Toward Enterprise Autonomy

AI will be a foundation for businesses, not a tool

AI will move information technology from the periphery to the center of value propositions of every successful institution in the world. IT will evolve from supporting business functions to being a cornerstone asset of every business. As this occurs, the compounding advantages of digitally-native enterprises will be available for real world business environments in every sector. This opportunity signifies a quantum leap in business efficiency, enabling businesses to perfect, and evolve, their operations in a continuous, AI enabled, manner. Enterprise Autonomy.

Seizing the potential requires reimagining the relationship between technology and business. This essay provides an initial framework for that new relationship, alongside tactical advice on how to begin to integrate AI into your strategic operations as a collaborative partner.

1 | Introduction

Enterprise Autonomy requires an AI powered operating system that defines the value proposition of a business not just supports it. A system like this has three foundational pillars:

• Decisions must be enabled by and captured in Software. AI will learn from what can be captured in a system. The most important data a business generates is it’s decisions, the rationale and their consequences. Capturing the information used to inform a decision, the decision itself, and the downstream outcomes is the fundamental building block for an applied AI system.

• Software that isn’t getting better is getting worse. The relationship between humans and enterprise software is extremely inefficient today. Humans absorb the clash between complexity of the real world and rigid boundaries of software systems. Humans are continuously encountering new situations and devising solutions. However that solutioning typically never makes it back into enterprise software systems. AI has changed this relationship. AI can learn from us in previous impossible ways that will make systems improve versus decay. Humans will be empowered to focus on more strategic and creative challenges to drive faster evolution of the system. We humans are still features not bugs.

• An operating system not a use case engine. AI systems can reintegrate operations that have been disintegrated by specific point solutions. An AI business requires well constructed circuits that can be integrated with other circuits to create a circuit board. Integrated value chains not isolated applications.

2 | The State of the Relationship between Technology and Business Today

Making a decision requires the ability to know the state of the world, to predict the impact of potential paths, and to execute against it. Omniscience, prescience and omnipotence are easier said than done. Each decision requires its own specific frame of reference, dynamically shifting for each decision, with data sourced from myriad systems and formats. Predictions require incorporating logic from simple algebra up to stochastic models and human intuition. Execution requires writing to systems of action in a governed, auditable, and secure manner. In order to have one integrated circuit, an application needs to accomplish each of these tasks. In order to develop a more ambitious circuit board, independent circuits need to integrate with each other. Managing this complexity is the challenge and the opportunity.

The opportunity is attainable. Integrated, intelligent operating systems have defined the explosive de facto monopolies of the consumer internet. Well executed consumer internet applications are designed for AI from the ground up. They build upon simple data models with full control over the inputs and outputs. The best executed products create systems where the data, intelligence, and product are indistinguishable from one another, and the results are explosive.

Outside of the consumer internet the reality is quite the opposite. Functions and industries have been created from whole cloth just to try to make incremental progress against each of the know/predict/execute missions in isolation from one another. Data and data management are their own own category of software. Analytics and data science is entirely another. Transactional and automation systems operate almost exclusively in isolation from the previous two. And perhaps most importantly humans work around and in spite of all of the above to actually make everything add up to something coherent. The result is no system linking execution and intelligence together, creating no foundation for AI to learn.

The first generations of cloud native enterprise software companies have made virtually no progress against this challenge, and in fact have likely regressed by creating ever smaller and more niche categories. The result has been a parasitic relationship between software companies and their end customers. The enterprise software ecosystem has rewarded innovation that has defined success narrowly and incrementally. Customers require bigger and bolder.

Why is this relationship broken?

Systems are designed to serve a business, but over time, businesses often find themselves serving the systems because the world changes faster than the software. As organizations grow, complexity naturally increases—more staff, more processes, and more localized automation from software. This software takes many forms. ERP (Enterprise Resource Planning), CRM (Customer Relationship Management), MES (Manufacturing Execution Systems) and many others are deployed to obtain local efficiencies from software automation. Yet, with each new system introduced, there's an unintended consequence: a subtle erosion of the business's ability to evolve. For every system added, there's a material loss in responsiveness for any process that relies on multiple systems. This sprawling ecosystem of systems works when things are very stable. Of course reality is anything but.

On the best of days, this federated complexity leads to inefficiencies in the 'business as usual' scenario. In order to make this system work and enable coordination at scale, orgs need to plan. Whole business processes and software categories to support them are constructed solely to hydrate systems and manage hand-offs between them. A significant proportion of our everyday work is preparing data for systems so they will run.

