AI Pulse

Uber Fired 10% of Its Support Team. OpenAI Shipped a Platform With Humans Built In. Both Stories Dropped This Week.

By Felix Maru · July 24, 2026 · 7 min read

On July 23, 2026, Uber announced it was cutting 10% of its customer service workforce. The internal memo cited AI as the reason. Two days earlier, on July 22, OpenAI launched a product called Presence, an enterprise platform for deploying AI agents in support and voice workflows, built with human escalation paths as a required design feature, not an afterthought. These two stories are not contradictions. Together, they are a precise picture of the two directions a company can go when it decides to "embrace AI" in support. One of those directions leads somewhere worth going.

What Uber Actually Said

Uber's VP of Global Community Operations, Megha Yethatika, sent a memo telling the community operations team that the company's structure had become, in her words, "too complex and siloed." According to reporting from Outlook Business, Yahoo News, and Engadget, the memo acknowledged that the team had already been adopting AI but stated that unlocking more potential required a simpler organizational structure layered on top of it. Remote employees were also told to relocate to hub offices.

This was reported as the first time Uber has directly and explicitly linked workforce reductions to AI-driven efficiency. It followed a separate 23% cut in Uber's HR operations earlier in June.

I want to be precise about what the memo actually said and what it did not say. The VP did not say "AI is handling the work so we need fewer people." She said the org had become too complex to get AI working well inside it. The stated goal was to simplify the structure so that AI could be layered in effectively. That is a meaningfully different framing, and it is the one worth paying attention to. The headlines calling it "replacing workers with AI" may be reading the result backward from the cause.

That said: the outcome is still 10% of a customer service team out of work, and the lesson for every other company watching is the same regardless of the sequencing.

What OpenAI Presence Actually Is

OpenAI launched Presence on July 22, positioning it as an enterprise product for deploying trusted AI agents in production support and voice workflows. Coverage came from VentureBeat, Street Insider, and Help Net Security, among others.

Presence bundles several components that support teams have historically had to assemble themselves: company policies and standard operating procedures baked into the agent's behavior, permission controls defining exactly what data and systems the agent can touch, pre-deployment simulations for testing edge cases before going live, and escalation rules that route conversations to human agents when needed.

The product uses OpenAI's Codex coding assistant to analyze production data and escalations, then propose updates. Human teams review and approve those updates before they go live. So the improvement loop is continuous but human-gated.

OpenAI said Presence powers its own English-language phone support line and resolves around 75% of inbound issues without human assistance. A Codex-driven improvement process reduced human handoffs by roughly 15 percentage points within 10 days of launch. (Paraphrase of reporting by MLQ.ai, July 22, 2026.)

Currently available through a limited general availability program with OpenAI Forward Deployed Engineers leading each implementation, this is not a self-serve product. Deployments go through OpenAI or select systems integrators. That pricing model, as The Register noted, means OpenAI is effectively charging enterprise consulting rates to deploy it.

The 75% Number. And the 25%.

Seventy-five percent auto-resolution is a strong number. I am not going to dismiss it. But the way you should read it depends entirely on what happened in the other 25% of conversations.

In the Presence architecture, that 25% did not fail. Those conversations hit the escalation rules that were designed in from the start and moved to a human agent. The question every support team should be asking is not "can we get to 75%?" but "what does our escalation path look like when AI hits its ceiling, and is a human there to catch it?"

When OpenAI describes human handoffs as a "first-class feature" of Presence, that is not marketing language. It reflects a real architectural choice. The escalation logic is not bolted on because something went wrong. It was planned because the product team understood that the 25% of conversations that need a human are, by definition, the most complex, the most sensitive, and the most consequential ones. Those are the conversations where getting it wrong costs you a customer, a lawsuit, or a reputation. Those are the conversations you absolutely do not want to hand to an AI that is not ready for them.

This is the design pattern worth copying: AI handles the volume, humans own the edge. The edge is where the real support work is anyway.

Why "Cut Headcount to Embrace AI" Is the Wrong Goal

Here is what I have seen happen, more than once, when a team uses AI adoption as the justification for reducing support headcount before the AI is fully capable.

You shrink the team. The AI handles 60%, maybe 70%, of volume that used to go to humans. That part works. But the remaining 30-40% of tickets are the genuinely hard ones, and now there are fewer people to handle them. Response times on complex issues go up. Escalations pile up. CSAT holds initially because the easy tickets are getting faster answers, then starts dropping because the hard ones are getting slower, worse answers from an understaffed team.

The AI did not cause that outcome. The sequence did. You reduced human capacity at the same time you introduced a new system that generates escalations, needs human oversight on edge cases, and requires someone to review and approve its improvement updates. The humans doing that work are not overhead. They are the reliability layer the AI runs on top of.

Uber's VP framed this correctly, even if the headline did not: she said you need an effective human organization in order to layer AI on top of it. That is the right sequence. Streamline the org, clarify what the humans need to do, build the AI layer to handle what it can, and measure whether your remaining human agents are now spending their time on higher-value work. Headcount reduction can be an outcome of that process. It should not be the stated goal going in.

What to Actually Do If You Are Evaluating This Now

CX Dive reported in July that US customer service job postings have dropped roughly 10% below pre-pandemic levels. The jobs disappearing are the purely transactional ones. AI is simultaneously creating demand for agents who can handle knowledge-intensive, judgment-heavy work. The mix is changing, not necessarily the total headcount, and that distinction matters if you are managing a support team right now.

Here is the checklist I would run through before any Presence-style deployment:

Sources

If you are working through how to frame AI deployment to your own support team, or thinking about where to set escalation thresholds, drop me a note. This is the conversation I find most useful to have right now.

Share 𝕏 in ✉

Comments