ChatGPT Work ships apps on its own. Here's the part it still hands back to you

What did OpenAI actually ship?

On July 9, OpenAI released ChatGPT Work: an agent that runs for hours without you. You give it an outcome — "pull last month's numbers from Drive and build the board deck" — and it decomposes the job, connects to your stack (Google Drive, Slack, Salesforce, Gmail, Teams) and comes back with the finished artifact. Docs, sheets, slides, websites, working web apps. It runs on GPT-5.6 and folds in Codex for writing and reviewing code.

It doesn't suggest. It delivers.

Sam Altman, OpenAI's CEO, framed the moment it lands in: "Every enterprise now is thinking about spend and the value they're getting in exchange for AI" (via Forbes). Right question. It cuts both ways — including at whatever the agent hands you.

One honest technical note, because the hype here is thick: frontier models are not interchangeable and they are not equally good at the same things. On independent complex-coding benchmarks GPT-5.6 trails other frontier models, and developers who pushed it hard — Simon Willison among them — didn't find it better than the competition on genuinely difficult work. Remarkable, yes. Uniform, no.

So can you cut the dev team?

Part of it, yes. Some of that work never should have cost what it cost.

The trap is assuming all of your software lives in that bucket.

Which layer can you hand to an agent?

The useful question isn't "AI or humans." It's what happens when the thing breaks on a Tuesday at 3pm:

Notice the table never asks how hard the code is. An agent will build you a beautiful member portal this afternoon. It asks who eats the error. That's the axis product launches never mention.

Why the code was never the expensive part

Our own example, and it comes with an uncomfortable ending.

Kynoz built a membership system from scratch for a yoga studio: online purchase and renewal, class-pack sales, gated access for active members only, email verification, automatic expired-to-renewal flow. Delivered complete in about two weeks. The studio's founder — previously a country manager at one of the largest technology companies in the world — told us his own teams back there would have quoted that project at three to six months.

But speed of typing isn't what made it work. Before the first line of code we researched the end user: age range, where they came from, socioeconomic profile. A large share of the members turned out to be older women with little comfort around technology. That finding rewrote the whole design — every screen built so that getting it wrong was nearly impossible. Research, analyze, map, validate the logic and the usability. Code last.

An AI agent would have out-typed us on that project. What it would not have done is ask a single person who was going to use it. It would have shipped something immaculate, modern, with an elegant login — and a 68-year-old stuck on step three, who doesn't renew online, who calls instead. At which point your automated system is just another employee.

Now the uncomfortable part: the studio closed. Not because of the system — the system and the site worked — but because its marketing strategy, contracted with another firm before we ever met, never brought customers in. The business never reached its audience. It's the lesson that stuck with us hardest: good software doesn't save a business without a strategy. It's why Kynoz doesn't sell code by the yard — technology, process and marketing ship together or they don't ship. If typing is now free, the question "who is this for, and how will it reach them?" didn't get cheaper. It got more expensive.

Isn't this just what a vendor would say?

Fair suspicion. You're hearing it from the people who get paid to decide what to build. So hear it from the other side instead.

On July 15 — six days after ChatGPT Work shipped — Anthropic, the company that builds the Claude models, launched Ode together with Blackstone, Hellman & Friedman and Goldman Sachs, valued at $1.5 billion. Its business isn't models. It's putting AI inside companies that aren't tech companies. They bought an applied-engineering firm and staffed it for exactly that.

Read it again: the people who manufacture the AI just bet a billion and a half that the AI isn't the hard part. Eddie Siegel, Ode's chief technologist, put it plainly: "model selection matters, but it's not where the majority of calories are spent" (via TechCrunch). His co-founder Chris Taylor, the CEO, named the part that lands on you: doing it right takes top-caliber applied AI talent, "which is not something most companies have."

That's the whole thesis, funded — and it's also the nearshore argument in one line. If the scarce input is applied talent that sits close to your business, the question stops being who types cheapest and becomes who you can actually get in the room.

Don't you use AI yourselves?

Every day. Pretending otherwise would be silly, and worse, dishonest.

We write code with AI assistants, we automate with agents, and we've spent months putting AI inside our clients' processes. The difference was never whether you use the tool. It's who answers when it breaks. When a company whose actual business is artificial intelligence — it processes satisfaction surveys for large-scale retail chains — needed their website and the application their clients use, they didn't prompt an agent for it. They hired us. An AI company knows exactly where the part AI does alone ends.

What this means if you're buying nearshore

It sharpens what you should be buying, and it doesn't touch the reasons you came to Mexico in the first place.

Mexico's IT sector holds roughly 700,000 professionals, per AMITI (via Expansión, 2025) — and 62% of US companies are weighing a move of production to Mexico (Kearney FDI index via FreightWaves, 2026). None of that was ever about who types faster. It's about a team in your time zone — we're in Puebla, on US Central — that picks up the phone at 9am your time, that you can fly to in a couple of hours, and that operates inside the USMCA framework, now under annual review, where documented process is what survives a rule change.

So the honest version: if a nearshore vendor's pitch is cheap hands on keyboards, ChatGPT Work is coming for that pitch, and it should. What doesn't commoditize is the team that shows up on your Tuesday morning call, tells you the feature you asked for is a bad idea, and can explain why in your language and your regulatory context.

Use the agent for the whole disposable layer. Guilt-free, starting today. For the layer that touches your money, your customers or your auditors, buy the thing that got scarcer, not the thing that got free.

AI has learned to type better than we do. It still hasn't learned to ask you who's going to use this. That's still somebody's job.

FAQ

What is ChatGPT Work?

Launched July 9, 2026, it's an agent rather than a chatbot: it connects to your apps (Drive, Slack, Gmail, Salesforce), works independently for hours, and returns finished deliverables — documents, spreadsheets, presentations and working web apps. It runs on GPT-5.6 with Codex integrated for writing and reviewing code.

Does this replace a nearshore development team?

It replaces the part of the work that was pure execution — and that part was always the easiest to buy anywhere. What it doesn't replace is user research, process design, architecture decisions, regulatory judgment and accountability when something breaks. If that's not what your vendor is selling you, the agent is the smaller of your problems.

Can we let it touch our production systems?

Only with the boring controls in place: a distinct identity for the agent, permissions scoped to what it actually needs, and an audit log of what it did. An agent with blanket access to your Drive is the same risk as an employee with every key in the building — one that works faster and at 3am.

Why keep a team in Mexico if AI writes the code?

Time zone overlap (Puebla runs on US Central), a two-hour flight, USMCA-aligned contracting, and people who can sit in your Tuesday call. Those were never code-typing advantages, so nothing about ChatGPT Work erodes them.

What should we do with software an agent already built for us?

Audit three things before it fails on you: what happens to customer data, whether it integrates with your existing billing and inventory, and who maintains it when it breaks. Three "not sure" answers means rebuild now — that's far cheaper than a rescue after an incident.