Ally Financial's Ally.ai platform writes post-call summaries for customer care reps using OpenAI models on Azure, with rep acceptance rising to 90% and minutes saved per call
Ally Financial built Ally.ai, an internal platform that routes work to LLMs including OpenAI models on Microsoft Azure, and its first production use case drafts summaries of customer care calls that reps review and accept or edit instead of writing notes by hand (reported by American Banker). After about a year and a half in production, CIO Sathish Muthukrishnan said rep acceptance of AI summaries rose from the mid-teens to 90% and summary accuracy from 40% to above 80%, and the use case expanded from banking to insurance and auto (reported by American Banker, July 31, 2025; accuracy is Ally's own internal measure). In 2026 he put accuracy at around 85% with three to five minutes saved per call across roughly 10,000 calls a week (reported by American Banker, June 1, 2026; per-call savings are Ally's estimate, not an audited figure).
Customer care associates had to take notes and write summaries after each call, which competed with the time and attention they could give customers (reported by Microsoft customer story and American Banker)
Ally.ai sends the conversation to an LLM with system prompts for summarization
LLM drafts a structured call summary
Associate reviews, edits or accepts the summary
Summary is saved to the customer record; edits and thumbs up/down feedback are captured to improve prompts
Human-in-the-loop: Associates review every AI summary and can edit it before saving; pilot associates initially ran in parallel and rated summaries; ongoing testing and bias checks (per American Banker).
Inaccurate or incomplete summaries saved to customer records if reps over-trust output
Bias introduced through prompts or training conversations (Ally runs ongoing checks)
Limitations
Summarization assistant with rep review rather than an autonomous agent
Accuracy and per-call time savings are Ally's own figures relayed in trade press; no external audit; per-call savings reported as 30 seconds to 2-3 minutes in 2025 and 3-5 minutes in 2026
Exact OpenAI model version not named in the sources
Rep acceptance of AI call summaries rose from about the mid-teens to 90% after about 1.5 years in production Reported by Ally CIO Sathish Muthukrishnan via American Banker, July 2025 · Source (americanbanker.com)
Summary accuracy rose from 40% to above 80%, later put at around 85% Reported by Ally via American Banker; Ally's internal accuracy measure, method not published · Source (americanbanker.com)
Three to five minutes saved per call on an average of about 10,000 calls a week averaging 15 minutes each Reported by Ally via American Banker, June 2026; Ally's estimate · Source (americanbanker.com)
Among customer care reps, the acceptance of summaries provided by AI has gone from about mid teens to 90% so the team has been able to perfect how AI works, and the accuracy and comprehensive nature of summarizing the conversation has gone up from 40% to north of 80%.
: Added in batch-34; source fetched 2026-10-27 and key figures confirmed in page text; figures attributed.
: Lumi spot-check: added a verbatim quote to the Microsoft evidence entry (page re-fetched 2026-10-27; quote confirmed in page text; supports the problem statement); kept Verified-production.
Air India began building AI.g on Azure OpenAI after Microsoft launched the service in November 2022, and calls it the first airline generative AI customer-service deployment at scale. Microsoft's customer story (February 11, 2026) says AI.g handles about 40,000 queries a day across more than 1,300 question types, from booking changes to refund requests. It has resolved 13 million+ conversations at a 97% success rate, so only 3% go to a human (reported by Microsoft, the vendor). In the same story, CDTO Dr. Satya Ramaswamy cites '97% of 4 million-plus customer queries', which doesn't match the 13 million figure. It is mainly an answer assistant; agents that run full workflows such as refund processing are still in trial
Airbnb started its AI strategy at the bottom of the funnel with customer service. It rolled out an AI customer service agent in spring 2025 and expanded it to all US English users in May 2025. The agent is built on 13 models from OpenAI, Alibaba, Google and open-source providers, relying mostly on Qwen because it is 'fast and cheap'. It shows action buttons and links so guests can complete tasks such as reservation changes or cancellations. It must also work across languages, follow nearly 100 policies and escalate accurately to humans, especially for trust and safety incidents. In October 2025 Airbnb said the agent had cut users needing a live representative by 15% and average resolution time from nearly three hours to six seconds (reported by Fortune/Bloomberg). On the Q1 2026 earnings call CEO Brian Chesky said it resolved over 40% of inquiries, up from about a third the prior quarter. He also said cost per booking fell about 10% year over year, which can't be attributed to the agent alone (reported by CX Dive). The figures are company-reported, but they come from earnings disclosures over about a year of production
easykind, a telehealth start-up launched in 2023, had two Patient Support Coordinators manually triaging the inbox each morning and could only promise 24-hour replies, with phones limited to 9-to-5. Lorikeet's agent now reads every incoming email, tags it Tier 1, 2 or 3 and routes it to the right coordinator queue. 'Buddie', embedded in the logged-in patient experience, answers FAQs around the clock under guardrails required by Australian rules: it never names a product, and if a patient reports concerning side effects it tells them to stop the medication and call emergency services, and escalates to a human. Lorikeet built a Salesforce Service Cloud integration. Email response time fell 92% from 24 hours to roughly two (reported by Lorikeet); Head of Patient Support Steph Galileos estimates four additional hires would be needed to match Buddie's capacity (an estimate of avoided hires, reported by Lorikeet; vendor-published)
Riachuelo, a Brazilian fashion retailer, replaced its previous chatbot with a Genesys agentic virtual agent on WhatsApp. Intelligent CIO reports self-service at 84% versus 30% with the old chatbot, customer satisfaction up 60% and technology costs down 26% (reported by Intelligent CIO LATAM, October 15, 2026, César Nicoletti quoted). The article gives no measurement window, and the CSAT and cost changes also reflect the wider platform change, not the agent alone.
Australian broadband provider Superloop has deployed two agentic AI customer assistants, Teddy and Mo, plus two self-service diagnostic tools in its app (Refreshify and X-Ray). CEO Paul Tyler told investors that the agentic systems now handle more customer interactions than human employees (reported by iTnews from Superloop's FY26 half-year results). Tyler said the automation apps cut inbound support calls by 30% over the past 18 months and improved both customer experience and cost-to-serve (reported by Superloop via iTnews; the 30% covers the diagnostic apps and AI together, so it is not attributable to the assistants alone).