The 60,000-Contact Problem
Every B2B SaaS company eventually runs into the same uncomfortable truth: there's a graveyard of leads sitting in your CRM that nobody has the time to touch.
For PriceLabs - the dominant dynamic pricing and revenue management platform for vacation rentals, with over 600,000 properties priced and 60,000+ paying customers - that graveyard had grown to 60,000 ICP-fit contacts who'd never been activated. The audience was right. The intent signals were there. What was missing was the human bandwidth to actually talk to any of them.
Sharon Biggar, VP of Marketing at PriceLabs, faced a familiar dilemma: hire someone, wait three weeks for onboarding, hope they ramp, and burn at least three months before a single email went out - or try something different.
She picked door number two. The result is an AI agent the team named Sherlock, and in a recent webinar hosted by Nivas Ravichandran (Head of Marketing, Spendflow) and Sayanta Ghosh (CEO of Nrev), Sharon walked through exactly how it was built, what broke along the way, and the ROI it's delivering today.
Here's the full playbook.
Why "Sherlock"? Because the Research Comes First
The team named the agent Sherlock for one reason: it spends most of its energy investigating the prospect before it ever drafts a sentence.
For every contact pulled from the database, Sherlock researches 12 distinct signals. Some are universal, some are deeply specific to the vacation rental industry - and that specificity is the point.
Among the things Sherlock looks for:
- Is this company expanding the number of properties they own or manage?
- On their direct booking website, is the calendar full or are they struggling to fill nights?
- Are they already using dynamic pricing? (You can often tell by checking whether the nightly rate varies day-to-day.)
- What kind of reviews are their properties getting?
- What are they posting on LinkedIn or their corporate blog?
- Are there events, new airline routes, or new regulations affecting their region?
- Are they actually getting reservations?
Where did the list of 12 come from? Sharon was refreshingly honest: it started as a wish list. The team sat down and asked themselves, "If we were about to get on a call with this customer, what would we want to know?" They wrote it all down, kept the questions that were realistically researchable, and handed the list to the agent.
Some signals turned out to matter less than expected - LinkedIn posts, for example, are surprisingly thin in the vacation rental industry because operators are out running properties, not posting hot takes. But the wish-list approach is the lesson. You don't need to engineer a perfect signal taxonomy upfront. You need to write down what a thoughtful human would look up, then let the agent operationalize it.
The takeaway for other marketers: Your 12 signals will be different. The exercise is the same - what would a great rep look up before a call?
The Three Workflows Behind Sherlock
Under the hood, Sherlock isn't one agent. It's three workflows stitched together.
1. Sherlock the Researcher. Once a week, Sherlock pulls a defined batch of contacts from the database and goes hunting. It scrapes LinkedIn, parses direct booking sites, looks at reviews, and dumps everything into a (very large, very unglamorous) Google Sheet. Sharon laughed about this on the call: "We have these sophisticated LLMs, and behind it all is a Google Sheet." The Sheet acts as both the memory layer and the input for the next step.
2. Sherlock the Writer. Using the research, the agent generates a personalized 5-email sequence for each contact. The structure was set in the prompt:
- Email 1 — The Hook. Highly personalized, anchored in something specific the agent found.
- Email 2 — The Education. Helps the prospect understand they have a problem (i.e., they're leaving money on the table without dynamic pricing).
- Email 3 — The Customer Story. A relevant case study or blog matched to the prospect's situation. This is the email that gets the most clicks.
- Email 4 — The Nudge. A slightly more sales-leaning push - "you haven't tried this yet, and here's why it's time."
- Email 5 — The Polite Step-Back. Reminds them the free trial is always available, no pressure.
If a prospect responds at any point, the sequence stops and they're routed differently.
3. Sherlock the Approver. Every email is scored by the agent against a set of criteria. If it fails, it's regenerated. If it passes, it ships via SmartLead.
The Stack
For anyone trying to map this onto their own setup, here's the tooling:
- Nrev - builds and orchestrates the agent workflows; also handles the LinkedIn scraping and research scaffolding.
- ChatGPT - the LLM doing the actual writing.
- Google Sheets - the (humble, beloved) data store.
- SmartLead - email deliverability platform.
- An in-house email infrastructure team - managing domains, warming, and deliverability across roughly a million emails a month.
That last piece deserves emphasis. Sharon was clear: deliverability isn't something Sherlock solves. It's the foundation Sherlock sits on top of. No matter how good your copy is, if your domains aren't warm, none of it lands.
The Three Challenges Nobody Talks About
Sharon highlighted three problems the team hit that aren't covered in the typical "build an AI agent" tutorial. These are the most useful part of the talk.
Challenge 1: ChatGPT really wants to be a salesperson
Out of the box, the LLM constantly tried to push for calls, audits, "free consultations," and "independent assessments." It loved exclamation marks. Everything was "amazing" and "awesome." None of that matched PriceLabs' tone, and worse, it didn't match the product-led motion at all - PriceLabs wants people in the free trial, not on a sales call.
The fix was a long, growing list of what to avoid baked into the prompt. By the time the team exported the prompt to a Google Doc, it was 34 pages long. Things like:
- Don't offer a call.
- Don't propose an audit.
- Don't propose a consultation.
- Don't use exclamation marks like a hype machine.
- Don't write to enterprise length when concise lands harder.
The prompt became a living document, accumulating constraints every time the agent did something off-brand.
