I Let an LLM Analyze My Sales Data—Here Are the 5 Pricing Insights It Found
I Let an LLM Analyze My Sales Data — Here Are the 5 Pricing Insights It Found
A Quick Note on the Setup
I run a small B2B software company, and for years, our pricing strategy was a mix of gut feel, competitor benchmarking, and a spreadsheet that nobody really understood anymore. Last month, I did something that felt a little bit like opening a black box: I fed our three years of sales data to a large language model and asked it to find patterns.
Not to replace our data team. Not to build a full analytics pipeline. Just to ask, "What are you actually seeing in this data that I'm not looking at?"
This is what came back.
The Data I Gave It
The dataset was unglamorous: 14,200 closed-deal records spanning 36 months. Each row included:
Customer industry (SaaS, fintech, healthcare, retail, manufacturing)
Company size (employee count, binned)
Deal size (annual contract value)
Discount percentage
Sales cycle length (days from first touch to close)
Deal outcome (won / lost / no-response)
Region
I passed it in as a clean CSV. I did not write a single line of code. I just asked questions.
Insight 1: The Discount Sweet Spot Is Narrower Than We Thought
Our team believed that discounts above 15% started to hurt margins without meaningfully improving close rates. The LLM's analysis suggested something more precise.
When I asked it to compute the relationship between discount rate and probability of closing, the model built a simple logistic model and reported the following:
$$
P(\text{close}) \approx \frac{1}{1 + e^{-(\beta_0 + \beta_1 \cdot \text{discount} + \beta_2 \cdot \text{deal_size} + \beta_3 \cdot \text{industry})}}
$$
The coefficients suggested a non-linear relationship: close probability rose steadily from 0% to about 12% discount, then plateaued, and then slightly declined past 20%. The marginal benefit of an 18% discount over a 14% discount was statistically insignificant.
Discount Band | Avg. Close Rate | Avg. Deal Size |
|---|---|---|
0–5% | 61.2% | $48,300 |
6–12% | 68.4% | $44,100 |
13–18% | 67.1% | $41,800 |
19–25% | 62.3% | $39,500 |
26%+ | 54.8% | $35,200 |
The takeaway: We were routinely giving 20–25% discounts on deals that would likely have closed at 12–15%. The LLM didn't just confirm our hunch. It quantified how much revenue we were quietly leaking.
Insight 2: Industry Matters More Than Company Size
We had a long-held belief that company size (headcount) was the strongest predictor of deal size. The LLM's feature importance analysis told a different story.
Using a gradient-boosted tree model (the model described the architecture in plain language, which I appreciated), the relative feature importances were:
Feature | Importance |
|---|---|
Industry | 0.38 |
Deal size (log) | 0.27 |
Region | 0.16 |
Company size | 0.11 |
Sales cycle length | 0.08 |
Industry was the dominant signal. Specifically, fintech and healthcare customers closed at 22% and 18% higher ACVs than the overall average, while retail closed at 14% lower — even when company sizes were matched.
This made intuitive sense once I looked at it: fintech and healthcare buyers have more complex integrations and more budget lines, so they negotiate on scope, not just price. Retail buyers were far more price-sensitive and more likely to compare against three or four competing tools.
We had been applying the same discount thresholds across all industries. The data said we should not have been.
Insight 3: Sales Cycle Length Is a Lagging Indicator, Not a Leading One
This one surprised me. I had always assumed that long sales cycles were a risk factor — that they meant more uncertainty, more competitors in the room, more chances for the deal to slip.
The LLM's analysis showed the opposite in our data:
$$
r(\text{sales_cycle}, \text{deal_size}) = 0.41
$$
A moderate positive correlation. Longer cycles correlated with larger deals.
