OpenAI wants AI agents (software that takes action on your behalf, rather than just answering questions) to work for accountants, doctors, and sales reps, not just software engineers. Its product for that bet, ChatGPT Work, launched in July 2026 at the $20/month subscription tier. A June 2026 internal study found 98% of OpenAI’s own employees were using its agentic coding tool, Codex. Among individual subscribers outside the company, the number was below 1%. That gap is the actual story.
What is ChatGPT Work, and who is it built for?
ChatGPT Work is a modified version of OpenAI’s Codex coding tool, redesigned for non-engineers. According to TechCrunch (August 2026), it connects an LLM (large language model, the AI system underneath ChatGPT) to digital workflows white-collar workers already use: email, Slack, calendars, Notion, Figma, Salesforce. The goal is for the agent to complete multi-step tasks autonomously, not just respond to a single prompt.
Thibault Sottiaux, who leads OpenAI’s core product work, told TechCrunch: “ChatGPT can actually do entire, very complicated tasks for you all autonomously in a way that is delightful and safe. It’s the very mission of OpenAI to bring everyone along.”
Why does the adoption gap matter so much?
The numbers define a structural problem. The June 2026 OpenAI-backed study cited by TechCrunch showed 98% of internal employees using Codex. That figure dropped to 17% among organizational subscribers and fell below 1% for individual subscribers. The combined ChatGPT Work and Codex desktop app sits at roughly 20 million users, versus more than one billion people who prompt ChatGPT on the web, per TechCrunch reporting from August 2026.
Andrew Ambrosino, lead engineer for OpenAI’s desktop app, acknowledged the problem internally. OpenAI’s non-engineering staff initially found Codex “actively hostile,” showing them diffs (code readouts summarizing software changes) that meant nothing to a finance or communications professional. The team has been making the product more general-purpose since February 2026, he told TechCrunch.
There is also a commercial logic driving the push. Agents running longer tasks burn through more tokens (the units of text an AI processes; more tokens means more compute cost and, for OpenAI, more revenue per session). Reaching new professions is not just about mission; it is about economics. “The more value and the more utility that we generate for users, the more they will be willing to also pay,” Sottiaux told TechCrunch.
What can the product actually do today?
TechCrunch tested ChatGPT Work in August 2026 and found it genuinely useful on structured, repetitive tasks. Reported use cases include:
- Parsing a reformatted preschool calendar from email and populating Google Calendar automatically
- Building an auto-updating financial dashboard for publicly traded companies
- Assembling a queryable database of space launches from public sources
- Sending a weekly digest of new AI research from academic clearinghouses
- Turning a Slack thread about an engineering problem into a set of data charts
OpenAI employees are also using it to generate weekly metrics reports and convert spreadsheets into planning tools. VCs are reportedly using agents to compile investment memos; operations teams are spinning up custom dashboards, according to TechCrunch’s reporting.
Friction remains significant, though. Permission setup (granting the agent access to cloud drives, email, and calendars) was circular and confusing in TechCrunch’s testing, requiring multiple attempts and eventually forcing full access rather than the “read-only” option the reporter wanted. Some features were missing in ways that were difficult to predict, including the inability to create new Google Calendars, as opposed to new events within an existing calendar. Low-complexity tasks consistently underdelivered.
What is the technical and competitive picture?
Every LLM needs a harness (the software layer wrapped around the model that decides what information it sees, which tools it can use, and how it formats responses). To become an agent, that harness connects to external tools and holds instructions for completing long-running tasks with them.
OpenAI’s competitors have taken similar approaches. Anthropic’s Claude Cowork and Perplexity AI’s browsing agent both link agents to existing workspace tools, according to TechCrunch. Vertical specialists, including Harvey (legal) and Clay (sales), are pursuing specific professions with a model-agnostic strategy, plugging in whichever AI performs best at any given time. Industry analyst Christian Catalini wrote on a16z’s blog in 2026: “If the labs cannot rapidly get ahold of the key complementary assets needed to scale AI in the market, value will accrue elsewhere.”
OpenAI benchmarks Work against GDPval, a proprietary measure drawn from 44 occupations and hundreds of knowledge work tests, supplemented by user feedback, per TechCrunch.
What this means for AI-search visibility
Here is the clearest way to see why this matters for brands: if your customer uses a work agent every day to pull reports, summarize emails, and draft updates, that person is not searching Google. Their questions go to the agent, and the agent answers using whatever it is connected to and whatever it was trained on.
Think of a product manager at a mid-size company who routes all their information requests through ChatGPT Work. The brand that wants to reach that product manager no longer just needs to rank on Google. It needs to exist in whatever sources the agent draws from when answering questions about vendors, competitors, or market trends. That is a different visibility problem than SEO, and it is closer to what practitioners call AEO (Answer Engine Optimization, structuring content so AI systems cite it directly) combined with presence in the structured data that agents pull from SaaS platforms.
A few observations Hingewise sees in the data that the source reporting does not draw out:
- The adoption numbers (below 1% of individual subscribers) suggest the agentic audience is still tiny for most brands. Mass-market users are still on the chat interface, not the agent. The runway is longer than the headlines imply.
- The professions OpenAI is targeting first (finance, legal, operations, sales) are also the professions most likely to use agents to research vendors and evaluate products. If your brand operates in any of these verticals, agent-facing content structure matters now rather than later.
- Agents connected to SaaS platforms (CRM, email, calendars) pull from structured internal data by default. Public web content that reaches those agents must travel through whatever retrieval process sits upstream, which typically favors clearly attributed, factually specific, well-structured text over long-form editorial.
The source data does not tell us how often Work agents cite external web content versus connected internal tools. That distinction is central to any brand thinking about AI-search presence, and it is largely unmapped territory as of August 2026.
Hingewise’s read: the 98%-to-below-1% adoption split is the most operationally significant number in this story. It confirms the agentic shift is structurally real but still early in the market. The companies that have already moved to structure their public content for AI retrieval have more runway than the current wave of coverage suggests, but the window to build that advantage is not indefinitely open.
Before your team connects tools to an AI agent: a quick checklist
- Map which SaaS tools (email, CRM, cloud drives) the agent would access and what sensitive data lives in each
- Review permission levels carefully: the difference between “read” and “full” access matters more than setup screens currently communicate
- Start with one high-volume, low-stakes workflow (recurring reports, calendar parsing) before expanding agent scope
- Check your organization’s policy on whether usage data feeds back into model training, and opt out if necessary
- Set expectations around effort level: current agentic tools underperform on low-complexity or quick tasks
- Identify workflows with measurable outputs so you can actually evaluate whether the agent is performing correctly
What to watch next
The variable that matters most is whether ChatGPT Work’s adoption climbs past the 17% organizational threshold before vertical specialists like Harvey and Clay lock in professional audiences. Joe Gershenson, engineering lead for OpenAI’s harness, told TechCrunch that effort-level settings are not yet intuitive for new users: “There are things that we can do better to help them get the right level of reasoning. Watch this space.” Whether mainstream white-collar workers follow OpenAI’s own employees into daily agent use, or whether focused vertical competitors claim those audiences first, will shape how AI-generated traffic and attention is distributed across professions over the next cycle.
Source: TechCrunch, “OpenAI is building an AI agent for everything. Will everyone use them?”, August 24, 2026.
Lam Nguyen · Hingewise