glin trains an Explainable Boosting Machine (EBM) on your table and gives it to Claude as an MCP tool. This case study measures what that gives you, compared with Claude alone. All data is synthetic, so the true answer is known for each lead.

Summary

Approach Right segment Tokens for each lead Same answer in 3 sessions
Claude alone 33% 2,000 88%
Claude with 1,000 labeled leads in the prompt 60% 83,000 78%
Claude with glin, 2 models 70% 12,800 98%
Claude with glin, 1 model 71% 12,300 100%
  • Claude with glin got 11 points more than Claude with the pasted table. It used 6.7 times fewer tokens.
  • The best possible score on these leads is 68%. glin is at the limit of the data.
  • glin gave the same answer for the same lead in each new session.
  • In a what-if test, glin was correct to within 3 points of the true effect of each sales action.
  • One 4-class model and two binary models gave the same accuracy.

The task

A B2B software company has 100,000 leads in its CRM. For each new lead, it wants to know one of four segments:

Segment Meaning
high_value_converted The account is $50k a year or more, and it became a customer in 120 days
high_value_not_converted The account is $50k a year or more, and it did not become a customer
low_value_converted The account is less than $50k a year, and it became a customer
low_value_not_converted The account is less than $50k a year, and it did not become a customer

Each lead has 23 columns. Examples are company size, revenue, industry, lead source, demo attended, budget confirmed, discount requested and days since first touch.

Why synthetic data

The first glin case study used real survey data (NFHS-5). Claude already knew much about that subject, so glin added only 6 to 8 points. Also, real data has no known answer, so you cannot check the explanations of a model.

For this study, I wrote the rules that make the data. The rules are specific to this company, so Claude cannot know them. Because I know the rules, I can check each prediction, explanation and what-if against the truth.

The rules in the data

Different columns control value and conversion:

  • Company size, revenue, industry, job level, features requested and an SSO requirement control value.
  • Lead source, demo, budget, competitor, pricing-page visits, trial users, rep tenure and time in the pipeline control conversion.

Some conversion rules are not what a sales manager expects:

  • More stakeholders help, up to 4 or 5 people. A group of 6 or more stops the deal.
  • Webinar leads convert less than content downloads.
  • A discount request of up to 25% has no effect. Above 25%, each point decreases the chance of a sale.
  • After 90 days, the chance of a sale decreases quickly.
  • A hidden “urgency” value increases both value and conversion. This value is not in the data. It connects the two outcomes.

The problems in the data

The CRM export is not clean, as in real work:

  • Revenue is in three formats: $12,500,000, USD 12,500,000 and 12500000.
  • Percents have a % sign. Dates are in two formats. Leads without a trial show n/a.
  • The crm_stage column is set after the outcome. It gives the answer (target leakage).
  • The lead_id column is an identifier.
  • Two columns have no effect. They only follow company size, so they look useful.

Step 1: Claude trains the models

I gave a new Claude session the training file (90,000 leads), the glin MCP tools and Python. I asked it to compare one 4-class model with two models (value and conversion), to train both and to recommend one.

The session took 7.3 minutes and cost $0.20. It did these steps:

  1. It found that crm_stage gives the answer, and it removed that column and lead_id.
  2. It trained three glin models: seg4 (4 classes), value_tier and converted.
  3. It recommended two models, and it said that the result was close.
  4. It found that conversion increases at the end of each quarter.
  5. It found that the two columns without effect get almost no weight.
  6. It found a fault in my data. In some leads, days_since_first_touch does not agree with first_touch_date. The session told me to examine this column. In a real CRM, this is the correct question to ask.

Step 2: Claude answers for 200 new leads

For each of 200 test leads, a new Claude session predicted the segment. Each session saw only the columns of one lead.

Approach Right segment Value right Conversion right Cost for each lead
Claude alone 33% 58% 55% $0.010
Claude with 1,000 leads in the prompt 60% 81% 72% $0.033 (cache in use), $0.34 (no cache)
Claude with glin, 2 models 70% 85% 83% $0.029
Claude with glin, 1 model 71% 87% 82% $0.024

Random selection gives 25%. On all 10,000 test leads, both glin designs got 69.2%, which is equal to the best possible score.