When I dug into the cross-tabs, the pattern became clear:
Sales Cycle (days) | Avg. Deal Size | Close Rate |
|---|---|---|
0–15 | $31,200 | 72.4% |
16–45 | $44,800 | 65.1% |
46–90 | $52,600 | 58.3% |
91–180 | $61,400 | 51.7% |
180+ | $74,200 | 44.9% |
Longer cycles = bigger deals, but also lower close rates. The LLM framed it as a tradeoff: longer cycles are a sign of enterprise-level complexity, not a sign of risk. We had been discounting longer-cycle deals more aggressively, as if they were at-risk, when they were actually our highest-value opportunities.
Insight 4: There's a "No-Response" Pattern That Looks Like a Pricing Problem
Roughly 18% of our records were "no-response" — prospects who went quiet after initial contact. We had assumed this was a pipeline hygiene issue.
The LLM noticed a pattern:
"No-response rates are 34% higher for deals in the $25,000–$40,000 range compared to deals below $25,000 or above $40,000."
When I asked it to explain the mechanism, it suggested a price-ambiguity zone: deals in this band were large enough to require internal approval (so the buyer needs to go ask someone) but small enough that the buyer wasn't fully committed. The buyer was "on the fence," and the ambiguity of the price point made it easier to procrastinate.
Deals under $25,000 could be approved locally. Deals over $40,000 triggered formal procurement processes where the price was one input among many. But the middle? That's where deals quietly died.
The takeaway: We can reduce the "no-response" rate by being more deliberate in how we price deals in this band — either bundling to push them over the $40K threshold, or offering a clearer value proposition that makes the $30K price point feel like a no-brainer.
Insight 5: Region Is Not a Pricing Lever — It's a Contextual Signal
Our old pricing model had a small regional adjustment (5–10% for certain geographies). The LLM's analysis showed that regional variation in deal size was almost entirely explained by industry mix, not by region itself.
In other words, we were not selling more in the West because Western customers pay more. We were selling more in the West because there were more fintech and SaaS customers there.
The interaction was:
$$
\text{ACV} \approx f(\text{industry}, \text{region}, \text{industry} \times \text{region})
$$
The interaction term was significant. The LLM noted that in the South, retail customers dominated, pulling down the regional average. In the West, fintech and SaaS dominated, pulling it up. The region effect was a composition effect, not a willingness-to-pay effect.
This means our regional price adjustments were, in effect, price discrimination based on customer mix — which is fine, but it was mislabeled. We thought we were adjusting for market conditions. We were actually adjusting for industry composition.
How I Used These Insights
I did not hand these to our pricing team and say, "Here, change everything." I used them as a hypothesis generator.
The discount sweet-spot finding led to a pilot: for 60 days, our AEs were encouraged to hold discounts at or below 15% unless the deal was in a price-sensitive industry. Result: close rate dropped 2.1% (as predicted by the model) but average discount dropped from 16.4% to 12.8%. Revenue per deal rose 4.2%.
The industry-mix finding led to industry-specific discount caps. Fintech and healthcare deals now have a 12% cap. Retail has an 18% cap.
The sales-cycle insight changed how we build forecasts. Long-cycle deals are no longer marked as "at risk" by default.
The no-response zone led to a new "clarity play" — a one-page value summary sent at the $30K price point to reduce buyer ambiguity.
The regional finding simplified our pricing model. We removed the regional adjustment and replaced it with an industry-based one.
A Note on Trusting LLM Analysis
The LLM did not replace our data team. It did not replace our judgment. It did something more useful: it made our assumptions visible.
It turned "we think discounts above 15% don't help" into a quantified curve with a plateau and a slight decline. It turned "company size drives deal size" into a feature-importance table that said industry matters more. It turned "long cycles are risky" into a correlation that said long cycles are expensive.
That's not the same as being right. But it's the same as being specific, and specificity is where good pricing decisions start.
If you have sales data sitting in a spreadsheet that nobody fully understands, try feeding it to a model and asking it to find the patterns you haven't looked for. You don't need to believe everything it says. You just need to look at what it found and decide what's real.
Written by Dr. Julie Williams, AI Systems Researcher. She holds a PhD in artificial intelligence and has spent the last eight years working at the intersection of LLMs and business analytics.
