Skip to main content
All posts
|8 min read

What OpenAI and Anthropic Think Happens Next — and What It Means If You Don't Run a Tech Company

Sam Altman and Dario Amodei have both told you, in their own words, where they think AI goes next. But the clearer signal is what their companies built this spring: deployment firms that put engineers inside ordinary businesses to redesign the work. Here's how to read the labs' near-future vision if you run a brokerage, a clinic, or a contracting business — not a tech company.

Agentic AIStrategyFuture of Work

If you want to know where AI is heading, the people building it have been unusually direct about it. The trick is reading both what they say and what they do — because this spring, the two lined up.

The Founders Are Telling You The Timeline

In his essay "The Gentle Singularity," Sam Altman wrote that "2025 has seen the arrival of agents that can do real cognitive work; writing computer code will never be the same." He laid out a rough sequence from there: 2026 as the year of systems that generate novel insights, 2027 as the year AI moves into the physical world through robotics.

Dario Amodei, in "Machines of Loving Grace," made the case from the other direction — that "most people are underestimating just how radical the upside of AI could be." Two CEOs who compete on almost everything agree on the shape of the next few years: agents that do real work, arriving faster than the org charts around them can adapt.

You can take the timelines with a grain of salt — founders are professionally optimistic. But the direction is the signal, and they're pointing the same way.

But Watch What They Built, Not Just What They Said

The louder signal came in May, and it wasn't a model release. Within a week of each other, both labs built consulting firms.

Anthropic launched a dedicated enterprise AI services company with Blackstone, Hellman & Friedman, and Goldman Sachs, explaining that "enterprise demand for Claude is significantly outpacing any single delivery model." OpenAI stood up its own "Deployment Company," describing engineers who "work closely with business leaders, operators, and frontline teams to identify where AI can make the biggest impact, redesign organizational infrastructure and critical workflows around it."

Read that last line twice. The companies with the best models on earth just spent billions to put humans inside customer businesses to redesign the workflows. That is an admission, not just an expansion. If the model alone were enough, you wouldn't need to embed an engineer to rebuild the work around it.

The Quiet Admission: The Bottleneck Moved

Put the words and the actions together and the message is clear. The labs believe the next phase is agentic — AI that operates across your systems and does whole tasks, not just answers questions. Gartner, watching the same shift, predicts that at least 15% of day-to-day work decisions will be made autonomously by AI agents by 2028, up from zero in 2024.

But the labs also know — from their own deployment data — that getting there isn't a model problem anymore. It's an organisational one. That's why the same Gartner analysis warns that over 40% of agentic AI projects will be cancelled by 2027, undone by cost, unclear value, and weak controls. The frontier moved from "can the model do it" to "can the business absorb it." The deployment firms exist to close exactly that gap, for customers who can afford them.

What This Means If You Run A Brokerage, A Clinic, Or A Contracting Business

You are not going to hire a forward deployed engineer from OpenAI. You don't need to. But you should read their bet as a map, because the thing they're being paid to install is the same thing you can build at smaller scale.

Their whole model assumes the customer's business is agent-ready: workflows clear enough to hand off, context organised enough for an agent to use, and a human in place to supervise the result. Most businesses aren't there yet — which is the entire reason the consulting firms have customers. BCG found 90% of CEOs expect AI agents to deliver measurable returns in 2026. The ones who get there will be the ones whose operations were ready to receive an agent, not just whose budgets were ready to buy one.

Three things make you agent-ready without a frontier-lab contract:

  1. Document one workflow end to end. An agent can only take over work that's been made explicit. The thing the forward deployed engineer does first is write down how the work actually flows. You can do that yourself, on one workflow, this month.
  2. Put your context where an agent can reach it. The labs win by grounding the model in the customer's own data and systems. Your small-scale version: keep your pricing, your templates, and your standard answers in one organised place instead of scattered across inboxes and heads.
  3. Name the supervisor. Every serious agent deployment has a human who owns the output. Decide now who watches the agent, catches its mistakes, and signs off. Without that role, you're in the 40% that gets cancelled.

The Honest Read

OpenAI and Anthropic think the next two years belong to agents — AI that does the work, not just describes it. You don't have to share their optimism about the timeline to act on the structure of their bet. The most revealing thing they did this spring wasn't predict the future; it was spend billions telling you, through their actions, that the model was never going to be enough on its own.

The businesses that win the agent era won't be the ones with the best model. Everyone gets the same models. They'll be the ones whose work was organised enough to hand to one.

Sources

We help organisations build the context infrastructure, harness design, and skills architecture that make AI actually work in production. If this resonates, let's talk.

Back to all posts