I Spent 14 Hours a Week Routing Inbound Leads. Here's How I Got It Down to 4 Minutes

I run RevOps at a mid-market B2B SaaS company. Last quarter I sat down and counted exactly how much of my week was being consumed by one task: taking inbound leads from our website forms, enriching them with firmographic data, scoring them, and routing them to the right account executive. The number was 14 hours and 20 minutes a week. That's 36% of a full work week. Doing data entry. By hand. In 2026.

This is the story of how I got that number down to 4 minutes a week (yes, four), what it actually cost to build, and the numbers I can finally show our CFO.

The Manual Workflow That Was Eating My Life

Let me describe what I was doing, step by hateful step, so the rest of this makes sense.

We get an average of 187 inbound leads per week through our marketing site — demo requests, pricing inquiries, ebook downloads with sales intent flags, and webinar registrations. Each one would land in HubSpot as a stub contact with maybe an email, a name, and a free-text "company" field. From there:

1. Open the lead, copy the email domain. ~15 seconds.
2. Paste into Clearbit, wait for enrichment. ~25 seconds.
3. Cross-check the company size against our ICP tiers. ~30 seconds. (Is this Tier 1 enterprise? Tier 2 mid-market? SMB? Disqualified?)
4. Look up the industry against our vertical specialization map. ~20 seconds. (FinServ goes to the financial AE pod. Healthcare goes to the regulated industries AE pod. Retail goes to Eric, who actually knows retail.)
5. Check LinkedIn for the lead's seniority. ~45 seconds. (VP+ gets fast-lane treatment. Manager-level goes through nurture first.)
6. Look at the round-robin tracker. ~20 seconds. (Who's next in the rotation for this segment?)
7. Update the HubSpot record with the enriched fields and the assigned owner. ~40 seconds.
8. Post in Slack so the AE sees it. ~25 seconds.
9. Log it in our internal scoring spreadsheet for the QBR. ~30 seconds.

Total per lead: about 4 minutes 30 seconds when nothing went wrong. Things went wrong constantly. Clearbit timed out. LinkedIn rate-limited me. The round-robin tracker was a Google Sheet that someone else was editing. The Slack auto-mention picked the wrong AE because the Slack user IDs in our routing table were stale.

187 leads × 4.6 minutes = 14.3 hours a week. Roughly $1,100 of my time, every week, on a task that a 22-year-old with a laptop could do, badly.

The Numbers I Could Already See Bleeding

The time cost was bad. The opportunity cost was worse.

I went back and pulled six months of data on lead response time vs. close rate. The numbers were brutal:

The InsideSales / Harvard Business Review data has said this for years — under 5 minutes is the magic threshold — but seeing my own pipeline data hit me harder than any third-party study ever did. My average response time was 6 hours and 12 minutes. Most of the high-intent leads that came in on a Friday afternoon didn't get touched until Monday morning. The Friday-to-Monday cohort had a close rate of 1.8%.

Math on the leakage: if I could move the median response from 6 hours to 5 minutes, the same 187 weekly leads would close at roughly 18% instead of the blended 6.4% they were actually closing at. That's an extra ~22 closed deals per quarter at our average deal size, or about $480,000 in net-new ARR per quarter I was leaving on the table by routing leads by hand.

I showed that number to my CRO. We had a budget for a fix the next day.

What I Actually Built (And What It Cost)

I evaluated three categories of solution.

Category 1: A traditional iPaaS (we already pay for one). The vendor wanted a six-figure professional services engagement plus 11 weeks to ship something the analyst who built it would have to babysit. The scoring logic would live in a visual flow editor that nobody on my team could read.

Category 2: A point solution from a lead-routing vendor. The pricing was per-lead, the scoring was a black box, and the integration with our internal vertical-specialization logic required a "custom development engagement." I asked how it would handle the Slack stale-user-ID problem I described above and got back marketing copy.

Category 3: An AI workflow on a general-purpose AI platform. I went with Swfte because the workflow primitive was already designed to compose calls to an LLM, a CRM connector, a Slack connector, and a database lookup into a single durable run. The setup, end to end, took me 6 working days with a part-time engineer for two of them. (Disclosure: I work in RevOps, not engineering. I'd never deployed a workflow before this project.)

