Your CRM Is Clean. Your AI Still Sucks. Here's Why.

AI needs to know what "good" looks like for YOUR business—here's how to teach it. Part 2 of a 5-part series on building AI-GTM infrastructure that actually works.

Your CRM is clean.
Your systems are connected.
Your data flows properly between platforms.

Now what?

You can feed AI perfect data and still get terrible results if you haven’t taught it what matters. Most companies skip this step entirely. They assume AI will figure out patterns on its own, or that their existing lead scoring logic will translate it somehow automatically.

It doesn’t work that way.

AI needs explicit instruction on what “good” looks like in your specific GTM motion. That means defining your ICP with precision, mapping your messaging hierarchy, structuring your intent data so AI can actually learn from it, and sharing examples of ideal outputs.

This is Part 2 of our 5-part series on building AI-driven GTM infrastructure. Part 1 covered data hygiene and system architecture. Now we’re covering the inputs: the frameworks and structures that teach AI what qualified means for your business.

---

👋 Hiii, it’s Kaylee Edmondson and welcome to Looped In, the newsletter exploring demand gen and growth frameworks in B2B SaaS. I write this newsletter every Sunday, and wildly, a few thousand of you read it each week. I’m grateful. If there’s ever anything in particular I can help explore, research, and write about, hit me up! I’d love to chat.

---

The Problem with Generic ICP Definitions

I reviewed a demand gen strategy last month where the ICP was defined as: “B2B SaaS companies, 50-500 employees, $5M-$50M revenue, using Salesforce.”

When I asked the VP of Marketing why those were the criteria, she said: “That’s where most of our customers are.”

Fair enough. But when I looked at their closed-won deals from the last 12 months, the pattern was different. Yes, most customers fit those firmographic criteria. But the ones with the shortest sales cycles and highest ACV all had three additional characteristics:

• Recently hired a VP of Revenue Operations

• Currently using at least 2 of their top 5 competitor tools

• Had engaged with their content on LinkedIn in the previous 60 days

Those behavioral signals mattered more than company size. But nobody had documented them. They existed in the pattern-recognition part of the sales team’s brain, not in any system AI could learn from.

This is the gap. Firmographics tell you who could buy. Behavioral signals tell you who will buy. Anything involving AI needs both layers to be useful.

Building Your Behavioral ICP Layer

Most companies stop at firmographics: company size, industry, location, revenue, tech stack. It’s a great start, but that’s table stakes. The competitive advantage comes from the behavioral layer.

Here’s how to build it:

Start with your best customers. Pull your last 20 closed-won deals with the highest ACV and shortest sales cycles. These are your gold standard accounts.

Identify pre-purchase behaviors. For each account, document what they did in the 90 days before they became a customer:

• What content did they consume?

• Which product pages did they visit?

• Did they attend webinars or events?

• What was their social media engagement pattern?

• Were they hiring for specific roles?

• Did they mention pain points publicly?

• Were they researching competitors?

Look for the patterns. You’re not looking for universal behaviors (every customer did X). You’re looking for signal clusters. Maybe 60% of your fastest deals came from accounts that did 3+ of these things:

• Downloaded your pricing guide

• Viewed integration documentation

• Attended a product demo webinar

• Had an executive engage with your LinkedIn content

• Posted a job opening for a relevant role

Translate patterns into data fields. Each behavioral signal needs to become a trackable field in your CRM. Not just “downloaded content” but which content. Not just “visited website” but which pages and how many times.

Weight the signals. Not all behaviors predict conversion equally. Use your historical data to weight signals based on correlation to closed-won deals. In my client’s case:

• Competitor tool usage: 2x weight

• RevOps hire in last 6 months: 3x weight

• LinkedIn engagement: 1.5x weight

• Content downloads: 1x weight

This weighted behavioral scoring becomes how you teach AI what “high intent” actually means for your business.

One client took this approach and discovered that accounts with a new CMO hire in the previous 90 days closed 40% faster than accounts without that signal—regardless of company size. That single behavioral attribute became their highest-weighted ICP criteria.

One additional criteria, especially in enterprise SaaS, is depth of behavioral signals exposed over how many ICP contacts. If one person from Pepsi visits your website, that’s amazing (or maybe even a fluke, right?), but if 8 core personas visit your site, gold.

Defining “Qualified” in Data Terms

I’ve sat through countless meetings where marketing and sales argue about lead quality. Marketing says leads are qualified. Sales says they’re garbage. Nobody can point to specific data that proves either side right.