Claude used the glin result correctly in all 400 glin sessions. Each prediction in a session was the same as a direct call to the model.

Step 3: the same answer in each session

I sent the same 40 leads three times, each time in a new session.

  • Claude alone changed its answer for 1 lead in 8.
  • Claude with the pasted table changed its answer for about 1 lead in 4.
  • Claude with glin (1 model) gave the same answer each time.

With 2 models, one lead had a different answer one time. In that session, Claude also used an old test model that the training session did not delete. Delete old models, or give them clear names.

Step 4: the explanations are correct

I compared the effect that glin learned for each column with the true effect.

  • For all nine rules that I checked, the correlation between the learned effect and the true effect is 0.97 to 1.00.
  • glin found each limit and each change of direction. Examples are the stop at 6 stakeholders and the drop above a 25% discount.
  • The learned effects are approximately 0.8 times the true size. The hidden urgency value causes this. The shape and the order of the effects are correct.
  • The two columns without effect get less than 0.5% of the total importance.

Step 5: what must the sales rep do?

I selected one open lead: a finance company with 537 employees, no demo, budget not known and a 34% discount request. I asked which action increases the chance of a sale most: confirm the budget, give a demo or decrease the discount to 10%.

Action True chance Claude with glin Claude with 1,000 leads Claude alone
No change 27% 29% 30% 7%
Discount decreased to 10% 56% 56% 36% 10%
Budget confirmed 63% 66% 42% 15%
Demo attended 67% 69% 47% 13%
Budget and demo 91% 91% 58% 24%
  • Claude with glin called predict once for each action. All of its values are within 3 points of the truth. It selected the demo first, which is correct.
  • Claude with the pasted table selected the same action, but its effects are approximately half the true size.
  • Claude alone put the lead in the wrong segment. Its values are 20 to 67 points too low.

One model or two models?

  1 model (4 classes) 2 models (value and conversion)
Accuracy on 10,000 leads 69.2% 69.2%
Log loss (lower is better) 0.716 0.721
Training time, 90,000 leads 248 s 127 s
predict calls for each lead 1 2
Explanation One list of causes for the segment One list for value and one list for conversion
  • The accuracy is the same.
  • One model has slightly better probabilities, because it learns the connection between the two outcomes. It also needs only one tool call.
  • Two models train in half the time. They give a separate explanation for each outcome. Sales can use the conversion model, and finance can use the value model.
  • Use two models when people act on each outcome. Use one model when you need only the segment.

Why this is better at scale

  • Cost. A pasted table costs approximately 81 tokens for each row in each question. All 90,000 leads need approximately 7.3 million tokens, which is much more than a context window. glin uses approximately 12,500 tokens for each question, for any quantity of data. You train the model one time.
  • Accuracy. glin learns from all 90,000 leads. A pasted table shows Claude only a small sample.
  • Privacy. The training data stays on your computer. Claude sees only the lead in the question.
  • Consistency. The same lead gets the same answer in each session and for each user.
  • Explanation. Each prediction shows the exact contribution of each column, so you can check the answer to “why” and “what if”.

Limits of this study

  • The data is synthetic. On real data, Claude can know more about the subject, and the difference can be smaller.
  • The test has 200 leads for each approach. A difference of less than approximately 7 points is not certain.
  • A different pasted sample, more rows or better instructions can change the result of the pasted-table approach.
  • The rules in this data are causes, because I wrote them. On real data, a model finds associations, not causes. However, that’s where this data-backed output can let a smart model like Claude Opus 5.5 think through causality now that it has statistically accurate correlations.
  • A glin session takes 5 to 7 seconds for each lead. Claude alone takes 3 seconds. Claude uses most of this time to copy the columns of the lead into the tool call.

The test ran on 2026-10-11 with claude-sonnet-5-5 and glin-ml 0.1.4 from PyPI. The 1,128 Claude sessions cost $25.08 in total.