Signal-Based Prospecting: How AI Research Agents Are Killing Spray and Pray

The old cold email playbook is dead. Not dying, not under pressure. Dead. If you are still running a "200 emails a day, generic opener, five-step sequence" outbound strategy in 2026, you are not generating leads. You are burning your domain and annoying people who will never buy from you.

The shift has been building for a few years, but it accelerated sharply in the past twelve months as AI research agents became genuinely capable. The operators who adapted early are booking more calls on lower volume. The ones who didn't are chasing deliverability fixes on domains they've already poisoned.

Signal-based prospecting is the replacement. It is not a tweak to the old model. It is a different philosophy about what cold outreach is actually for. This piece explains what it is, why it works, and how to build a basic signal-based workflow using Clay's Claygent feature, even if you're running a two-person operation.

Why Spray-and-Pray Is Now a Deliverability Liability

First, let's be precise about what "spray and pray" actually means, because the people still running it often don't think of it that way. It means: selecting a large list based on basic filters (job title, company size, industry), using a generic or lightly personalised opener, and relying on volume to generate responses. It was never a great strategy. In 2026, it has specific, measurable costs.

Spam filters have become significantly more sophisticated. Google and Microsoft's filtering algorithms now assess content quality and relevance signals, not just technical factors like SPF and DKIM alignment. An email that looks templated, lacks genuine context, and matches patterns associated with mass outreach will land in spam regardless of your warm-up infrastructure. The old deliverability playbook (warm up the domain, rotate inboxes, stay under 50 sends per day) is necessary but no longer sufficient.

Reply rates on spray-and-pray campaigns have compressed dramatically. Where a well-configured cold email campaign might have generated a 3-4% reply rate three years ago, the same approach today is producing 0.5-1% for most operators. The math stops working. At 200 sends per day across five days per week, a 1% reply rate gives you ten replies per week. How many of those are interested? How many booked calls does that actually produce?

More importantly: every low-quality send costs you something beyond the immediate non-response. It trains the filters. It generates spam reports from prospects who are increasingly willing to hit that button. And it builds a negative signal on your domain that compounds over time.

What Signal-Based Prospecting Actually Means

A signal is a specific, recent, observable event that tells you something about a prospect's current situation. Not a demographic fact ("you're a mortgage broker in Manchester"). Not a generic observation ("I noticed you're growing"). A signal is something that happened: they posted about a specific challenge, they hired for a role that implies a particular problem, they were mentioned in an article about a topic adjacent to what you do, they just launched something new.

The power of a signal is threefold. First, it gives you a credible reason to reach out right now, which dramatically reduces the "why are you emailing me" friction. Second, it demonstrates that you've actually looked at their situation, which increases the likelihood they'll give you 30 seconds of attention. Third, it allows you to make a relevant connection between their current situation and the specific outcome you can help them achieve.

Contrast two openers:

Generic: "Hi James, I came across your profile and noticed you're the MD at Apex Mortgages. I wanted to reach out because we help mortgage brokers generate more leads."

Signal-based: "Hi James, saw your post on LinkedIn last week about struggling to fill your pipeline after the base rate announcement. We've been helping brokers in exactly that position over the past few months. Happy to share what's been working."

The second opener is doing three things simultaneously: acknowledging a specific situation, demonstrating that you're not a robot blasting a list, and making a relevant, timely offer. The prospect has to make a conscious decision to ignore it rather than just pattern-matching it to spam and deleting it.

Claygent: The AI Research Agent That Makes This Scalable

The obvious problem with signal-based prospecting is that it seems to require manual research for every prospect. And it does, if you do it manually. At any real volume, that is not sustainable.

This is the problem Claygent solves. Claygent is Clay's AI research agent: you give it an instruction, it browses the web for each row in your table, and returns a specific finding. Not a scraped data field. A synthesised insight. The difference matters enormously.

Here is what a basic Claygent instruction looks like for a mortgage broker outbound campaign:

"Research this mortgage broker's company. Find the most recent relevant signal: a LinkedIn post from the past 30 days, a news mention, a recent hire, or a new service announcement. If you find one, summarise it in one sentence. If you cannot find a specific recent signal, write NONE."

Claygent runs this for every row in the table. For a list of 500 prospects, it will find a genuine signal for roughly 200-300 of them. The ones where it returns NONE get filtered out of this campaign cycle. You are now sending to 200-300 people you have something specific to say to, rather than 500 people you're hoping will respond to a generic pitch.