Resources are then allocated according to plans made in the past. These plans allocate resources using static logic representing the best way to do things, on average. It is too laborious to account for in-the-moment context from more than one system is too laborious to even consider. The result is resource allocation is virtually always sub-optimal. I regularly hear from customers running the most sophisticated supply chains in the world that planning is not intended to be optimal, it is intended to merely get to something feasible. Plants use dramatically too much energy because they rely on simple automation for when a blower turns on. Backorders and excess inventory occur simultaneously. Insurance companies underwrite the wrong policies while ignoring profitable ones. Best fit lines that underpin the ability of the myriad systems to execute hand offs do not reflect the diverse and dynamic nature of business operations.

The huge amount of effort also leads to familiar problems. There is variance in quality, headcount grows just to manage processing of data into systems, and operations run on the premise that all the assumptions made—often without a basis in real-time truth—are accurate because it is just too hard to check if they are.

On the worst of days, this complexity breeds fragility. A vivid example from 2023 is the operational meltdown experienced by Southwest Airlines, where a cascade of disruptions led to an unprecedented number of flight cancellations and delays. The systems, rigid and unable to adapt to the rapid changes in weather, staffing, and scheduling, became a liability rather than an asset. Organizations find themselves unable to coordinate at scale, evolve, or respond immediately to the incessant flux of the business environment. The provocative question with a near universal answer is how many enterprise systems enabled companies to respond to the massive disruption of Covid. The answer: None.

The crux of the issue is that businesses have become ensnared in their own web of systems, each adding a layer of complexity that obscures a clear view of operations. The original intent of these systems—to streamline and simplify—has been overshadowed by the weight of their maintenance and the inertia of their integration.

These challenges lead to myriad tactical use cases to apply technology towards. Low hanging fruit abounds. But this “use case” orientation also creates a “forest-through-the-trees” obfuscation that the point of new tech is to fix these tactical breakages. The fruit is so low hanging that we forget the ambition of where we could end up and stack band-aids on band-aids.

The challenge, then, is to reimagine how these systems can be restructured or replaced by AI-driven solutions that prioritize agility, learning, and a return to the fundamental purpose of the business: to transact effectively and efficiently in the service of its mission.

3 | Implementing an AI OS

This problem is tractable, even for the most complex organizations in the world. The requirements for a system to enable this are conceptually simple and anchored in the decision-making process—Know, Predict and Execute. These requirements have to be met for single decision and serve as a framework for the efficacy of any software application. Does the application provide perfect awareness to a human or model decision maker? Are expected consequences of a given action understood and are those assumptions continuously evaluated? And perhaps most importantly, are you actually doing something in the application?

The requirements also must be met for the entirety of the enterprise with an intentional approach to bringing together data, logic and action. We are defining this as an AI Mesh Architecture. The AI Mesh must bring together data, AI, and worfklow by design. At Palantir our bet is that the AI Mesh Architecture is the foundation from which the most efficient companies we have ever seen will be built.

We will be sharing our view on the technical requirements to achieve this in great detail over the coming weeks and months.

4 | Software that improves versus decays

Software systems that are not improving are getting worse. As the world changes and the code you wrote doesn’t, it’s ability to achieve it’s original goal breaks. Fortunately the latest AI systems can now learn in far more comprehensive ways than previously possible. This means the ability for software to keep up with the world has been forever altered. Software can move from creating increasing drag over time to creating increasing thrust. This is a very big deal.

AI systems will be able to enable the internal kaizen from the micro to the macro for every organization they power. At a minimum the system should enable continuous improvement in: repetitive decision making, horizontal collaboration, and strategic design.

Repetitive decision making: The simplest form of feedback to continuously evaluate decisions and their effects. Evaluating decisions and feeding back learnings in an effortless manner requires a thoughtful integration of data, logic and action to capture context and results.

Horizontal Collaboration: Complex operations are always managing competing objective functions, i.e. service quality and cost to service. Organizational preferences for how to manage those tradeoffs are often hidden in complex data or solved for at an extremely coarse level. An integrated system enables those tradeoffs to be quantified and a foundation for iteratively adapting strategy to balance tradeoffs.

Strategic Design: Your system should enable not only operating within constraints but suggest changing constraints. For example not only optimizing production of an oil rig given physical constraints, but suggesting CapEx projects to change the most valuable constraints.

This system-wide learning capability empowers organizations to continuously refine their operations and strategies, leading to a more agile and intelligent enterprise.

5 | Conclusion

The future is bright. AI is going to put technology back into a role of serving the world not extracting from it. This is going to provide a reacceleration in productivity with broad participation across industries -and not only digitally native ones. Looking forward to the revolution.