EGGai

Benchmark / Lead ranking / June 2026

EggAI · Lead ranking benchmark · June 2026 campaign

Claude Code vs EggAI Lead Ranking

Each method picks 500 customers per call list. We check who really spent enough in June.

Claude Code

32.5%churn · toploss 96.1%3 runs ranged 28.6 – 36.2% on churn

EggAI pretrained

38.4%churn · toploss 97.6%single reads 34.6 – 37.0% on churn

EggAI extended thinking

48.4%churn · toploss 98.2%single reads 47.4 – 49.4% on churn

EggAI agent

50.6%churn · toploss 97.8%single reads 47.8 – 50.4% on churn

Score = share of the 500 picked customers who reached their June spend target. Random pick: 5.8% churn, 38.1% toploss.

Claude Code · data goes outside

What the agent reads from customer data is sent to an outside AI service.

EggAI · data stays inside

All data stays on our servers or in the client's private environment.

Same input for every method

A wholesale client has two June call lists. Churn: customers who stopped coming, target 5,000 THB. Toploss: big spenders whose spend dropped, target 1.3× their recent monthly spend. We can call 500 per list. All four methods get the same 3 files.

Transactions · one row per item bought · Jan–May 2026

day_sid	Month	invoice_no	cust	division	class	prod_num	product_name_th	Sales	Qty
20260412	4	INV-0418827	C-104772	FRESH FOOD	Pork	210033871	หมูสามชั้นสไลซ์ 1 กก.	1,890.00	10
20260412	4	INV-0418827	C-104772	DRY FOOD	Cooking oil	110502114	น้ำมันปาล์ม 18 ลิตร	1,045.00	1
20260412	4	INV-0418827	C-104772	NON FOOD	Packaging	310877405	กล่องข้าวกระดาษ 500 ชิ้น	690.00	2
20260419	4	INV-0433105	C-208315	DRY FOOD	Rice	110100045	ข้าวหอมมะลิ 15 กก.	2,350.00	5
20260503	5	INV-0450911	C-208315	FRESH FOOD	Vegetables	220410382	ผักกาดหอม 1 กก.	-85.00	-1

Candidates · churn 70,908 customers · toploss 48,631 customers

row_id	cust	offer_month	offer_min_spend	file
8000001	C-104772	2026-06-01	5,000	candidates_churn
8000002	C-311940	2026-06-01	5,000	candidates_churn
8000003	C-208315	2026-06-01	14,300	candidates_toploss
8000004	C-415006	2026-06-01	6,000	candidates_toploss
8000005	C-500218	2026-06-01	31,800	candidates_toploss

Example rows are made up. Real customer data stays on the secure server.

Four ways to pick the 500

Claude Code

Same 3 files

AI agent as data scientist

Looks at the data

data it reads → Outside AI service

Writes code

Trains a new model

Top 500

Next run starts over.

EggAI pretrained

Same 3 files

Customer history summary

Similar past examples

EggAI model, already trained

Top 500

EggAI extended thinking

Same 3 files

Extended thinking on our proprietary data

Database

spider-engineconsolidates data during inference

EggAI model, already trained

Top 500

EggAI agent

Same 3 files

AI agent on our servers sets up the task

Finds the target, the list and the history to use

Extended thinking on our proprietary data

EggAI model, already trained

Top 500

The agent runs inside. No data goes out.

One test, answers locked away

Data up to 31 May 2026

Claude Code

EggAI pretrained

EggAI extended thinking

EggAI agent

Each picks 500 per list

Score: share who reached target

Real June spend, locked

No method sees June data.
Each method ran 3 times.
June answers are checked once, at the end.

Results

Accuracy

Churn list · axis 0 – 55%

Random pick

5.8%

Claude Code

32.5%

EggAI pretrained

38.4%

EggAI extended

48.4%

EggAI agent

50.6%

Toploss list · axis starts at 90%

Claude Code

96.1%

EggAI pretrained

97.6%

EggAI extended

98.2%

EggAI agent

97.8%

Random pick: 38.1% (below the axis).

Top 10,000

Potential redeemers = customers whose June spend reached the promotion target. How many does each method catch in its top 10,000?

	Random pick	Claude Code	EggAI extended	EggAI agent
Churn · 4,127 potential redeemers in 70,908
Caught in top 10,000	582	1,682	2,033 (+351)	2,002 (+320)
Share of all redeemers caught	14%	41%	49%	49%
Toploss · 18,551 potential redeemers in 48,631
Caught in top 10,000	3,815	7,005	7,688 (+683)	7,644 (+639)
Share of top 10,000 who redeem	38%	70%	77%	76%

Claude Code only hands in 500 names, so we re-ran each Claude run's own final code to score every candidate (the re-run gives back exactly the same 500). Claude Code = mean of 3 runs (churn 1,674 / 1,680 / 1,691).

