An AI agent is now resolving 65% of customer calls at 90% lower cost. That's the number Northwestern Ontario should hold vendors to.

OpenAI's own customer story about Ringg shows an AI agent platform resolving up to 65% of customer calls without a human, at roughly 90% lower cost after switching model generations, running at more than 7 million calls a month. For a Northwestern Ontario business weighing an AI phone or chat vendor, the capability question is now answered by real production numbers — what is left is ordinary vendor discipline, and that part still has to be done carefully.

OpenAI's own customer story, published September 23, says an AI agent platform called Ringg now resolves up to 65% of customer service calls without a human — using OpenAI's GPT-5.6 models, at roughly 90% lower model cost than the previous generation it replaced, at real production scale: more than 7 million connected calls a month, with an average customer-satisfaction score of 4.8. That is one vendor's self-reported results serving large consumer platforms, not a promise about any single Northwestern Ontario business. But the direction of the numbers matters more than the specific vendor: whether AI phone and chat support actually works at scale is no longer a live question, and it just got a lot cheaper to run. For a Thunder Bay clinic, tourism operator, or trades business that has put off an after-hours phone or chat setup because the technology felt unproven, the honest read is that the unproven part is mostly gone — what remains is ordinary vendor due diligence, and that part still has to be done carefully.

What OpenAI actually published

Ringg builds AI agents that handle customer requests across voice, chat, WhatsApp, and web. According to OpenAI's write-up, Ringg's agents now resolve up to 65% of customer requests without human involvement, and migrating selected real-time workloads from GPT-4.1 to GPT-5.6 cut Ringg's model costs by roughly 90% while holding the quality and latency the workloads need. For live interactions, GPT-5.6 Luna interprets what a customer is asking, selects the right tool, and walks them through multi-step tasks; an orchestration layer then executes actions across CRMs, ticketing platforms, payment systems, and scheduling tools, escalating to a human with a conversation summary when it can't finish the job. OpenAI's story also cites specific customers: Policybazaar, an Indian insurance marketplace, reports 67% of requests resolved without a human across more than 57,000 connected requests, with average response time down from 8-12 minutes to under 60 seconds; Practo, a healthcare booking platform, reports 85% first-call resolution and a 70% cut in operating cost versus its prior human-led workflow; Groww, an investment platform, reports 72% self-service resolution on account questions. All three are large consumer platforms operating at a scale no Northwestern Ontario business will match — they're cited here to show the pattern repeats across sectors within one vendor's own customer base, not as a number a local business should expect to hit.

Why 'compared to what' is the number that actually matters

The more durable story here isn't the 65% resolution figure — that depends entirely on how routine a given business's call mix already is. It's the cost curve: a roughly 90% drop in what it costs to run a real-time AI voice workload, produced simply by moving to a newer model generation, is the kind of shift that changes the math for a much smaller operation than a national insurance marketplace. AI voice and chat bots earned a real reputation, not long ago, for being unable to handle an actual customer call without a human quietly finishing the job behind the scenes. This case study is evidence that reputation is now describing where the technology was, not where it is. That doesn't make the buying decision simpler — if anything it raises the stakes, because the question a Northwestern Ontario business needs to ask a vendor has shifted from "does this actually work" to the standard software-procurement questions this hub has already laid out: where is customer data stored and processed, what does the contract lock you into, and what does the pricing actually look like once you're past the demo. Read that checklist before signing anything.

What actually changes for Northwestern Ontario

For a local business or institution with real after-hours or overflow call volume — a clinic booking appointments, a tourism operator fielding the same questions about hours and rates, a trades business missing calls while on a job site — the cost per automated interaction has genuinely moved in the last year, and that's a different calculation than the one that made sense 12 to 18 months ago. A municipal service desk handling routine, non-emergency inquiries faces a similar cost curve, but public-sector adoption carries a higher bar on data handling and public accountability that a falling model price does not resolve on its own. For anyone in a phone-answering or front-desk-style role, the honest framing is that the economics of automating routine, repeatable inquiries just moved — not the judgment calls, complaints, and local-knowledge conversations that still need a person. An employer evaluating this should be specific, with staff and with themselves, about which share of their own call volume is genuinely routine, rather than importing a vendor's headline number.

A practical first step

Before shopping a vendor, pull your own numbers: count what share of last month's calls or chats were genuinely routine and repeatable — hours, pricing, booking status, availability — versus what needed real judgment. That routine share is exactly what's now cheap to automate, and it tells you what resolution rate to actually expect instead of importing someone else's headline figure. Then run any vendor through the data-residency, contract-lock-in, and pricing questions this hub laid out for AI vendor selection generally. On funding: a pilot project like this — software, consulting, and integration costs to adopt an AI tool — sits squarely in what NOIC's Building Blueprints for AI Adoption program is built to support for Northwestern Ontario SMEs, and FedNor's Regional AI Initiative is a second program worth checking, though it excludes retail and service-based businesses by its own published eligibility terms. Confirm current eligibility, intake status, and funding amounts directly with each program before applying — neither of those categories confirms itself. Scoping a pilot against your own actual call volume, before signing anything with a vendor, is exactly the kind of groundwork a local implementation partner can help with.

Sources: Ringg's AI agents resolve up to 65% of customer calls with OpenAI — OpenAI, September 23, 2026: openai.com/index/ringg | Ringg: Multichannel Customer-Service Agents with Model Routing and Production Evaluation — ZenML LLMOps Database: zenml.io/llmops-database/multichannel-customer-service-agents-with-model-routing-and-production-evaluation | Ringg Says Its AI Agents Resolve Up to 65% of Routine Customer Inquiries — Superpower Daily: superpowerdaily.com/posts/ringg-says-its-ai-agents-resolve-up-to-65-of-routine-customer-inquiries | Building Blueprints for AI Adoption (BBAA) — Northwestern Ontario Innovation Centre: nwoinnovation.ca/programs/bbaa/ | Regional Artificial Intelligence Initiative (RAII) — FedNor: fednor.canada.ca/en/our-programs/regional-artificial-intelligence-initiative-raii-northern-ontario. This is commentary, not procurement or funding advice — confirm current vendor terms and program eligibility directly before acting.