Balaji Sreenivasan is the founder and CEO of Aurigo Software.
Every infrastructure organization has experienced people who know far more about how work actually gets done than any procedure or manual can realistically capture. They know the trigger points for a second review, what evidence a regulator or auditor might expect and which sequence of steps to follow to keep a project moving.
That knowledge has been built over decades, forming a “shadow rulebook” that keeps things running smoothly. When the “authors” of that shadow rulebook leave the organization, a lot of operational knowledge leaves with them.
According to Deloitte’s “2026 Engineering and Construction Industry Outlook” report, 41% of construction workers are expected to retire by 2031, while only 10% of the current workforce is under the age of 25. That means a lot of operational knowledge and “ways of doing things” is about to walk out the door.
This is especially poignant because capital projects have a reputation for being difficult to automate. They’re heavily regulated, procedural, highly specialized and dependent on people who have spent decades learning how to deliver them.
However, those same characteristics can make them well-suited to AI. Whether an owner is building a highway, bridge, airport, power plant or data center, much of the work operates within defined requirements, approval paths, design standards and prescribed outputs. That gives AI something valuable to work with: codified procedure and repeatable logic.
The primary obstacle here isn’t that AI isn’t capable enough but that the “formal” rulebook only tells half of the story. Every infrastructure owner develops their own way of applying those rules through interpretation and context, and while some of that practice is documented, much of it isn’t.
A September 2025 global RICS survey found that 45% of surveyed construction organizations weren’t using AI at all, while only 1% had scaled it across projects. Yet nearly 70% of project managers and quantity surveyors believed the technology could help them deliver greater value. The key is figuring out how.
In my experience, that gap between interest and adoption comes down to more than the technology itself. Capital owners are often working across fragmented data, legacy systems, security and privacy requirements and established workflows that are difficult to change. Many also lack the internal skills to deploy and govern AI at scale, while experienced practitioners are understandably wary of tools that don’t reflect how their organization actually works.
Successful adoption requires AI to be connected to trusted data, integrated into existing processes and grounded in the institutional knowledge that sits behind the formal rulebook.
Capital delivery is more deterministic than it looks.
Capital programs are indeed complex, but they’re not chaotic or random. Many of the workflows involved have well-defined inputs, steps, requirements and outputs. Contracts, inspections, payment processing and compliance processes are far from open-ended exercises.
This is precisely where AI agents can shine. Agentic AI can use context to determine what information a process requires, retrieve the relevant records, compare them against a rule or historical action, identify what’s missing and prepare an output for review with supporting evidence.
Consider the National Environmental Policy Act (NEPA) process for federal aid highway projects. Although categorical exclusions and environmental reevaluations must satisfy established federal requirements, completing them still requires practitioners to gather project records, compare current conditions with prior documentation and determine whether anything has changed that affects the environmental decision.
A specialized AI agent can carry out the initial analysis, surface potential issues and assemble supporting evidence for an experienced practitioner to review. Across 500 projects, our NEPA agent is expected to save approximately 53,000 hours of work. The agent isn’t inventing environmental policy but is reasoning within an existing structure of rules, evidence requirements and historical activity to help users reach a defined outcome faster.
All of this may lead decision makers to ask, “If so much of this work is already structured, why not simply deploy a generic AI agent to get on with it?” Because the “official rulebook” is only the starting point.
There’s knowledge in the shadowlands.
Consider two organizations that follow the same regulation but take vastly different amounts of time to reach a conclusion.
One agency may require an additional review before advancing an environmental decision, while another may interpret a reevaluation trigger differently depending on how similar cases have been handled in the past. That experience tells practitioners what triggered reviews on similar projects, what a regulator expects to see, which exceptions warrant escalation, how particular risks are treated and what an acceptable submission looks like.
Very little of this exists in one clean procedure manual. Instead, it sits with program managers, engineers, planners, environmental specialists, contract administrators and project leaders who know not only which rule applies but how the organization applies it in practice.
It also explains why the next generation of AI for capital programs can’t simply be switched on out of the box. The regulation may be common, and the underlying model may be common, but the way each owner actually operates isn’t.
Turn the shadow rulebook into an asset.
The encouraging part is that undocumented knowledge isn’t unknowable. It’s simply undocumented. Capturing it requires going beyond policies and process diagrams and working alongside the people who actually perform the work.
This is why a forward-deployed engineering model is so important. Engineers working alongside customer teams observe how decisions are actually made, document the unwritten rules and precedents people rely on, translate them into instructions and evaluation criteria for agents and test those agents against real work. Those agents will operate within the system of record the organization uses. Project history, contracts, approvals, schedules, documents and prior decisions provide the context needed to understand the formal rule and how that organization applies it in practice.
Once that happens, knowledge that existed only in the heads of experienced people becomes a codified institutional asset. It can be repeated, taught, applied and enriched by agents and passed to the next generation of employees.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?