Most Companies Have AI Tools, But Many Don't Rewire Themselves For AI

Gaurav Singal is Chief Technology Officer at ConstructConnect, a Roper Technologies company.

getty

A 2025 McKinsey survey puts AI adoption at almost 90%, yet few are capturing meaningful business value from it. I've found that this gap is because most organizations are running AI on top of their existing operating model. They installed a tool, but they didn't "rewire" for it first.​

By mid-2025, more than 90% of our engineers at ConstructConnect were using AI coding tools, mostly Copilot or Cursor. Individual tasks were faster, but delivery speed as a team hadn't changed. Features still took the same number of weeks. Our cycle was still our cycle.

Then, two engineers started having an outsized impact. One of our AI-forward engineering leaders resolved a deployment problem in a matter of days that the team had been blocked on for two years. Another delivered a feature set in five weeks that the team had scoped at 17, with 87% of commits coming from his AI agent. Both had the same tools and hours as their peers. The difference was that they had reorganized their workflow around AI agents rather than adding AI on top of what they were already doing.

Our CEO, Buck Brody, and I had been tracking this shift since early 2025. When a new generation of models arrived in December, the decision came quickly: Go all-in, organization-wide.

A productivity tool makes you faster at what you were already doing. An AI agent can completely change what you're doing.

In late 2025, AI agents arrived at scale in software development. When Andrej Karpathy, one of the founders of OpenAI, posted in early 2026 that “it is hard to communicate how much programming has changed due to AI in the last two months: not gradually and over time, but specifically this last December,” he was describing this transition. This wasn't just a tool getting better—it was a category changing.​

Foundation 1: The Data Layer

We run a cloud data warehouse with a semantic layer. When an agent reaches for business data, it reads meaning, not raw tables. It knows what ACV means, what the retention calculation assumes and what a churned customer is in our context. An agent on an undocumented data model will fabricate. An agent on a semantic layer will reason. That difference determines whether AI is useful or dangerous.

Foundation 2: The Knowledge Layer

Alongside the data layer, we built a knowledge layer: a single source of truth for how we work, and a context graph indexing the organization across systems, tickets and decisions. Without it, agents answer the narrow question and miss everything else. With it, an agent can surface the customer complaint filed last week, the design decision from three sprints ago and the data definition that resolves the ambiguity. The codebase is one layer. The organization around it is what agents need to see.

Foundation 3: The Operating Model

Our product teams already ran as small, empowered squads with continuous validation built in. When agents arrived, we didn't have to teach teams to iterate. We had to teach them to bring an agent into the iteration. Organizations still running large, slow, consensus-heavy delivery models will find that agents accelerate the individual contributor without changing the team’s output. The operating model is the bottleneck, not the tool.

The three rollout decisions were:

1. Preparing every major codebase before training anyone, with context files and clear agent boundaries across all active repositories

2. Building an AI champions network with a shared skill marketplace, peer credibility front-running the formal rollout

3. Redefining what we expected from engineers and shifting the high-value work from writing code to making better decisions

The proof that the culture had genuinely shifted came at the quarter’s end: a two-day hackfest of 40-plus prototypes in 33 hours. Engineers built a conversational interface for our takeoff estimation product, natural language search for construction professionals and optimizations for our computer vision models that delivered what would normally take a full quarter. That's what happens when engineers have the tools and the permission.

None of this required a large technology budget. It required a belief, held consistently, that clean data, accessible knowledge and fast-moving teams were worth the discipline.

​Today, at ConstructConnect, technology, product management, design and every other department are operating in fully agentic mode. The agentic era doesn't reward early adopters of tools. It rewards the ones who rebuilt for it.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?