That filtering alone changes the economics of outbound. You are sending 40-60% of the volume. Your reply rate on the filtered list will be meaningfully higher. Your deliverability is better because a higher proportion of recipients are engaging rather than deleting. And the people who do reply are better qualified because they responded to a specific, relevant message.

Building the Workflow: From Apollo to Inbox in Four Steps

Here is the end-to-end workflow we run at Levity for signal-based prospecting campaigns. This is a real workflow, not a theoretical one.

Step 1: Pull the list from Apollo. Define your ICP filters carefully. For mortgage brokers: UK-based, 5-30 employees, owner or MD title, independent (not part of a major network). Export the list as a CSV. Aim for 400-600 contacts per campaign batch. The quality of your ICP definition matters more than the size of the list.

Step 2: Import to Clay and run Claygent. Create a new Clay table, import your CSV, and verify the data (email addresses, company names, LinkedIn URLs where available). Add a Claygent column with the research instruction above. Let it run. This takes time: for 500 rows, expect 30-60 minutes depending on Clay server load. Filter the table to remove any rows where Claygent returned NONE or found only generic information. You want specific, recent, relevant signals only.

Step 3: Generate personalised first lines with Claude. Add a Clay column that calls the Claude API (or uses Clay's built-in Claude integration). The prompt: "Using this signal: [Claygent output], write a one-sentence cold email opener for a message to [First Name] at [Company]. The opener should acknowledge the signal naturally and bridge to lead generation for mortgage brokers. Maximum 25 words. No em dashes. Direct, conversational tone."

Review a sample of the generated first lines. You will need to tweak the prompt a few times to get consistent quality. This is normal. Once the output is solid, the rest of the process is automated.

Step 4: Push to Instantly and send. Export from Clay (or use the native Clay-to-Instantly integration) and load into your campaign. The email template uses the dynamic first line as the opener, then transitions to the core message: the problem you solve, a credibility signal (specific results, client type), and a low-friction call to action. Three emails maximum in the sequence, spaced 3-5 days apart.

What the Results Actually Look Like

Numbers I'm comfortable sharing from campaigns we've run over the past few months, without identifying specific clients:

• Reply rates on signal-based campaigns: 4-7% (versus 1-2% on generic campaigns from the same ICP)
• Positive reply rate (interested, not "remove me"): 2-4% of total sends
• Calls booked per 500 sends: 8-15 (versus 2-4 on generic)
• Spam complaints: significantly lower, consistently below 0.1%

The comparison that matters most: on 300 signal-filtered sends, we are booking more calls than we were on 700 generic sends. Lower volume, better results, healthier domains. The upfront cost is higher (Clay credits, Claygent run time, prompt refinement). The return is considerably better.

There is also a compounding benefit that is harder to quantify: the conversations you have when someone responds to a relevant, specific message are better quality than the ones where someone responds to a generic pitch. They came in with context. They know you actually looked at their situation. The sales cycle is shorter.

The Objections (And Why They Don't Hold)

"This takes too long to set up." The initial Clay workflow takes 2-3 hours to configure correctly, including prompt refinement. After that, running a new campaign batch takes about 30 minutes. The setup cost is a one-time investment.

"Clay is too expensive." Clay's credit model means you only pay for enrichment you actually use. For a 500-person prospecting batch running Claygent plus one Claude integration, you are looking at roughly £30-50 in Clay credits. Against the cost of a single new client, that math is not serious. The question is not whether Clay is affordable: it is whether your current outbound approach is generating enough return to bother optimising.

"Claygent doesn't always find useful signals." Correct. That is the point. If there is no useful signal, the prospect does not get an email in this cycle. They go into a slower, longer-term nurture sequence or get picked up in the next quarterly batch when something may have changed. Sending without a signal is the old model. The whole point is to stop doing that.

Signal-based prospecting is not magic. It requires upfront thinking about what a relevant signal looks like for your specific ICP, and it requires prompt work to get the personalisation quality where it needs to be. But once the workflow is running, it compounds. Every campaign teaches you more about what signals convert, which refines the next batch.

The operators still doing spray-and-pray in 2026 are going to keep seeing response rates decline. The ones who have rebuilt around signals are already running a fundamentally different business. The gap between those two groups is only going to widen.