Resources mentioned in this newsletter:

Outbound Pipeline Inspection

We’re all under immense pressure to implement AI in GTM, and AI SDRs are one of the primary use cases everyone is talking about.

It’s easy to see why. Almost every company says generating more pipeline is one of their top priorities and what better way to do that than implementing an AI SDR? The benefits aren’t just for outbound. If we get this right, we can convert more marketing leads, ABM accounts, and existing customers to more pipeline and revenue.

Easy peasy. Right? Unfortunately not.

While many companies have implemented AI SDRs, most have nothing to show for it yet. In SaaStr’s recent survey, only 3% of companies that had implemented an AI SDR could show any revenue from it, let alone a sustainable ROI and repeatable, scalable motion.

This is because AI SDRs are just a tool, and they’re often just one of the tools, (alongside solid strategy, process, analysis, and human execution), required to make outbound work. In this article, we’re going to lay out how to make this work, and how to continuously improve it.

Design the Engine Before You Buy Anything

Let’s start with first principles. Outbound works when you reach the right people with the right messaging the right number of times. To achieve this we need to define the “right” people, companies, messaging, and sequence of outreach, and calculate how many people we can actually work.

ICP and Personas: Define as narrowly as your data allows. These accounts are the easiest to sell to, easiest to serve, happiest, most profitable, and easiest to retain and expand.

Sequencing: Decide how many emails/calls/messages it takes, based on what has worked on YOUR buyers.

Research and Messaging: Define the research process and messaging that’s worked best to help you land the customers you have today.

Capacity: Determine how many prospects we can work with the above process so we know how many accounts and contacts we need for each rep.

The Six-Agent List-Building Stack

Now for the fun part.

Build each step as its own, separate agent, instead of one agent trying to do everything. When (not if) something breaks, you’re fixing a single agent rather than untangling a monster. Other GTM motions can also reuse some of these agents without rebuilding them.

Keep in mind, although we’re suggesting specific tools for some of these agents, but they may not be right for your company today, and may change dramatically by tomorrow.

Every agent writes back to Salesforce, so nothing lives in an AI tool’s memory.

1. Targeting Agent

Tools: Clay as the control layer, with ZoomInfo feeding it.

• Select sources based on the data your business actually needs
• Build a waterfall per field so a miss falls through to the next source
• Identify accounts by describing your ICP in plain language, not filters
• Identify contacts at those accounts who match your buyer personas

This gives you a base of high-fit ICP accounts the rest of these agents run against. 

2. Triggers and Signals Agent

Tools: ZoomInfo, 6sense, and the like.

• Define the signals: hires, funding, M&A, intent, etc.
• Select sources that push signals to you as they happen
• Build the data structure so every signal lands on the account record
• Automate it so it runs on a cadence and catches ICP accounts you missed

This adds signals to the target accounts and may identify additional ICP-fit accounts. 

3. Enrichment Agent

Tools: Narrow the list with Claude before spending on contact level data with ZoomInfo or Clay

• Determine if and how accounts from agents above need enriched
• Define the data you need, starting with revenue, industry and the basics
• Build a waterfall that runs free filters before anything that burns credits
• Integrate tools so paid credits go to direct dials and email addresses only
• Automate it against existing records, since titles and jobs change yearly

This may or may not already be covered in “Targeting” and “Triggers and Signals” Agents.

4. Scoring Agent

• Score accounts on firmographics and behavior weighed together
• Score contacts inside each account so reps know who to call first
• Automate it, feeding opens, clicks and rep flags back into the model

Prioritize which Accounts to work and enrich further with contacts.  

5. Contact Agent

Tools: Claude for web research.

• Identify the accounts that scored well, and spend contact data only there
• Build a waterfall that sends Claude to leadership pages when titles fail
• Automate it so job change signals skip ahead and land straight here

6. Segmentation Agent

• Segments are often titles by industry, but not always
• Define segments around buyers’ problems and solutions
• Enrich data to fill the gaps your segment logic depends on
• Automate it, rerunning with marketing when a new problem shifts the audience

Segmentation makes it FAR easier for humans and/or AI to research and message effectively. 

