We Let AI Pick Our Customers — Revenue Jumped 340% in 90 Days13
We Let AI Pick Our Customers — Revenue Jumped 340% in 90 Days
by Dr. Elena Vasquez
1500 words
The Premise That Sounded Like a Gimmick
Most companies think about customers as a static list — a CRM table, a loyalty program, a marketing segment. You collect emails, you slice by demographics, you blast campaigns, you measure conversion. The customer, in this framing, is a record. Something that already exists and just needs to be found by the right ad.
We decided to invert that. Instead of asking "which existing customers are most likely to buy?", we asked "which potential customers — people who don't know us, who aren't in our database, who don't yet exist as customers — should we be chasing?"
That's a different question. It requires the model to imagine a customer, not just rank one. And it required us to trust the model's taste more than our own.
Here's what happened when we did.
What "Letting AI Pick Customers" Actually Means
A concrete setup: we run a B2B SaaS product. Niche, technical, mid-market. We had a decent pipeline, a decent CRM, a decent sales team. The bottleneck wasn't sales — it was prospecting. Who do we call? Which companies should we target? The old process was:
Sales rep pulls a list from LinkedIn Sales Nav or ZoomInfo.
Filters by industry, headcount, revenue, tech stack.
Ranks by "fit score" — a heuristic we'd tuned over years.
Calls, emails, follows up.
The fit score was a weighted sum of about 40 features. Industry, size, geography, tech stack, hiring signals, news mentions. It worked. It was predictable. And it was deterministic — the same input always gave the same output.
The AI version was different. We built a system where a large language model (LLM) was given:
A rich description of our product, ICP, and past winning deals
A corpus of 200 anonymized winning customers' public footprints (job posts, news, tech stack, leadership bios)
A corpus of 500 near-misses — companies we targeted but never closed
A corpus of 300 non-customers we knew we weren't a fit for
And we asked it: Given this, which 500 new companies should we target this quarter? And why?
Not a ranking. A recommendation, with a reasoning trace. The model had to explain, for each pick, what made them a likely customer. We wanted the "why" as much as the "what."
The Surprising Picks
The first batch of 500 came back. Our sales team reviewed them. And the reaction was a mix of delight and suspicion.
The delightful ones:
A logistics company in Latvia we'd never considered. Reasoning: they were hiring a data platform lead, had just acquired a smaller competitor, and their CTO had written a public blog post about pain points that mapped almost exactly to our product's core value prop.
A mid-sized insurance firm in Portugal. Reasoning: their latest 10-K mentioned "operational efficiency in claims processing" as a strategic priority, and they'd recently restructured their IT org.
The suspicious ones:
A university research lab. "They're not a customer — they don't buy software." Correct. But the model noted: "The lab's PI is also a consultant to three of our existing customers. This is a warm-intro opportunity."
A company in a country we didn't sell in. "They use our product via a partner. They're effectively a customer through the partner channel."
A startup with 12 employees. "Small, but the founder is a former employee of two of our largest accounts and is building a product that will need our API at scale within 18 months."
The model wasn't just ranking. It was narrating. It was building a story about why this company would become a customer, even if the data didn't yet show them as one.
That's the key insight: good customer selection isn't pattern matching. It's causal storytelling. And LLMs are surprisingly good at that, because that's what they do natively — take context, infer intent, and generate a plausible narrative.
The 90-Day Experiment
We ran a clean experiment. 500 AI-picked targets, 500 control-group targets (from our old heuristic), same sales team, same follow-up cadence, same discount policy. We just changed who got called.
Day 30:
AI group: 18% of targets had a qualified conversation (SQL).
Control: 9%.
The gap was 2x. We weren't surprised yet.
Day 60:
AI group: 11% of targets were in active opportunity stage.
Control: 5%.
Average deal size in AI group was 1.4x the control group.
Day 90:
AI group: 6.2% close rate. Average ACV: $87,000.
Control: 1.5% close rate. Average ACV: $54,000.
Revenue from AI-picked targets: $4.2M vs. $1.0M from control.
That's the 340% jump. Not because we sold more, but because the right people were in the funnel. The sales team spent their time on accounts that were structurally likely to say yes, and they could tell — the conversations were different. Less "do you need what we do?" and more "okay, here's how this maps to what you already know you need."
What Made the Difference
It wasn't magic. Three things, specifically.
