AI-Powered Betting Assistant for Sports Wagering Platform

fanduel · Media & Entertainment · 2025 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Sports wagering assistant

FanDuel wanted a generative betting assistant that could help users research and assemble wagers while respecting state-level gambling regulation, responsible-gaming safeguards, latency constraints, and customer control.

Constraints:

Architecture

FanDuel AceAI is an in-app betting assistant built on Bedrock/Nova models with state-rule handling, responsible-gaming guardrails, static and dynamic prompt tests, and a user-control loop that avoids autonomous betting.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  aceai-in-app-assistant["AceAI in-app assistant"]
  bedrock-model-layer["Bedrock model layer"]
  state-rules-layer["State rules layer"]
  responsible-gaming-guardrails["Responsible gaming guardrails"]
  customer-control-loop["Customer control loop"]
  static-and-dynamic-tests["Static and dynamic tests"]
  customer-feedback-loop["Customer feedback loop"]
  aceai-in-app-assistant --> bedrock-model-layer
  aceai-in-app-assistant --> responsible-gaming-guardrails
  aceai-in-app-assistant --> state-rules-layer
  responsible-gaming-guardrails --> customer-control-loop
  customer-control-loop --> aceai-in-app-assistant
  static-and-dynamic-tests --> bedrock-model-layer
  static-and-dynamic-tests --> responsible-gaming-guardrails
  customer-feedback-loop --> static-and-dynamic-tests
Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component Role Interacts with
AceAI in-app assistant Conversational betting assistant inside FanDuel customer experience. Bedrock model layer, Responsible gaming guardrails, State rules layer
Bedrock model layer Runs Nova Lite and Nova Pro for cost-effective generation. AceAI in-app assistant
State rules layer Applies state-specific wagering constraints and data residency considerations. AceAI in-app assistant
Responsible gaming guardrails Detects harmful prompts and links users to support resources. AceAI in-app assistant, Customer control loop
Customer control loop Keeps research, discovery, and wager decisions under user control. AceAI in-app assistant
Static and dynamic tests Continuously test prompts live and offline. Bedrock model layer, Responsible gaming guardrails
Customer feedback loop Feeds assistant quality and behavior learning back into development. Static and dynamic tests

Data flow

User requests go through the in-app assistant, Bedrock Nova models, state-specific rules, responsible-gaming responses, and live/static evaluation loops while customers retain control.

3 supporting quotes
“we finally settled on Nova Lite and Nova Pro because that delivers what we want for us and it's very cost effective as well.”transcript, stated
“AceAI is very sensitive to these kind of questions and will respond immediately with a link to the responsible gaming page”transcript, stated
“they have a list of static and dynamic tests where they're constantly testing the prompts that are coming in live as well as statically behind the scenes.”transcript, stated

Why this architecture

The design rejects full autonomy because regulated wagering products require user control, responsible-gaming safeguards, and state-specific constraints.

“But in this case, our customers wanted to retain control. So throughout the betting experience, they wanted to be able to do the research themselves, stop, consider things, do the discovery themselves”transcript, stated
“all of those elements were seamlessly handled while maintaining that control at every single step of the journey.”transcript, stated

Models used: Nova Lite · Nova Pro · Amazon Bedrock models

Scale

Alternatives considered (roads not taken)

End-to-end autonomous betting agent

Rejected because: FanDuel explicitly did not go to full autonomy because users wanted to research, stop, consider, and control each step.

“We did not go to full autonomy. And when you're considering what to build, you're gonna hear a lot about autonomous agents that handle things end to end.”transcript, stated

Model choice without cost and scale constraints

Rejected because: The team settled on Nova Lite and Nova Pro because Bedrock scaling and token-cost control fit the product.

“Being serverless, Bedrock scales very well. And if you speak to Santosh very nicely, he'll put up the limits for you. And costs obviously are your tokens. So as long as you control the token, your costs will be controlled as well.”transcript, implied

Prototype responsible-gaming handling

Rejected because: The prototype mishandled harmful prompts, so AceAI added immediate responsible-gaming responses.

“in the prototype, if somebody asked, say, "Hey, I'm gonna put all my money and lose the house," it would say, great, good luck. It doesn't handle that.”transcript, stated
“AceAI is very sensitive to these kind of questions and will respond immediately with a link to the responsible gaming page”transcript, stated

Unbounded token use for every betting prompt

Rejected because: FanDuel chose cost-effective Bedrock Nova models and relies on token control for cost management.

