CData
Whitepaper

Serving Context to Probabilistic AgentsCorrect Answers with 93% Fewer Tokens

27 AI models chose their own steps through two MCP gateways with the same business context. Through Connect AI Gateway, more models reached the same correct answers with less effort and at a lower cost.

Download Full Report
Download Executive Summary
To Reach the Same Correct Answer
93% fewer tokens
1/10 the tool calls

Connect AI Gateway against the MCP gateway, correct answers only: medians across the 44 model-and-task pairs both got right.

Median Tokens Per Run, Across All Runs
Connect AI Gateway 87K
MCP gateway 914K

What we found

27 models from 12 developers, three governed-data tasks across live CRM, warehouse, and ITSM data, 810 runs.

Measure
Connect AI Gateway
MCP gateway
Tasks completed correctly
83%
36%
Cost per correct task, all 27 models
$0.26
$5.67
Models that completed every task
16 of 27
3 of 27
Confidently wrong answers, read tasks
7%
27%
Runs with unauthorized writes
10%
29%
Median tokens per run, across all runs
87K
914K
Median tool calls per run, across all runs
8
47

Cost per correct task counts failed runs. Counting correct answers only, models used 93% fewer tokens and a tenth of the tool calls through Connect AI Gateway.

How each gateway served context

Both gateways had the same business context, with identical business definitions. They differed in how that context reached the model and where the work on the data happened.

Connect AI Gateway
MCP gateway
Business context
The Context Engine serves only the entries a question needs, when the agent asks
All of it, about 8,600 characters, loaded into every session up front
Getting the data
One request across all three systems
A separate request to each system, with raw records returned a page at a time
Processing the data
The Virtual Data Layer matches, filters, totals, and ranks server-side and returns a short result
The model matches accounts, filters, totals, and ranks in its own reasoning

The study doesn't separate the two mechanisms, so it credits the result to both together.

Methodology

810 runs 27 models from 12 developers, three tasks, five runs each through each gateway.

Live connections, synthetic data A CRM, a cloud data warehouse, and an ITSM system, with data built to behave like a real company's.

Same business context System context, semantic definitions, and company knowledge on both gateways.

Same limits Same models, prompts, reasoning effort, 40-turn cap, and scoring on both gateways.

Get the research

Executive Summary (13 pages)
The results, how each model tier held up, and questions to ask when you evaluate a gateway.

Full Research Report
Full methodology, results for all 27 models, and the governed-write findings.

Testing Harness (GitHub)
Run the benchmark against your own systems and models.