GPT-Live replaces ChatGPT's voice mode with one that listens while it talks/GPT-5.6 Sol ships a day later, 54% cheaper on agentic work/Salesforce's Shopper, Buyer & Merchant Agents go live inside ChatGPT/Meta auto-labels AI-made and AI-edited ads on Facebook and Instagram/Klaviyo's Composer agent enters public beta/The squeezed shopper · Michigan 49.5, CPI 4.2%/GPT-Live replaces ChatGPT's voice mode with one that listens while it talks/GPT-5.6 Sol ships a day later, 54% cheaper on agentic work/Salesforce's Shopper, Buyer & Merchant Agents go live inside ChatGPT/Meta auto-labels AI-made and AI-edited ads on Facebook and Instagram/Klaviyo's Composer agent enters public beta/The squeezed shopper · Michigan 49.5, CPI 4.2%/
This week OpenAI answered two separate questions on back-to-back days: what should talking to AI feel like, and what should thinking with AI cost. GPT-Live changes the first — it listens while it talks, the way a person does, and quietly hands off anything hard. GPT-5.6 Sol changes the second — the same reasoning now runs at roughly half the compute. Read next to Salesforce's new shopping agents and Klaviyo's marketing agents going live the same week, the pattern for marketers is one agent layer getting a cheaper brain and a more human voice, all at once.
The Marquee
E-Commerce + CX
OpenAI built the agent's voice and its brain in the same week.
The hand-off
The hand-off, not the upgrade. GPT-Live keeps the conversation natural while GPT-5.6 Sol does the thinking, then hands the answer back mid-sentence. Source: OpenAI.
What happened. On July 8, OpenAI launched GPT-Live, a full-duplex voice model replacing Advanced Voice Mode in ChatGPT — it listens and speaks at once, and for anything hard it quietly delegates to a frontier model and folds the answer back into the conversation. The next day, OpenAI previewed GPT-5.6 Sol, its new flagship — CEO Sam Altman told CNBC it's 54% more token-efficient on agentic coding work than what came before.
Why it matters. For two years, the "AI voice assistant" lived in one product and the "AI reasoning engine" lived in another — marketers picked a chatty front end or a smart back end, not both. GPT-Live collapses that choice. It's built to sound present in the room — soft acknowledgments, natural pauses, room to think — while routing anything that needs real reasoning to GPT-5.6 Sol behind the scenes. That's the architecture every AI customer-service or voice-shopping agent will copy within a year. And because Sol does that reasoning for roughly half the compute, running an always-on voice agent for every customer conversation — not just the ones that escalate to a human — suddenly pencils out. Your brand's phone tree, hold music included, just got put on notice.
If your brand still runs a phone tree, a chat widget, or an IVR script, you have about two years before "it sounds like a person" stops being your differentiator.
Three Quick Hits
Three other things that moved this week.
SFShopper · Buyer · Merchant Agents
E-Commerce + CX
Salesforce put a shopping agent inside ChatGPT.
On July 6, Salesforce made Agentforce Commerce generally available — a Shopper Agent that carries a customer from browsing through checkout and into service, a Buyer Agent that closes B2B deals over WhatsApp and SMS, and a Merchant Agent that manages catalogs in plain language. All three are wired natively into ChatGPT and Google's AI Mode from day one.
What to do: ask your Salesforce rep whether your storefront qualifies for an early pilot before a competitor's does.
METAAI-ad labels · C2PA detection
Ethics + Privacy
Meta just started grading your ads' honesty.
Meta updated the "About this ad" disclosure on Facebook and Instagram to auto-label ads made or heavily edited with generative AI. It catches Meta's own tools and detects outside edits through C2PA metadata from tools like Photoshop and DALL·E.
What to do: pull last month's paid creative and check it for undisclosed AI edits before Meta's detector flags it for your customers.
$200BVerified AI-ad spend · 65 countries
Marketing Data + Media
Someone finally has receipts on AI-platform ad spend.
Guideline extended its Ad Intelligence dataset to capture verified, transaction-level ad spend on ChatGPT and Perplexity — about $200 billion in annual media investment across 65 countries, built from real transactions instead of platform self-reports.
What to do: bring one Guideline number into your next budget meeting the next time a vendor cites a self-reported AI-ad figure.
Brand Spotlight
KLAVIYOComposer + Customer AgentPublic beta · shared data
Klaviyo showed what happens when two agents share one brain.
Klaviyo moved Composer, its AI marketing agent, into public beta on June 30, alongside a major upgrade to its existing Customer Agent. Composer audits every live campaign, flow, and segment, ranks the revenue opportunities it finds, and builds the campaign itself instead of just suggesting one. Customer Agent now handles support across web chat, email, text, and WhatsApp — and, this is the part worth copying, the two agents share the same real-time customer record. A conversation Customer Agent has this morning makes the campaign Composer ships this afternoon smarter. Spanx and Dermalogica got early access in private beta before this week's wider rollout.
Before you buy the next AI marketing point solution, ask what it shares with the tools you already own — that's the question Klaviyo just answered for its own products.
The Marketer's Playbook
Three things to consider this week.
Three concrete moves a marketer could make. If you do nothing else from this issue, consider these.
01
Pilot a full-duplex voice agent on one low-stakes channel.
Order status, appointment confirmations, simple returns — pick something low-stakes and see what "listens while it talks" actually changes about containment and satisfaction, before you bet the whole call center on it. By end of month.
Re-run your agentic workflow math against GPT-5.6 Sol's efficiency numbers.
A workflow that wasn't worth the API bill in June, at roughly half the compute cost, might be worth shipping in Q3. Pick one and rerun the numbers. By next issue.
The more the machine decides, the more the human deciding is worth.
This week's news is a story about delegation. GPT-Live hands hard questions to a bigger model. Klaviyo's two agents hand each other customer context. Salesforce's agents hand a shopper from browsing to checkout to service. None of that works without a person deciding what's worth handing off — and what isn't. That's the finding behind McKinsey.org's new piece on the human skills 2026's AI-driven workplace will actually reward: judgment, creativity, and the willingness to ask a better question keep gaining value exactly as the routine work disappears. Wake Forest's motto is Pro Humanitate — for humanity. This week's agents make the case for it without meaning to: the more the machine does, the more the human doing the deciding is worth.
A note from Ged
This is not an assignment. It's for your use as you see fit — read it, skip it, forward it, file it. The Playbook is a set of options, not a checklist.
Michigan sentiment finished June at 49.5, prices are still running 4.2% hotter than a year ago, and NRF is forecasting 4.4% retail growth this year — the shopper your new voice agent is chatting up is inflation-tired but still spending.
What to Watch Next Week
Three things to follow.
BLS June CPI · Tuesday, July 14May's headline hit a three-year high at 4.2%. If June holds or climbs, expect the squeezed-shopper story to dominate every voice-agent pitch this quarter.
Gemini 3.5 Pro · Ships July 17Google delayed its rebuilt flagship to the same day this issue publishes, after a scramble to fix coding and reasoning benchmarks. Watch whether it actually lands, or slips again.
GPT-Live's video and screen-share supportOpenAI says full-duplex voice-with-video is coming "soon." Watch for the rollout — it's the difference between a phone agent and a screen-share agent.
Wake Forest University · School of Professional Studies
The thinking behind this brief comes from industry.
AI in Marketing Weekly is written by Ged King alongside the Master of Digital Marketing & AI — a program where faculty and students alike are working professionals, applying this in real jobs every week, not studying it from the sidelines.
