One of the things we get asked most often by broker partners is some version of the same question: what is the AI actually doing, exactly, in those 90 seconds between form submit and quote? The honest answer is that it is doing fewer things than the hype suggests, and doing them more reliably than a typical human-led workflow can.
This post walks through what we mean by lead qualification at Coverly, what the qualification agent decides in real time, and the tradeoffs we have accepted to keep the agent deterministic enough that a partner can run it in production.
The framing problem: most teams conflate qualification with routing
When broker operators say they want to qualify leads, they almost always mean two different things stacked on top of each other:
- Is this lead real? — does the contact info look usable, is the vehicle or life product identifiable, is the ZIP plottable on a rate map.
- Is this lead valuable? — bucket the lead into the campaign-appropriate tier (hot / warm / cold) so it lands in front of the right person at the right cadence.
The mistake we see shops make is collapsing both of these into one human-led step. The producer who picks up the phone is doing both jobs at once, with incomplete information, in the first thirty seconds of the call. That is why the average producer touches a fresh lead and decides "this one isn't going anywhere" before any signal is in front of them.
What the agent decides, in 90 seconds
Coverly's qualification agent runs a small fixed sequence of steps. It is not a general-purpose model that decides everything; it is a deterministic pipeline with one narrow model call per step, each of which produces a structured field on the contact record.
1. Identity & data hygiene
The agent validates the email shape, normalizes the phone number if present, and confirms ZIP code maps to a U.S. state. Anything that fails validation gets tagged bad_data on the contact record — we do not throw the lead away, but it does not enter the nurture pool. In practice, fewer than 4% of submissions fail this step.
2. Product line & vehicle enrichment
For auto submissions, the agent enriches the make / model / year against a vehicle lookup table and stamps year, body style, and a stability score on the record. For life submissions, it stamps health-class guess, coverage-amount band, and term length if the user specified one. None of this is a credit pull — it is just structured public data plus the form fields the user already typed.
3. Tier scoring
The model call. We pass the enriched lead, the partner's prior tier distribution from the past 90 days, and the partner's stated capacity (e.g. "this team closes 12 hot leads a week"). The model returns a tier label and a confidence. We do not return a probability — we return a tier, because the downstream routing fire-and-forgets on that label.
4. Routing & followup
Once the tier is on the record, the rest is mechanical. Hot leads page the partner's on-call producer. Warm leads drop into a 7-day nurture cadence. Cold leads land in a monthly-newsletter track. Every event lands back on the contact record so attribution reporting is reading a real-time log, not a monthly spreadsheet rebuild.
The tradeoffs we accepted
We chose this shape deliberately and three tradeoffs are worth naming out loud.
The tier labels are coarser than a probability. We did not return a 0.0–1.0 conversion-likelihood score for downstream consumers to interpret, because we saw partners making the wrong call when given one. "What does 0.62 mean?" turns into "uhh, I guess on the fence", and the lead dies in no queue. Three tiers — hot / warm / cold — is coarser, but it forces a routing decision every time.
The model is fail-closed, not fail-open. If the model errors out — rate limit, timeout, anything — we drop the lead to warm rather than rejecting it. We accept that some bad leads will land in nurture rather than risk dropping a real lead because the model had a bad day. We surface this as a failure mode in the partner dashboard so capacity gets tuned against the real distribution.
The agent's outputs are auditable, not freeform. Every decision lives as a column in the contact record. We can replay it, dump it, audit it. We cannot do that with a chat-style "what did the AI think?" agent. For an industry that is going to be increasingly regulated, the audit trail is a feature, not a tax.
What this is not
This is not the agent making the close. The producer still closes. The agent is not deciding who to call, what to script, which carrier to push, or whether the lead deserves a callback. All it is doing is turning untyped form fields into a structured record with a routing label on it — reliably, in 90 seconds, at scale.
If you are evaluating agent stacks for a broker operation, the question we would push you to ask is not "which model scores best on my historical data?" It is "which agent has the smallest, most auditable set of decisions, and which one has the cleanest failure mode when something goes wrong?" The teams winning on this are not winning with the most sophisticated models; they are winning with the most boring pipelines.