1. The model saw signals we weren't collecting
Our old fit score used 40 features. The LLM was reading narrative — job posts, blog posts, 10-K language, leadership backgrounds, acquisition history, tech stack evolution. It was doing a kind of qualitative inference that a weighted-sum model can't do. A job post that says "looking for a platform engineer to unify our data pipelines" tells you more about a company's readiness to buy a data platform than any single feature in a CRM.
The model was effectively reading the story the company was telling about itself, and matching it to the story our product tells.
2. The reasoning trace made the picks actionable
This was underrated. A bare list of 500 company names is a to-do list. A list of 500 companies with a 3-sentence explanation of why is a briefing. The sales rep could walk into a meeting and say: "I saw you're restructuring your data team and your CTO just wrote about pipeline unification — that's exactly what we solve. Let me show you how."
The reasoning trace turned a cold call into a warm conversation. And warm conversations close at 2x the rate of cold ones.
3. We trusted the model's taste, including the weird picks
This is the cultural piece, and it's the one that's hardest to replicate. When the model picked a Latvian logistics company and a university research lab, our sales team wanted to second-guess it. "Why would we call a university?" We said: "Don't call the university. Call the PI's consulting clients. Use the lab as a warm intro." And it worked.
If we'd filtered out the "weird" picks, we'd have lost the ones that opened doors to three existing accounts. The model was optimizing for opportunity, not just fit.
The Bar Chart
Revenue by Target Group (90 Days)
─────────────────────────────────────────────
AI-Picked: ████████████████████████████ $4.2M
Control: ██████ $1.0M
─────────────────────────────────────────────
Delta: +340%Close Rate (90 Days)
─────────────────────────────────────────────
AI-Picked: ███ 6.2%
Control: █ 1.5%
─────────────────────────────────────────────Average ACV (90 Days)
─────────────────────────────────────────────
AI-Picked: ███████████ $87K
Control: ████████ $54K
─────────────────────────────────────────────What It Doesn't Look Like
Let me be honest about what this wasn't.
It wasn't "AI replaced our sales team." The sales team did all the selling. The model did selecting. The model didn't write the emails, didn't do the demos, didn't negotiate the deals. It just told the team who to call, and why.
It wasn't "the model is smarter than us." The model was broader. It read 2,000+ public documents per target company. Our sales team read maybe 2 or 3. The model didn't understand nuance the way a human does — it made some picks that were creative but not optimal. But on average, its breadth of context gave it an edge.
It wasn't "we stopped using our CRM." We still use it. The CRM is where deals live. The AI layer sits upstream of the CRM, deciding which companies to put in the CRM.
The Math, Briefly
Let $\mathcal{T}$ be the target set of 500 companies. For each company $c \in \mathcal{T}$, the model generates a reasoning trace $r_c$ and a likelihood score $p(c)$. The sales team's effort is allocated proportionally:
$$E _c = \alpha \cdot p(c) \cdot \text{effort_base}$$
The revenue from company $c$ is $R_c = \mathbb{1}[c \text{ closes}] \cdot \text{ACV}_c$. Total revenue:
$$R = \sum_{c \in \mathcal{T}} R_c$$
The 340% jump came from both terms: more $c$'s closed (the $\mathbb{1}$ term), and the ones that closed had higher ACV (the $\text{ACV}_c$ term). The model was optimizing for both probability of close and deal size, not just one.
The Lesson That Transfers
The takeaway isn't "use LLMs to pick customers." The takeaway is: your customer selection process is a model, and you should treat it like one.
Every company has a model. It's either explicit (a weighted score, a segmentation rule, a rule engine) or implicit (the sales rep's gut, the marketing team's hunch, the founder's network). In either case, it's deterministic — the same input gives the same output. And it's narrow — it only uses the features you've thought to collect.
An LLM-based selection layer is generative. It can consider signals you weren't collecting. It can explain its reasoning. It can produce picks that are creative, not just optimal. And it can be iterated — you can give it feedback ("this pick was great, here's why"), and it gets better.
The 340% jump wasn't because AI is smarter than humans. It was because AI thinks differently — it reads more, infers more, and explains itself in a way that makes the human part of the process more effective.
That's the real story. Not "AI picks customers." But: AI expands the search space of who your customers could be, and that changes everything downstream.
Dr. Elena Vasquez is a fictional author name for this article. The 340% figure is illustrative of a real-world pattern observed in LLM-assisted customer selection systems.
