They Are Coming to Deploy Us

On July 2, Microsoft announced the Frontier Company, a new unit with $2.5 billion behind it and 6,000 engineers whose mission is to install themselves inside client companies to build and operate their artificial-intelligence systems — ensuring, in the process, that the AI Microsoft deploys actually takes root in the business.

Two days earlier, Amazon Web Services had announced $1 billion for something similar. In May, OpenAI created a deployment company with more than $4 billion in capital from a consortium led by TPG. Anthropic partnered with Blackstone, Hellman & Friedman, and Goldman Sachs on a $1.5 billion vehicle to embed engineers inside midsize companies.

Four giants arrived at the same conclusion in the same quarter: The way traditional consulting has been approaching the problems of AI is inadequate.

A study from MIT’s Project NANDA found that 95 percent of generative-AI pilots in the enterprise produce no measurable impact on the bottom line. The models have been improving for two years, and the number has not moved. The failure lies in integration with the actual processes of the business.

The consulting firms understood the moment before many of their clients did. In February, OpenAI launched the Frontier Alliances, multiyear partnerships with McKinsey, BCG, Accenture, and Capgemini. Each firm is building teams certified in OpenAI’s technology, in order to sell it and implement it. McKinsey says so plainly in its own press release: Its deployed teams will create a feedback loop between client projects and OpenAI’s product development.

This reduces the consultant in your office to something else as well — a certified distribution channel for the model vendor — while your transformation project feeds that vendor’s product. As a bit of color: Microsoft’s new unit and OpenAI’s platform share the same name, Frontier. The industry does not even bother to differentiate its conquests.

Here, dear reader, is where the story becomes ours. Mexico and Latin America are the perfect market for this offensive, and not for the reasons we would like.

This wave of neo-consultancies is breaking over a corporate structure that spent 20 years learning to outsource its technological muscle. The technology departments of many large companies hollowed out from within until they became administrators of contracts. The system is operated by a third party; the development is done by a software factory. The IT department gathers requirements and signs annexes. This is true from the government to private equity.

That structure worked, more or less, when technology was a commodity. With artificial intelligence, it is not a model that will produce a good implementation inside the company.

Everything of value in an AI system is built through use. There are the fine-tuned prompts and the evaluations that measure whether the system works. There is the record of what failed, and why. That knowledge is the company’s learning curve. When the project is executed by an outsider and no one absorbs it internally, the learning curve walks out the door when the contract ends.

Satya Nadella wrote it in a recent essay: The models are becoming interchangeable, so the durable asset is ownership of the deployment relationship. Whoever is standing in your building when the AI finally works keeps the access to the data and the institutional memory of what it cost to get there.

Closing the door on the consultant would be naive. What must change is what gets contracted and what gets retained. Three concrete rules.

First, the assets of the pipeline belong to the company. Prompts, evaluations, test datasets, and the backlog of use cases must live in the company’s own repositories, under contracts that say so in plain language. We have to make certain that all of this development remains in the company’s hands.

Second, every external project needs an internal counterpart that absorbs rather than receives. One business expert and one engineer per workflow, with a mandate to keep the system when the outsider leaves. Palantir invented this model 15 years ago, and today everyone copies it because it works.

Third, do the full math. In many cases, an internal team of 10 people costs less than a year of a large contract with a global firm. And unlike the contract, that team stays — and accumulates the context of the business with every mistake.

The new era of consulting will seek to ride AI. We have to make certain that the value stays inside the company, and not in the service of outsiders.