The commercial test for AI is shifting from model performance to whether a system can hold up inside a working business. OpenAI and Anthropic’s 2026 moves into enterprise deployment suggest that leading AI companies see implementation talent and access to customer workflows as increasingly valuable parts of the market.
In May, OpenAI launched a deployment company backed by more than $4 billion. Its engineers work inside customer organizations to connect AI with company data and operating processes, staying through implementation. Days earlier, Anthropic announced a similar venture aimed at mid-sized businesses that may lack the technical staff to carry out that work themselves.
The roles at the center of these efforts are known as forward-deployed engineers, a model associated with Palantir. Their job is to work alongside a customer’s team, understand how its operations function and adapt AI systems to real information and real tasks. The measure of success is practical: whether the system continues to work when it encounters routine orders, customer needs and exceptions.
Deployment is becoming part of the product
For years, the AI race centered on building a better model. Companies competed to make systems faster, more capable and less expensive. That work still matters, but the new deployment businesses point to a different commercial constraint: many buyers need help turning model capability into reliable operations.
The shift also changes the traditional software economics. A product built once and sold broadly can scale efficiently, but enterprise AI often has to fit a company’s data, workflows and decision rules. When the companies building leading AI businesses put engineers into customer operations, they are committing resources to work that is specific to each client. The bet is that successful implementation can create enough value—and customer knowledge—to justify that investment.
Across the spring announcements, billions of dollars are being directed toward deployment. That concentration signals a change in where AI providers expect value to accrue: not only when a model is released, but when it becomes useful in a customer’s daily work and keeps delivering results. The market is starting to treat implementation capacity as a competitive asset alongside model quality.
Why pilots struggle to become routine
The case for deployment has been strengthened by the gap between polished demonstrations and real operations. A pilot may use tidy data, a narrow task and a small group of cooperative users. Once a system enters a live business, it faces conflicting records, inconsistent processes and information that has not been maintained for automated use.
A customer may appear in several databases under slightly different names. Prices can be stale, and departments may keep separate versions of key records. An AI system can process those inputs quickly while still producing an unreliable result. The problem is often not model capability; it is that the organization has not agreed on which information is accurate or how the system should handle uncertainty.
Gartner predicts that through 2026, companies will abandon 60% of AI projects that are not built on data ready for AI—accurate, current information that is agreed across the organization. In the same research, 63% of companies said they either do not manage data in a way suited to AI or are unsure whether they do. Those figures point to an operational weakness that can limit returns even when the underlying technology performs as expected.
Gartner has also warned that more than 40% of AI agent projects could be canceled by the end of 2027 because of rising costs, unclear business value and weak controls. These risks sit around the technology: the quality of the data, the design of the workflow and the boundaries on what an AI system can do without a person’s approval. Deployment teams are being asked to address those conditions as part of getting AI into production.
For companies, that can mean cleaning and reconciling records, defining what a good output looks like, and setting clear escalation points. It may also mean redesigning a process rather than inserting AI into an existing one unchanged. A forward-deployed engineer can help uncover those requirements, though the organization still has to own its data and decide how its work should run.
Customer knowledge may shape the next sale
The commercial implications extend beyond the initial implementation. Engineers who spend time inside a business can learn where delays occur, which errors carry the greatest cost and which tasks consume staff time. That operational understanding may inform later product decisions and help providers identify where customers are willing to pay for automation.
This helps explain why AI companies are moving quickly into deployment. A provider that understands a customer’s actual work may be better positioned to tailor its systems and keep the relationship as the customer’s needs change. At the same time, this approach can be labor-intensive. The economics will depend on how much work can be repeated across customers and how much requires bespoke engineering.
Smaller businesses may also benefit from services designed for organizations without large in-house technical teams. For founders, the opportunity is not limited to creating new models. Companies that make AI work for a particular industry or workflow could build useful businesses by solving implementation problems that general-purpose tools leave to customers.
Readiness will determine how far adoption spreads
Gartner expects that by the end of 2026, 40% of enterprise applications will include task-specific AI agents, up from less than 5% in 2025. It also forecasts that by 2028, at least 15% of everyday work decisions will be made autonomously by AI. These are forecasts, not settled outcomes, but they indicate how quickly software vendors and buyers are testing task-level automation.
As AI features become more common in business software, choosing a model may become a less distinctive decision for many buyers. The harder questions will concern readiness: Is the information dependable? Can teams describe a successful result? Which actions can an agent take on its own, and where should human judgment remain in the process?
Those questions connect AI spending to basic operating discipline. Companies that can answer them are more likely to move beyond a demonstration and assess the effect on costs, speed and service. Those that cannot may find that adding an agent increases complexity before it improves performance.
New models will continue to attract attention, and improvements in capability and cost will shape the market. But the next phase of competition looks likely to turn on whether businesses can prepare their data, adapt their workflows and set sensible decision boundaries. The clearest measure of progress will be what AI can reliably do in ordinary operations—and whether the resulting value justifies the effort required to get it there.