Consistency

Claude CodeEggAI pretrainedEggAI extendedEggAI agent

Each dot is one run. Tighter = more consistent.

Cost and time at scale

	Claude Code	EggAI pretrained	EggAI extended	EggAI agent
Cost
1 run	$0.60	$0.17	$0.20	$0.13
100 runs	$60	$17	$20	$13
1,000 runs	$598	$172	$198	$126
Time (Claude Code: 32 CPU cores, 32 GB RAM · EggAI: GPU batch inference)
1 run	4.6 min	1.6 min	2.5 min	1.2 min
100 runs	7.6 h	1.7 h	2.0 h	1.3 h
1,000 runs	76 h	17 h	20 h	13 h
Potential redeemers in the churn top 500 (vs Claude Code)
1 run	163	192 (+29)	242 (+79)	253 (+90)
100 runs	16,250	19,200 (+2,950)	24,200 (+7,950)	25,300 (+9,050)
1,000 runs	162,500	192,000 (+29,500)	242,000 (+79,500)	253,000 (+90,500)

1 run = rank both lists once. EggAI cost = GPU time of all reads (3 per list, 2 lists) at $1.25 per GPU-hour; data prep runs ahead of time. Claude Code = Claude Opus 5.5 API as billed + CPU at $0.05 per core-hour. EggAI time is GPU batch inference. Claude Code runs one agent at a time on 32 CPU cores and 32 GB RAM. 100 and 1,000 runs = measured run × 100 / × 1,000. EggAI agent also spends 23 min of one GPU (about $0.47) once, the first time it sets up a new kind of list; later runs reuse that set-up.

Side by side

	Claude Code	EggAI pretrained	EggAI extended thinking	EggAI agent
Data privacy
Private transaction data sent outside	Yes: everything the agent reads goes to an outside AI service	No: stays on our servers or private environment	No: stays on our servers or private environment	No: the agent runs on our servers too
Paid to an outside AI service, per 1,000 runs	$477	$0	$0	$0
Works with the client's private data	Only by sending it to the outside AI service	Yes: plugs into private data inside the client's environment	Yes: plugs into private data inside the client's environment	Yes: plugs into private data inside the client's environment
Accuracy
Churn score	32.5%3 runs ranged 28.6 – 36.2% (7.6 pts)	38.4%single reads 34.6 – 37.0% (2.4 pts)	48.4%single reads 47.4 – 49.4% (2.0 pts)	50.6%single reads 47.8 – 50.4% (2.6 pts)
Toploss score	96.1%3 runs ranged 93.4 – 98.2% (4.8 pts)	97.6%single reads 96.6 – 98.0% (1.4 pts)	98.2%single reads 96.8 – 97.6% (0.8 pts)	97.8%single reads 96.6 – 97.8% (1.2 pts)
Churn vs random pick	5.6×	6.6×	8.3×	8.7×
Churn: potential redeemers in top 10,000	1,682	1,244	2,033	2,002
Toploss: potential redeemers in top 10,000	7,005	7,050	7,688	7,644
Consistency (3 runs)
Churn: lowest to highest	28.6 – 36.2% (7.6 pts)	34.6 – 37.0% (2.4 pts)	47.4 – 49.4% (2.0 pts)	47.8 – 50.4% (2.6 pts)
Toploss: lowest to highest	93.4 – 98.2% (4.8 pts)	96.6 – 98.0% (1.4 pts)	96.8 – 97.6% (0.8 pts)	96.6 – 97.8% (1.2 pts)
Same method every run	No: 3 runs, 3 different methods	Yes	Yes	Yes
Cost
Per run	$0.60	$0.17	$0.20	$0.13
Per 1,000 runs	$598	$172	$198	$126
Per 1,000 potential redeemers (churn top 500)	$3.68	$0.90	$0.82	$0.50
Speed
Time per run	4.6 min (3.3 – 5.9)	1.6 min	2.5 min	1.2 min
1,000 runs	76 h	17 h	20 h	13 h
How it works
Input	Same 3 files	Same 3 files	Same 3 files	Same 3 files
Trains a new model each run	Yes	No	No	No
Set-up for a new kind of list	One sentence + retail/wholesale analyst and data scientist skill.md	One-time set-up	One-time set-up	AI agent on our servers sets it up itself (23 min, once)
Hardware	32 CPU cores, 32 GB RAM	GPU	GPU	GPU

One June campaign. Scores on a 500-customer list can move about ±2 points by chance.
EggAI extended thinking and EggAI agent also use our proprietary data, so EggAI pretrained is the like-for-like match with Claude Code.
EggAI agent and EggAI extended thinking are within chance of each other. The difference is the set-up: the agent writes it itself instead of a person.
EggAI agent was checked against June once. Its May practice check came in slightly below EggAI extended thinking.

EggAI · Lead ranking benchmark · June 2026 campaign