HomeLeadership

As Token Bills Climb, ConstructConnect’s CTO Sizes AI Model Costs To The Outcomes They Deliver

November 18, 2026

Gaurav Singal, Chief Technology Officer of ConstructConnect, described a cost discipline that funds only the AI use cases with a defined return and sends the rest to cheaper models or plain software.

As Token Bills Climb, ConstructConnect’s CTO Sizes AI Model Costs To The Outcomes They Deliver
Credit: CIOnews

Get the latest from CIOnews.

Enterprise AI, governance, risk, and leadership insights for CIOs, CTOs, CISOs, and technology leaders.

Thank you! Your submission has been received!
Oops! Something went wrong while submitting the form.
"At some point, the upside of the outcome you're driving is limited. If AI is giving a customer a status update, it doesn't really matter whether you use a top frontier model, because the value of that task is capped."

Gaurav Singal

CTO
ConstructConnect

Enterprise AI budgets are starting to buckle under a cost few teams priced in at the pilot stage: the token bill that multiplies once a workflow runs across thousands of users. Keeping it in check is a matter of discipline. Match each task to the smallest model its outcome can justify, reach for traditional software when that scales more cheaply, and fund only the handful of use cases with a defined return, and the cost question starts to answer itself.

Gaurav Singal is Chief Technology Officer of ConstructConnect, where he's moved the technology organization to agentic delivery and rebuilt how the company brings construction projects to market. Before ConstructConnect, he held technology leadership roles across financial services, logistics, and gaming, including a stretch as Vice President of Technology at Goldman Sachs and Executive Vice President and Chief Information Officer at Georgia Lottery Corporation. His read on AI spending starts from a single test of whether the outcome is worth the compute.

"As long as leaders know what outcome they're driving and what the ROI is, they can formulate that in terms of model cost or token cost. Then all this token optimization, these cost conversations, will fade away, and people will just be talking about the outcomes they're driving with AI," Singal said. That reframing changes what gets measured first. Instead of defaulting to the most capable model on the menu, his teams start from the value of the task and work backward to the cheapest way to deliver it.

  • Outcomes over horsepower: "At some point, the upside of the outcome you're driving is limited. If AI is giving a customer a status update, it doesn't really matter whether you use a top frontier model, because the value of that task is capped." The task's ceiling, not the model's capability, is what sets the budget for it. Knowing a task well enough to match it to a right-sized model is the lever, and work routed to a cheaper model has become a mainstream way to pull it. Singal treats the resulting spend as sound only when the return clears the cost.

  • When plain software wins: The case for a frontier model weakens as a workflow moves from a demo to daily production. "Frontier models are great for experimentation and getting quickly to a proof of concept. But if you're trying to scale something across an organization, it may not be the right cost-efficient solution, because now you're looking at thousands of people times that workflow, times the token cost," Singal said. For work that has to run at that volume, he points teams toward conventional software engineering when it can carry the load at a fraction of the price.

  • The new sprawl: Speed has a byproduct, and at ConstructConnect the pace is deliberate. "We're moving almost ten times the speed we were last year, so it's inevitable you create some tech debt and architecture sprawl. What's good now may not be good tomorrow, and as tech leaders you have to recognize that and address your tech debt more frequently than you used to." The discipline he pairs with it is a standing cleanup cycle, a sprint every three months aimed at tech debt and security to keep the foundation current before sprawl hardens into a constraint.

Cost discipline and a clean architecture still leave the hardest problem untouched: getting an agent to behave the way its designers intended once it meets real production data. In an industry like construction, that data's rarely tidy. A single project can arrive as thousands of pages of drawings and specifications, with competing versions and no reliable record of which one is current, and a model asked to reason over it will reason confidently from the wrong source. Closing that gap is where the next two priorities sit, and neither of them is a prompt.

  • Reliability starts with evals: What stops a system from reasoning off the wrong source is evaluation, treated as seriously as the product it governs. "It's very important to have evals and an evals framework for your products and anything you build intelligence into. You really need to put a lot of hard work into evals," Singal said. The payoff is visible downstream: estimators upload project plans that can run to thousands of pages, and computer-vision models trained against the wide variation in how architects draw them now quantify materials, counting doors, windows, and fixtures, in minutes rather than the weeks or months it once took. That same hard work is what flags answers that look right but aren't. AI carries most of the load, by his estimate 80 to 85 percent of it, and the last stretch still takes real effort.

  • Translating value upward: Getting a cost-conscious AI program funded means speaking the finance team's language, and Singal keeps three categories ready for it. "I always look at it through three lenses," he said. "One is how it accelerates revenue. The second is whether it improves the bottom line and our EBITDA through efficiency and performance gains. And the third is risk management, including reputation risk, like avoiding a cyber attack and the regulatory cost that follows." Put in real dollars, that translation is what has carried his case with CFOs, CEOs, and the board. It also argues for restraint on the input side: he caps active experiments at three to five use cases with a defined upside, ranked by business impact, rather than a registry of hundreds nobody can account for.

The thread running through all of it is a willingness to do the accounting before the building. Singal caps experimentation at a few high-value use cases, time-boxes the pilots, and names the upside of each before it starts. The results, when the discipline holds, are concrete: a new commercial real estate project that once took 60 hours to reach a customer now gets there in six.

The measurement, in his telling, follows from the setup more than the tooling. "If you do that hard work up front, it just makes it easier to measure, because now you've actually put a thought around it."

research report

From the Edge to the Core:
Bringing Agentic AI to the Heart of the Enterprise.