The problem is that “qualified” means different things to different people, and none of those definitions live in structured data that AI can use.

Here’s what I mean: Your AE knows a lead is qualified when they “have budget and authority and an active project.” But your CRM doesn’t have fields for any of those things. So AI can’t learn the pattern.

You need to translate your qualification criteria into actual data points:

Budget indicators:

• Company size (employees, revenue)

• Recent funding rounds

• Hiring velocity

• Tech stack (shows investment in tools)

Authority indicators:

• Job title/level of contacts engaging

• Number of stakeholders involved

• Executive-level engagement

• Decision-maker viewing pricing

Active project indicators:

• Product research behaviors

• Competitor comparison activity

• Integration documentation views

• Implementation timeline questions

• RFP downloads or sandbox signups

For each qualification criterion, ask: “What data would prove this true?” Then make sure you’re capturing that data.

One SaaS company I worked with defined “qualified” as accounts showing 3+ of these signals within a 30-day window:

• Technical decision-maker viewed API docs

• Economic buyer viewed pricing page

• Multiple stakeholders from same account engaged with content

• Account fit their firmographic ICP

• Currently using a competitor tool

They structured this as a scorecard in their CRM, with each signal clearly defined and tracked. Now when sales asks “why is this qualified?” marketing can point to specific data points, and AI can learn which combination of signals predicts closed-won deals.

Mapping Your Messaging Hierarchy for AI

Most marketing teams have messaging in slide decks and Google Docs. That’s fine for human consumption, but AI can’t learn from a PDF.

You need your messaging structured as data.

Here’s the hierarchy that matters:

Level 1: Category positioning What category do you compete in? How do you define the problem space?

Example: “Marketing automation” vs. “Revenue orchestration platform” vs. “GTM intelligence”

This matters because AI needs to understand what comparison set you’re in when analyzing competitive research or intent signals.

Level 2: Core value propositions What are your 3-5 primary value props? These should be consistent across all content.

Example for a sales intelligence tool:

• Find accounts showing buying intent

• Prioritize outreach based on engagement signals

• Personalize at scale with account insights

• Connect data across your GTM stack

• Measure what actually drives revenue

Level 3: Proof points for each value prop What evidence supports each claim? Customer stories, data points, features, integrations.

Example for “Find accounts showing buying intent”:

• Proof point: Customer story about 3x pipeline increase

• Proof point: Integration with 15+ intent data sources

• Proof point: 92% accuracy in predicting in-market accounts

• Proof point: Case study showing 40% faster sales cycles

Structure this as CRM fields and tags

Create content tags in your CRM that map to this hierarchy:

• Primary category: [category name]

• Value prop focus: [which value prop(s)]

• Proof points referenced: [specific proof points]

• Persona targeted: [which persona]

• Buying stage: [awareness/consideration/decision]

When someone engages with a piece of content, you’re not just tracking “downloaded whitepaper.” You’re tracking “engaged with content focused on [value prop 2], targeting [CFO persona], at [consideration stage].”

Now AI can learn patterns like: “Accounts that engage with ROI-focused content targeting finance personas convert 2x faster than accounts engaging with technical feature content.”

Without this structure, AI just sees generic “content engagement” and can’t differentiate between someone reading a blog post about industry trends vs. someone downloading your pricing calculator.

Labeling Intent Data for Machine Learning

Intent data is only valuable if it’s properly labeled and weighted. Most companies collect signals but don’t teach AI which signals actually matter.

Here’s the framework I use with clients:

Signal taxonomy: Create categories for different signal types:

• Research signals (viewing content, attending webinars, downloading resources)

• Comparison signals (viewing pricing, reading case studies, competitor research)

• Technical signals (viewing docs, testing integrations, sandbox usage)

• Buying committee signals (multiple stakeholders engaging, executive involvement)

• Timing signals (budget cycle indicators, contract renewal dates, hiring activity)

Intent strength scoring: Not all signals indicate the same level of intent:

• Low intent: General industry content, blog reading, social media follows

• Medium intent: Specific use case content, webinar attendance, competitor comparisons

• High intent: Pricing page views, demo requests, technical documentation, trial signups

Recency weighting: Signals decay over time. A pricing page view yesterday means more than one from 6 months ago. Build recency into your scoring (example below):

• 0-7 days: Full weight

• 8-30 days: 70% weight

• 31-60 days: 40% weight

• 61-90 days: 20% weight

• 90+ days: Minimal weight

Frequency patterns: Look for acceleration. An account viewing your pricing page once might be casual research. Viewing it three times in one week is a strong buying signal.

