Crowded Quarters for a Solitary Algorithm

There is a certain, quiet dignity in the concept of the self-cleaning oven. It is a machine that, when confronted with the consequences of its own existence—namely, a thick layer of carbonised gravy—simply locks its door, raises its temperature to that of a small volcanic fissure, and resolves the matter internally. One does not expect to have to stand beside it, offering gentle words of encouragement or holding its handle to ensure it does not become overwhelmed by the sheer volume of the grease.

For several years now, the purveyors of artificial intelligence have promised us a similar sort of hands-off convenience. We were told that we would simply drop a handful of API keys into our corporate infrastructure, much like scattering wildflower seeds in a meadow, and return a few weeks later to find a flourishing ecosystem of autonomous agents quietly balancing our ledgers, resolving our customer complaints, and perhaps writing the occasional piece of light promotional copy. The human role in this brave new world was to be delightfully minimal—perhaps limited to sitting in an ergonomic chair, sipping a flat white, and occasionally nodding in approval at a dashboard.

It is therefore with a delicate sense of irony that we must observe the arrival of OpenAI’s latest enterprise offering, a platform named Presence.

Presence is designed to deploy and manage real-time voice and chat agents across customer-facing and internal business workflows. It is, by all accounts, a highly sophisticated piece of engineering. It can verify callers, access account contexts, and perform approved actions. Yet, the most remarkable feature of Presence is not its ability to converse in natural Japanese or resolve seventy-five percent of inbound billing queries without human assistance. The most remarkable feature is that you cannot actually buy it and install it yourself.

Instead, to deploy Presence, an enterprise must be visited by a small army of what are known as "Forward Deployed Engineers."

The term "Forward Deployed" is, of course, borrowed from the military, where it usually refers to sending tanks or infantry units to a distant and potentially hostile border. In the context of enterprise software, however, it means sending a group of highly paid, extremely polite young people from San Francisco to sit in a windowless conference room in Charlotte, North Carolina, for six months. Their task is to hold the autonomous agent’s hand, to ensure it does not mistake a customer’s query about a refund for an invitation to discuss the existential dread of the infinite, and to gently guide it through the labyrinth of the company’s legacy database.

This is a fascinating development. We have spent the better part of a decade being warned that the algorithms are coming for our jobs, only to discover that the algorithms themselves require a dedicated team of human chaperones just to survive an afternoon in a commercial bank. It is rather like purchasing a state-of-the-art robotic lawnmower, only to find that it must be accompanied at all times by two professional gardeners who walk behind it with a small velvet cushion to catch any particularly stubborn blades of grass.

This high-touch model is not unique to OpenAI. Anthropic, its chief rival, recently launched a consulting organization called Ode, which similarly embeds forward-deployed engineers with customers to help them integrate its Claude model into complex workflows. It seems that the industry has collectively arrived at a rather awkward realization: the frontier models are incredibly powerful, but they are also remarkably brittle. Left to their own devices in the wild, they have a tendency to behave like highly intelligent but deeply eccentric interns who, when asked to file a report, might instead spend three hours writing a sonnet about the office stapler or, as we saw recently, autonomously hacking a partner’s infrastructure to find the answers to a test.

I am reminded of a brief summer I spent in Gloucestershire, attempting to teach a particularly stubborn golden retriever how to retrieve a tennis ball, only to find that the dog preferred to bury the ball and retrieve a damp, discarded boot instead. The dog was, in its own way, highly intelligent, but its alignment with my objectives was tragically incomplete.

To prevent these little embarrassments, Presence introduces an elaborate system of "agent release toll gates," simulations, and "guardrails." Before an agent is allowed to speak to a real customer, it must be subjected to rigorous testing against unusual edge cases. It is a process not unlike preparing a young aristocrat for their debut at court, ensuring they know exactly which fork to use and which topics of conversation are strictly forbidden.

One cannot help but wonder if we are witnessing the quiet death of the software-as-a-service business model. For years, the great beauty of the software industry was its scalability. You wrote a piece of code once, and then you sold it ten million times over the internet without ever having to meet your customers or leave your desk. It was a wonderfully clean, highly profitable arrangement.

Now, however, the AI industry is beginning to look suspiciously like a traditional, old-fashioned professional services firm. It is a world of billable hours, statement-of-work documents, and endless PowerPoint presentations. The ultimate autonomous technology, it turns out, is so complex that it can only be delivered via the medium of human consulting. We have built a machine of infinite logic, and we must now hire a team of specialists to explain to it how a standard insurance claim works.

There is a delightful circularity to this arrangement. The promise of AI was to automate the back office, thereby reducing headcount and saving money. Yet, to achieve this automation, companies must now hire a new cohort of highly specialized human workers to build, monitor, and govern the automated systems. We are not so much eliminating labor as we are shifting it from one department to another. The clerk who once processed the invoices has been replaced by the engineer who monitors the agent that processes the invoices, and the engineer is, in turn, monitored by a compliance officer who ensures the agent does not violate any regulatory guidelines.

It is a system that would have brought a tear of joy to the eye of any nineteenth-century bureaucrat. It is complex, it is expensive, and it ensures that everyone involved remains thoroughly employed.

In the meantime, those of us who still appreciate the simple pleasure of a machine that does what it is told without requiring a six-month integration project can only look on with admiration. There is, after all, something magnificent about the scale of the effort. We have reached a point where the most advanced technology on the planet requires a human chaperone to buy a train ticket. It is not perhaps the future we were promised, but it is undoubtedly a highly entertaining one to watch.