“Being serverless, Bedrock scales very well. And if you speak to Santosh very nicely, he'll put up the limits for you. And costs obviously are your tokens. So as long as you control the token, your costs will be controlled as well.”transcript, implied

Patterns

human-in-the-loop-gate (safety)

AceAI is intentionally assistive, leaving bet construction control and final action with the customer instead of acting autonomously.

“But in this case, our customers wanted to retain control. So throughout the betting experience, they wanted to be able to do the research themselves, stop, consider things, do the discovery themselves”transcript, stated

safety-policy-enforcement (safety)

The assistant detects responsible-gaming risk prompts and routes the user to responsible-gaming support instead of encouraging harmful behavior.

“AceAI is very sensitive to these kind of questions and will respond immediately with a link to the responsible gaming page”transcript, stated

model-routing (model-ops)

FanDuel evaluated model choices and selected Nova Lite and Nova Pro because they matched the product's cost and performance needs.

“we finally settled on Nova Lite and Nova Pro because that delivers what we want for us and it's very cost effective as well.”transcript, stated

eval-driven-development (evaluation)

The team runs release test suites and uses both static and dynamic prompt evals from live user feedback.

“they have a list of static and dynamic tests where they're constantly testing the prompts that are coming in live as well as statically behind the scenes.”transcript, stated

Tradeoffs

How autonomous the betting assistant should be.

Chose assistive recommendations with user control over full autonomy.

Rationale: FanDuel explicitly avoided full autonomy because customers wanted control and the product sits in a regulated gambling context.

“We did not go to full autonomy. And when you're considering what to build, you're gonna hear a lot about autonomous agents that handle things end to end.”transcript, stated

Which model stack to use for the assistant.

Chose Nova Lite and Nova Pro on Bedrock over other available model choices.

Rationale: The team chose models that delivered acceptable quality while controlling cost and leveraging Bedrock's serverless scaling.

“Being serverless, Bedrock scales very well. And if you speak to Santosh very nicely, he'll put up the limits for you. And costs obviously are your tokens. So as long as you control the token, your costs will be controlled as well.”transcript, implied

Gotchas (what bit them)

Sports wagering assistants inherit state-by-state regulatory and latency constraints.

Root cause: Some rules must stay within the state while central rules live elsewhere, creating a cross-region decision path.

Resolution: Design the architecture around state-specific rule placement and latency-sensitive decisioning.

“the central rules lie in a different place. The state-specific rules lie within the state itself. So obviously there's a latency problem”transcript, stated

Prototype assistants can miss harmful responsible-gaming prompts.

Root cause: Early prototypes may optimize betting assistance before policy guardrails are complete.

Resolution: Add explicit responsible-gaming handling before production use.

“in the prototype, if somebody asked, say, "Hey, I'm gonna put all my money and lose the house," it would say, great, good luck. It doesn't handle that.”transcript, stated

Full autonomy is inappropriate for this product experience.

Root cause: Customers wanted to retain control over betting choices in a regulated financial-risk setting.

Resolution: Keep the assistant in an advisory role and preserve control through every step.

“all of those elements were seamlessly handled while maintaining that control at every single step of the journey.”transcript, implied

Evaluation

FanDuel evaluates AceAI through release test suites, static and dynamic prompt tests, customer feedback loops, and product usage/drop-off instrumentation.

Metrics: 16.6 million bets during one Super Bowl · nearly 3 million active users during the game · peak of 70,000 bets · thousands of tests in the release test suite

“We had over 16.6 million bets just during that game. Nearly 3 million active users during the time of the game. And we peaked at 70,000 bets”transcript, stated
“We have thousands of test suites, sorry, thousands of tests in our test suite and we run those every time we make a release”transcript, stated

Outcomes

The transcript is a vendor/cloud-stage talk and gives platform scale numbers, but not controlled business-impact metrics for AceAI itself.

“With FanDuel's AceAI, a generative AI betting assistant, your experience for bets just got a whole lot easier.”transcript, stated
“The system is understanding how many times somebody has actually used this AceAI program, what are their drop-off rates?”transcript, implied