Label your historical data: This is the step most teams skip. Go back through your closed-won deals and label which signals appeared, when, and in what sequence. This labeled dataset is what trains your AI models.

One client labeled 200 closed-won deals with all pre-purchase signals. They discovered that accounts showing 3+ high-intent signals within a 14-day window closed at a 68% rate, vs. 12% for accounts without that pattern. That became their AI trigger for immediate sales outreach.

Creating Structured Content Performance Data

Your content analytics probably track page views, time on page, and download counts. But that’s not enough for AI to learn which content actually drives pipeline.

You need to connect content performance to revenue outcomes:

Tag content by strategic attributes:

• Persona targeted (not just “marketer” but “demand gen manager at Series B SaaS”)

• Buying stage (awareness, consideration, decision)

• Value prop addressed

• Content format (guide, case study, calculator, interactive demo)

• Pain point tackled

Track engagement by account, not just by person: When someone at Acme Corp downloads your guide, that’s an account-level signal, not just an individual lead. Track:

• Which accounts are consuming which content

• How many people from an account engaged

• What sequence of content did they consume

• Time between content engagements

• Mix of content types (blog → guide → case study → pricing)

Connect content to pipeline outcomes: For every deal that closes, map back to which content that account engaged with. You’re building a dataset that shows:

• Accounts that consumed [content X] closed 2x faster

• Accounts that engaged with 3+ pieces of [persona-specific content] had 40% higher ACV

• Accounts that viewed [competitor comparison content] before demo had 50% higher show rates

Measure content influence, not just attribution: Attribution tells you last touch or first touch. Influence tells you what content moved deals forward. Track:

• Did deal velocity increase after content engagement?

• Did more stakeholders get involved after certain content?

• Did accounts move from one stage to another after specific content?

Structure all of this as fields in your CRM, not just in your content analytics tool. AI needs to see the relationship between content engagement and revenue outcomes in one system.

A recent client discovered that accounts consuming their ROI calculator were 3x more likely to reach closed-won within 45 days. But that insight only became actionable when they structured it in their CRM and could trigger AI-powered plays when accounts hit that signal.

Building Your AI Training Dataset

Everything above feeds into your AI training dataset. This is the foundation that makes predictive scoring, automated campaign triggers, and intelligent routing actually work.

Here’s what your training dataset needs:

Historical won/lost deals: At least 100-200 deals (more is better) with:

• All firmographic data

• All behavioral signals they showed

• Content they engaged with

• Sequence and timing of signals

• Time to close

• Deal size

• Win/loss reason

Account progression data: For accounts that didn’t close (yet), track:

• Which stage they reached

• Where they got stuck

• What signals appeared before they stalled

• Whether they’re still active or went dark

Negative examples matter: AI needs to learn what not to prioritize. Include:

• Accounts that showed high intent but didn’t close

• Accounts that were poor fit despite high engagement

• False positive patterns to avoid

Regular dataset updates: Your training data should be refreshed quarterly as you close more deals and learn new patterns. Markets shift, buying behaviors change, and your ICP, more likely than not, might evolve.

One client built their initial training dataset in 6 weeks by going back through 18 months of deals. They labeled 180 closed-won accounts with all pre-purchase signals. That became the foundation for their predictive scoring model, which now routes high-intent accounts to sales 48 hours faster than their old manual process.

The Output: AI That Actually Understands Your Business

When you do this work, AI stops being a black box and starts being a system that understands:

• Which accounts match your ICP (both firmographic and behavioral)

• What “qualified” means in your GTM motion

• Which signals predict buying intent

• Which content moves deals forward

• Which patterns lead to closed-won deals

This is what makes AI useful. Not the tools, not the algorithms, but the structured inputs that teach AI what good looks like for your specific business.

Most companies skip this step. They plug in AI tools, feed them generic data, and wonder why the predictions are wrong and the personalization feels off.

The companies getting real results from AI spend weeks (and most of the time, months) doing this foundational work. They’re not smarter, they just understand that AI is only as good as what you teach it.

Next week in Part 3: We’ll cover workflow automation and decision loops—how to design the GTM engine where humans train, AI executes, and humans refine. But none of that works if AI doesn’t know what good looks like first.

See ya next week,
Kaylee ✌