Designing intake that survives past manual triage

Every intake process works at ten submissions a week. The interesting question is what breaks at two hundred — and the answer is almost never the form.

What breaks is the invisible human step between "a response arrived" and "the right person is working on it." At low volume that step is free: someone reads everything, and reading everything is fast. At higher volume it becomes the bottleneck, and the failure mode is specific — it isn't that things get done slowly, it's that things get done out of order. The urgent request sits behind forty routine ones because nothing distinguished them at arrival.

Three questions worth answering at capture time

Most triage work is spent recovering information that could have been established the moment the submission landed. Three questions carry most of the weight.

What kind of request is this? Classification determines almost everything downstream — who handles it, how fast, what's needed to close it. If it's done manually, it's done inconsistently, because two people reading the same free-text description will file it two different ways. Constrain the choice at capture where you can, and classify the rest automatically against defined categories rather than leaving it to whoever opens the queue first.

Who owns it? Unowned work is work that waits. Routing doesn't need to be clever to be valuable: a rule that assigns by category, region, or account is enough to eliminate the "someone should look at this" gap. What matters is that ownership is a property of the record from the start, not a decision made in a standup.

What does the handler need that the submitter didn't give? Account tier, prior history, whether this is a repeat of something closed last week. This is enrichment, and it's pure overhead when done by hand — someone opens another system, looks something up, and pastes it into a comment. It's exactly the kind of work worth handing to an agent, because it's mechanical, verifiable, and boring.

Answer those three at arrival and the queue arrives pre-sorted. Triage stops being a task and becomes a review.

Conditional capture beats a long form

The instinct when intake is under-specified is to add fields. It works for about two releases, then completion rates fall and people start typing "n/a" into everything, which is worse than no data because it looks like data.

Conditional logic is the better lever: ask a short set of questions, and let the answers open the branches that are actually relevant. The submitter experiences a form scoped to their situation. The handler receives a submission where the fields present are the fields that mattered. Same completeness, a fraction of the abandonment.

Structure the state, not just the answers

Automated routing only helps if there's somewhere for its output to live. If your platform stores answers and nothing else, a classification step has nowhere to write its result — so it writes to a Slack message, and the sorting you just did is invisible to the queue.

Intake that scales needs the operational fields — state, owner, priority, classification, decision — modelled alongside the response, and writable by every surface: the interface, the API, and any agent you connect. That's what makes partial automation safe. The agent handles the mechanical pass, its output lands on the record, and a person sees exactly what was decided and can override it in one click.

Instrument the drop-off, not just the submissions

The submissions you receive are a biased sample: they're the ones that made it. The ones that didn't are usually more informative.

Funnel data — where people stop, which field they abandon on, how long the form takes — tells you whether your intake is actually capturing the population you think it is. A required field that half your submitters can't answer doesn't show up as a problem in your response data. It shows up as silence, and silence is easy to mistake for low demand.

The sequence that works

If you're rebuilding intake, do it in this order:

1. Shorten the form with conditional branches, so completion is high and the fields you get are real.
2. Define the operational schema — the states, owners, and categories your team already uses informally — and make it structured.
3. Automate the mechanical pass: classify, route, enrich. Keep the output on the record where humans can see and correct it.
4. Watch the funnel and cut whatever is costing you submissions without earning its place.

None of these steps requires trusting automation with judgment. That's the point. The goal isn't an intake process that runs itself — it's one where the human effort goes into the decisions that need a human, and everything before that decision is already done when they arrive.