Challenge 2: A/B testing has to be reinvented for AI-generated email
Old-school A/B testing - same body, three subject lines, see which wins - falls apart when the LLM is generating fresh copy for every prospect. Every subject line is already different. There's no "control."
PriceLabs adapted by testing strategies instead of copy. Examples:
- Tone: more salesy vs. softer.
- Length: shorter vs. longer.
- Sequence assumption: problem-aware vs. problem-unaware.
You stop testing words and start testing principles. The "variant" becomes a directive in the prompt; the "control" is the existing prompt. It's a fundamentally different muscle, and most marketing teams haven't built it yet.
Challenge 3: The human-in-the-loop approval doesn't scale (and most of the value evaporates anyway)
In the early days, every email Sherlock wrote was reviewed in Slack. The team would approve or reject, and Sherlock would regenerate as needed. Then they dropped to one in two. Then one in three. Then they noticed something obvious in hindsight: humans get bored fast, and at the volumes where AI matters, approvals become theater.
It got worse when PriceLabs expanded to all six languages the product supports. Sharon could review English and one other language. The other four? "It looks like German, that's good enough" - and out it went. The check was no longer a real check.
So they did the smart thing: they read what they had been checking for, codified those checks into the prompt, and folded them into Sherlock's own scoring. Now the agent self-grades against criteria like "you didn't hallucinate the case study" and "you didn't propose a call." Approvals flipped from gating to monitoring - the agent now sends autonomously, and a sample is dropped into Slack so the team can read along, not block.
The arc to remember: approve everything → approve a sample → codify the checks into the prompt → let it run.
When Humans Reply to the Robot
A subtle but important wrinkle: the AI is on one side, but a human is on the other - and humans don't always do what your funnel diagram says they should.
Some prospects ignore the email. Some click the trial link. And some… reply. Sometimes they reply asking for a call, even though Sherlock never offered one.
PriceLabs handled this with a routing layer. Every reply is mirrored into a Slack channel (so the team can see whether Sherlock is actually generating heat) and automatically opens a support ticket. The support team takes the call, not sales - because Sharon was clear that these prospects are still too cold for a sales rep, and bad leads from marketing are how marketing earns the eternal resentment of sales.
It's a small piece of plumbing, but it solves a real problem: AI outbound creates inbound at the wrong stage of the funnel, and you need an answer for that before you turn the volume up.
How Long Did All of This Take?
Sharon's rough timeline:
- Early November: Started building.
- December: Sherlock was sending real emails (with human approval).
- January: Sherlock started sending autonomously.
Roughly 3–4 weeks to "good enough to ship," and another month or two to retire the human approval step. Nrev handled most of the technical lift - the prompt engineering, the scrapers, the workflow plumbing - while the PriceLabs team owned the strategy, tone, and content the agent draws from.
Volume now sits between 500 and 1,000 prospects per week, depending on what's being tested and what the deliverability team can support.
The ROI
After about three months in production, the headline numbers:
- ~1.1% conversion rate to free trial - strong for cold outbound, especially in a PLG context where most published benchmarks come from sales-led companies.
- A real return on investment versus the counterfactual (no human hire, three months saved, a previously dead database now working).
- Sherlock has "children." Other teams at PriceLabs have spun up sibling agents - including one that emails ahead of conferences and is allowed to be more salesy because the goal there is meetings at the booth, not free trials. Each child agent inherits the architecture and gets retuned for its job.
That last point is maybe the quietest but most important result. Once you've built one Sherlock, the marginal cost of the next one drops dramatically. The infrastructure investment compounds.
The Five Things to Steal From This Playbook
If you're a marketer staring at your own pile of dormant contacts, these are the lessons worth taking away.
1. Start with a wish list, not a model. What would a great human look up before a sales call in your industry? That's your signal list. Build for that.
2. Make the case study the centerpiece. PriceLabs found that the email featuring a matched customer story drives the most clicks. People react to "here's someone like you who solved this," not "here are our features."
3. Write the prompt like a living document. Sharon's prompt is 34 pages and growing. Every weird thing the LLM does becomes a new line item. Treat it like a runbook, not a one-shot configuration.
4. Test strategies, not subject lines. Tone, length, awareness level, CTA aggressiveness - these are your new variants. The LLM does the copy.
5. Don't bottleneck on human approval. Codify your checks into the agent's own scoring rubric. Reserve human attention for sampling and strategic course-correction, not gatekeeping.
The Bigger Idea
The most underrated point in the whole webinar wasn't about prompts or stacks or signals. It came from Sayanta at Nrev, almost in passing:
"The success is based on how custom you can get with your signals - how specific to your business and your industry."
The companies winning at AI-driven outbound aren't the ones with the fanciest models. They're the ones who can articulate, with precision, the exact things a human expert in their domain would notice - and then teach a machine to notice them at scale.
Sherlock works because PriceLabs knew what a great rep would look up. The agent is the lever. The domain knowledge is the fulcrum.
If you've got a graveyard of leads, the question to ask yourself isn't "can AI do this?" It's "do we know exactly what we'd want to know about each of these people if we had infinite time?"
If the answer is yes, you already have the hard part figured out. The agent is just the easy part.
This post is based on a webinar with Sharon Biggar (VP of Marketing, PriceLabs), hosted by Nivas Ravichandran (Spendflow) and Sayanta Ghosh (Nrev). Sherlock was built on Nrev's workflow automation platform.