How S&P 500 Companies Are Really Using AI: October 1, 2026 Update

“I can’t keep up.”

“I don’t know how to tell what’s hype and what’s real.”

Given the speed of AI transformation, it’s tough to keep up with developments within your own company, let alone across industries or the entire S&P 500.

That’s why the AIDE Institute tracks new ways AI is being deployed, along with updates to initiatives we’ve already documented.

Six AI deployments worth a closer look

This update includes 389 new applications from 116 companies and 429 changes to previously documented initiatives. Of the new applications, 189 support internal operations and 200 appear in products and services.

Amazon: driver coaching inside a broader safety system

• Application: In-vehicle AI monitoring surfaces real-time safety information so Delivery Service Partner owners can identify driving patterns and coach drivers.
• Reported result: Amazon reported that serious delivery vehicle crashes fell by more than 23% across its network last year.
• Read carefully: The 23% result applies to Amazon’s broader safety program, not to the camera or model alone.
• Source: Amazon delivery safety and AI-powered technology

Five pharma companies: one model trained without sharing data

• Application: AbbVie, Astex Pharmaceuticals, Bristol Myers Squibb, Johnson & Johnson and Takeda jointly trained a protein-ligand model on more than 20,000 proprietary structures. The model trained inside each company’s environment while only model updates moved between participants.
• Reported result: AISB-1-Fed produced high-quality protein-ligand interface predictions on 52.1% of 1,056 held-out private structures, compared with 35.6% for the public baseline and 40.9% for Boltz-2.
• Why it matters: The approach shows how companies can train across valuable private data without placing that data in a shared pool.
• Source: AISB federated drug-discovery model

DoorDash: a narrow agent task with hard checks

• Application: A multi-agent system starts from a Jira ticket, checks rollout data, finds the relevant repositories, edits code, runs validation and opens a pull request to remove stale feature flags.
• Reported result: Across 50 recent stale flags, the system produced usable pull requests for 45. Average cleanup time was 13.8 minutes and average cost was $4.79, compared with DoorDash’s estimate of one to two hours by hand.
• Human role: An engineer confirms the target value and performs the final merge after tests and static analysis.
• Source: DoorDash multi-agent feature-flag cleanup

Live Nation: automation handles volume, people handle exceptions

• Application: Venue Agent answers fan questions about parking, door times, accessibility and venue policy.
• Reported result: Live Nation reported that 95% of questions handled by Venue Agent were resolved without a handoff to a person.
• Earlier pilot: Melody logged more than 37,000 fan interactions and 17,000 customer service sessions during a 12-day BottleRock Napa Valley window. Live Nation reported that 85% of fans got what they needed within three responses.
• Source: Live Nation and Salesforce fan-support agents

Target: AI attached to an existing customer journey

• Application: AI supports teacher and college wish-list experiences inside Target’s existing shopping journey.
• Reported result: Target reported that wish-list creation rose more than 50% year over year, items added more than doubled, and conversion across key back-to-school pages rose nearly 20%.
• Read carefully: These outcomes were reported alongside the program, but they do not isolate the contribution of a single model.
• Source: Target AI personalization disclosure

Cisco: one inference job across two vendors’ chips

• Application: A distributed service connected eight NVIDIA H200 GPUs and eight AMD Instinct MI350X GPUs over a Cisco Silicon One Ethernet fabric.
• Reported result: The combined system produced 73,357.62 tokens per second, or 104.7% of the throughput measured across the two setups separately.
• Why it matters: Multi-vendor inference can help companies continue using older capacity as newer hardware enters the fleet.
• Source: Cisco multi-vendor MLPerf inference benchmark

Activity by sector

Information Technology led this update, followed by Communication Services and Consumer Discretionary.

What to take from this update

• Companies are applying AI to defined units of work rather than abstract transformation goals.
• Reported results are most useful when the source makes clear what was measured and what the number represents.
• People remain involved where review, judgment or accountability matters.

Method note

This update covers cases added from August 25 through September 24, 2026. An application is included when a company or credible public source describes a specific use of AI. Production status and recorded results reflect the public evidence available at the time of review.