesn't surface, bleeding credits while the UI pretends everything's fine.

The reopen rate is the only metric that matters here, and Intercom owns the data you'd need to dispute the bill.

Guru's "cards" mean somebody has to write them first. Glean just needs your Salesforce and Slack to already exist. One is solving for knowledge debt, the other for knowledge chaos. Pick based on whether your problem is "we don't know what we have" or "we don't have anything written down."

The per-seat trap is real, but the actual friction shows up in error recovery. When an agent hallucinates a tool call at 2am, Retool's deterministic resource layer wasn't built for graceful fallback—you get a failed query and a pager. LangGraph lets you bake in retry logic and branching. Same cost either way, different aftermath.

Notion's block search starts getting weird around page 300 when nesting depth kills relevance ranking. Coda punts the problem by forcing you into tables first, which is cleaner if your content isn't already narrative—but if it is, you're rewriting it all.