This essay was inspired by a generous conversation with Barry McCardle, cofounder of Hex and author of Understanding Forward Deployed Engineering.
—
In the past, a company’s surface area was limited by their pool of intelligence.
Each new product or feature incurred heavy costs in programming, product, design, marketing, sales, and support. A company could pull a few tricks to decrease these costs, such as buying third-party software or hiring contractors. But for the most part, the only way out was to increase the intelligence pool. In other words, they had to hire more people to do these jobs. And until a few years ago, each additional wallop of brainpower came with price tag of at least $100k per year.
Then came GPT-5 and Opus-4 AI models. Where previous large language models were mostly confined to question-answer chat interfaces, these new models could perform long-horizon work when paired with agent harnesses. Better, they were smarter than the average person and cost less than a single latte.
It’s little wonder that most companies are now experimenting with how to add AI to their workforce.
Unfortunately, most companies are not AI-shaped. In the same way that Oracle’s databases weren’t useful to companies that weren’t willing to digitize their paper records and train their staff to write SQL queries, AI models and agents can’t just show up to work and start executing. They can rewrite an email or create a quick demo with minimal human direction, but they can’t figure out which customers are about to churn, call them up, chat about fishing, and miraculously save the account, Michael Scott style.
Even the most AI-native companies haven’t fully figured out the structure of AI deployment within an organization.
Today, most efforts are individual-driven with outward ripples. Someone cobbles together a couple of different tools and becomes more effective at their job. The increase in outputs hits others at the company, who learn how to build their own AI systems to increase productivity. But as the ripple expands outward, the momentum of AI transformation begins to dissipate. Then it meets a process that’s resistant to AI acceleration and work begins to pile up in front of the bottleneck. In the end, the company’s intelligence pool only marginally expands.
So while AI has empowered individuals to make a massive impact, it hasn’t yet consistently created end-to-end organization-level outcomes.
To solve this, we’ve turned to application AI solutions and forward-deployed AI services.
Application AI
AI point solutions are thriving.
Application AI companies like Sierra and Harvey sport multi-billion dollar valuations after finding product-market fit in self-contained enterprise functions. Whether it’s customer support or document research, there’s a clear success criteria and task execution largely doesn’t depend on other functions within the company.
So these companies have taken frontier models and built two layers of infrastructure around them. The first is harnesses and evals that stay up to date as models improve, change, or drift. Then, they’ve built products around ingesting raw information and connecting to customer systems of record. In doing so, they’ve wrangled a general purpose AI model into a highly focused AI agent or solution that can operate independently without dipping into a customer’s intelligence pool.
As AI models become more powerful, we can imagine two futures for application AI companies.
The first is one where AI integration into new functions is the bottleneck. To integrate each new AI solution, a company has to expend a non-trivial amount of their intelligence pool for onboarding. This cost becomes too high to bear and companies instead choose to expand scope with existing vendors. With more context, the AI solution can perform better, which then puts the vendor in a better position to win more business.
For companies like Sierra, they’ve already begun the journey from support to analytics to even marketing. In this future, we’ll see companies working with a few large applied AI providers that have all-in-one suites of distinct solutions.
The second future is one where execution and having the right evals becomes the point of differentiation. AI models misbehave in ways that are hard to detect and fix. As models improve, the specific domain expertise required to build the correct AI product in each function will deepen, which means that it’ll be unlikely for a single vendor to be the best in multiple different functions.
On top of that, the value of a better solution can be exponential. In coding agents, the frontier has shifted from Cursor to Claude Code to Codex in a matter of months. Companies have followed this transition because the value gained from even a marginally better agent has outweighed the pain of switching providers. In this future, we’ll see companies buying dozens to hundreds of best-of-breed point solutions.
The overlap between both futures is that application AI solutions will carve out more and more functions from a company’s surface area. In doing so, a company can better leverage its intelligence pool. Even though the pool itself didn’t increase, the same company can now support more and larger companies without a corresponding increase in headcount.
Forward-Deployed AI
Every company seeks to control its own destiny.
The current model of application AI companies is to have a development team that sits in San Francisco or New York shipping product. Then, they send out a deployment engineer or team to a customer’s office to make sure the first few agents are correctly configured. While effective, this limits AI to the products being built by the central development team.
No two companies look alike. Even within the same industry and solving the same problem, two companies can have radically different processes. So far, applied AI companies have found success in relatively homogeneous problem spaces. As AI extends into operations, marketing, and mission-critical functions, companies will very likely want to control and manage their own internal iteration of the technology.
Two decades ago, Palantir coined the term “Forward-Deployed Engineer.” They would hire strong software engineers who normally joined large tech companies and send them straight to enterprise customers who normally couldn’t recruit that caliber of talent. More importantly, these engineers didn’t try to force a specific product that had been built from a central office. They acted as consultants dedicated to tackling a specific domain of problems.
For Palantir, that domain was data. Government agencies had great data collection and retention. What they were lacking was a way to bring all those different pieces of data together into something actionable. In fact, this problem extended beyond governments. The decreasing cost of storage meant that companies all around the world were logging more and more data without a proper motion to analyze or integrate those pieces of data.
We currently sit at a strange but similar point on the AI adoption curve. Most companies have both an abundance and drought of AI at the same time. Because of how quickly the market has developed, companies often have enterprise plans with large model companies like OpenAI or Anthropic, a plethora of point solutions, and even more pilot programs for internal agents or AI products.
Much of the bottleneck of AI is in documenting current processes, building the correct eval and guardrails into an AI harness, and deploying the model or agent with the full observability.
This can come in different forms. There’s visibility where companies can lose track of their active agents and their trajectories. As models become more expensive, these agents can incur non-trivial levels of cost each time they stray. There’s also access where agents have varying levels of access to internal data or execution capabilities. With the non-deterministic nature of AI agents and potential security vulnerabilities, this can become a larger liability. And finally, there’s syncing where model or tool behavior can drift and decouple from the actual work that needs to be done.
Like with Palantir and data, there is a distinct lack of talent in the world who can bridge the gap between AI expertise and domain-specific know-how.
OpenAI has partnered with BCG, McKinsey, Accenture, and Cap Gemini to implement, govern, and scale custom AI agents for enterprise customers. They’ve also created The OpenAI Deployment Company with partner private equity firms. On the other side, Anthropic also partnered with Blackstone, Hellman & Friedman, and Goldman Sachs to start a new AI services company.
However, I believe the future for forward-deployed AI will be slightly different. Palantir’s magic was in going to a client, partnering with them, and building something new for the sake of IP that might potentially be discovered. Instead of coming in with assumptions around what model to use or how to build the best harness, the most successful forward-deployed organizations will be in service of their customers.
Closing Thoughts
The core primitives of cloud, like S3 and EC2, were built out around 2010. And then it took a few more years for companies to learn how to apply the new technology to their business.
AI is in a similar state. We currently have technology that can turn electricity and silicon into machines that pass the Turing Test. It’s now cheaper and quicker than ever to “hire” intelligence. Yet, the lack of organization-level impact has shown that raw intelligence alone isn’t enough.
We’re currently in the messy middle as we fumble around how to best deploy AI into the world. However, both applied AI solutions and forward-deployed AI services are maturing. Application solutions will begin to carve out more and more neatly defined functions, while forward-deployed services will better interweave models into the messy, proprietary core of a company.
And, in the end, we’ll have found a way to permanently increase a company intelligence pool.