The Research and Messaging Agent Stack

We’ve defined the research process and messaging that generates the most meetings, pipeline, and revenue. Now we need to build the agents to do it.

We start, segment by segment, building an agent to source and analyze the information we need to feed our messaging. Then, we build an agent to leverage that research to make messages “relevant/personalized at scale” and passes to the rep to edit, approve, or go out automatically.

1. Research Agent

Tools: Claude for web research.

• Define the output as concise research with links to sources
• Define the process around what a rep would do: read the site, find the angle
• Select tools that can research at scale, since a rep can’t do this per email
• Automate it so research is waiting for the rep when they need it

2. Messaging Agent

Tools: Claude or your sequencer’s own AI. Outreach, Artisan, and Salesloft can all draft and send, so where the message gets written matters less than what you let it write.

• Start with templates for each segment that reflect good emails
• Have marketing or sales write the parts that don’t change (problems/solutions)
• No human or AI should reinvent solution statements for every email
• Then define the parts that do change based on the research
• Set guardrails on length and reading level, and tie research to the problem

Running Outbound

How much personalization and human touch you put into each email should track the value of the prospect.

• Tier 1: the rep writes it, working from research the agents already did
• Tier 2: AI drafts it, the rep reviews and approves before it sends
• Tier 3: AI writes and sends it

Each tier down saves time and costs response rate. If your price point justifies it, personalize everything but don’t make reps start their research from zero.

(An AI-Native Outbound Engine doesn’t have to include AI-written emails.)

Capture the Data, Then Update the CRM

Capture everything the engine produces, Outreach and Artisan sync emails to Salesforce, Otter and Gong give you call transcripts, a tool can pull LinkedIn DMs if you accept the terms of service risk, and your calendar shows meetings booked while the transcript confirms which were held.

Have another agent read that material and update the CRM: MEDDIC fields from the call, next steps on the opportunity, notes flagged from things mentioned in the call like a promotion, etc.

Decide in advance which fields it can overwrite and push the rest to Slack for rep approval. Keep the judgement calls human: no agent moves a deal to closed lost. Closed won off a signature in an e-sign platform is fine. Movement between earlier stages depends on your deal size and sales motion.

Analysis: AI, Then Human Analyst and Management

Build an AI Analyst to do the obvious heavy lifting. Then send that analysis to human analysts and management to go further with it.

Start with execution, not results. If the sequence called for 10 to 15 touches and the rep did 3, there’s nothing else to analyze. We have to get execution right before we can get insights from the data.

Then grade the messaging. We built an agent that scores every email on subject line, homework, problem, solution, CTA, word count and reading level, then ranks them best to worst in a PDF we can skim. No human can read hundreds of emails to coach a rep. (Trust me, I’ve tried.) This agent helps both me and our reps immensely.

Show reps the rubric behind the score. An unexplained number out of 100 gets dismissed, and faster once people know a machine produced it. Spell out what the scores mean.

Management drives the coaching. It has to come from the top. Few reps go hunting for this feedback on their own.

Have a human dig past the summary. Cut the results by segment, rep, team, and product alongside Claude to build the agenda for the GTM Council. Turn the analysis into insights about improvements to make.

GTM Council Meetings

Bring the insights from the data analysis to the entire revenue leadership team in a regular, focussed, meeting where you can make decisions about where and how to make improvements.

Get your GTM leaders in a room weekly or monthly. The group decides what to change, and those decisions go back into the targeting, signals, research and segmentation agents.

Where to Start

Notice how little of this revolves around an AI SDR?

The actual outreach portion of the system is just one part, with multiple processes working upstream and downstream to get the best results possible.

The tools you use matter much less than the structure you put them into.

Instead of asking which tool to buy or which agent to build, ask where in your outbound motion is that structure missing or unoptimized.

If you want a second set of eyes on how effective your existing structure is, check out our free Outbound Pipeline Inspection.

For a DIY read on what this structure should look like, check out our Outbound Efficiency Pyramid.