Here's the shape of the workflow we ended up with:

1. Webhook trigger from HubSpot form submission. Fires within ~200ms.
2. Enrichment step that calls a Clearbit-style firmographic connector from the marketplace, plus a LinkedIn enrichment provider for seniority and tenure.
3. Scoring step that hands the enriched payload to an LLM with a system prompt encoding our exact ICP rubric. It returns a JSON object with tier, vertical, seniority_band, score, and a one-sentence rationale. The rationale matters — we'll come back to that.
4. Routing decision step that joins the LLM output against a small Postgres table of "who currently owns which segment" (this used to be the messy Google Sheet) and picks the next AE in the round-robin for that segment.
5. Write-back to HubSpot with the enriched fields, the score, the rationale, and the assigned owner.
6. Slack DM to the assigned AE plus a channel post to the segment-specific room. The Slack user lookup goes through our SCIM-synced directory, so the stale-ID problem is gone.

End-to-end median latency: 3.8 seconds from form submit to AE pinged in Slack. The previous median was, again, six hours.

The Numbers, Six Weeks In

We turned the workflow on for 50% of inbound traffic for the first two weeks (A/B against my hand-routed control group), then went to 100% in week three. By week six the numbers were stable enough to share.

The misrouting number surprised me most. I assumed I was better at routing than an AI would be, because I knew the AEs personally. I was wrong by an order of magnitude. The LLM, given the rubric, applied it consistently. I applied it consistently on Monday morning and inconsistently by Friday at 4pm, when "let me just put this one in Eric's queue because it's almost a fit" became the operating mode.

Total platform cost across the six weeks: about $340 in compute and connector API calls, plus the one-time engineering effort. Annualized, the workflow runs for roughly $3,000/year. Against the $480K/quarter pipeline lift, the ROI math is, frankly, embarrassing in how lopsided it is.

The Three Things That Actually Made It Work

If you take nothing else from this post, take these three.

1. The rationale field. I made the LLM output a one-sentence reason for every tier and segment decision. This sounded like a small thing. It was the single most important design choice. When an AE pushes back on a routing decision, the rationale is right there in HubSpot — "Mid-market FinServ; CFO title at a 220-person fintech; high intent (pricing page + demo form within 4 minutes)." Arguments stopped. Trust in the workflow built up in about two weeks.

2. Named queries for the routing table. I did not let the LLM write SQL. The Postgres adapter we used exposes a small set of named, parameterized queries: get_active_owners_for_segment, get_next_in_round_robin, record_assignment. The LLM picks the segment; the database lookup is deterministic. This is the difference between an AI workflow that survives an audit and one that takes down your CRM. For the broader pattern, see how Swfte connects AI to any service.

3. Human-veto on a tiny slice. Anything the LLM scores above 85 and tags as Tier 1 enterprise gets a 2-minute human review window — a Slack message to me with Approve/Override buttons — before it routes. That's about 6 leads a week. It catches the rare miss (a self-described "VP" who's actually a VP at a 4-person consultancy) and gives me air cover with the AE team. For the pattern name and tradeoffs, see workflow orchestration patterns.

What I'd Do Differently

Two regrets, both about scope.

I should have started smaller. I tried to ship all six steps on day one. The enrichment and scoring steps were straightforward; the Slack integration ate two days because I tried to also rewrite our notification format at the same time. Ship the workflow with ugly Slack messages first. Pretty them up in week two.

I should have brought the AEs in earlier. I demoed it as a finished product. A few of them, justifiably, treated it as a black box being imposed on them. The fix was easy — I added the rationale field, gave them an Override button, and held a 30-minute walk-through — but I'd skip the political pothole next time.

The Conversation I Couldn't Have Before

The week the workflow stabilized, I had a one-on-one with my CRO. The agenda used to be "how are we going to keep up with inbound." It's now "we have 14 hours a week back; what do we use it for?" The answer turned out to be account research for the top of the pipeline — work that actually requires a human and was being starved of attention because I was doing data entry.

If you're routing leads by hand in 2026, you are leaking money in a way that is, today, embarrassingly cheap to fix. Pull six months of your own response-time-vs-close-rate data before you do anything else. The number will pay for the project.

Related reads if you want to keep going:

• How to connect AI to any service with Swfte — the connector tiers I used for HubSpot, Slack, and the Postgres routing table.
• AI sales automation and CRM — broader patterns beyond just inbound routing.
• AI lead generation workflows — the outbound side of the same machine.
• 6 AI workflow orchestration patterns — the human-veto and saga patterns I used here.