This is a part considered one of a three-part sequence on how HubSpot remodeled with AI. Phase two covers how we develop with Agent-first GTM. Phase 3 is how we function as an AI-first corporate.
The entirety we construct at HubSpot exists to assist our consumers develop. So when generative AI emerged, our engineering group didn’t simply see a productiveness device; we noticed a possibility to construct higher merchandise and get extra price into consumers’ arms quicker.
And when off-the-shelf AI gear hit their ceiling, we didn’t simply search for higher ones. We constructed the platform beneath them. That call compounded sooner than we anticipated. As a result of all of our AI is constructed on a shared basis, each and every new capacity we send makes the entire machine extra tough and consumers get a extra constant revel in throughout the entirety they use.
Lately, we’re in a position to innovate at a tempo that merely wasn’t imaginable ahead of. 100% of our engineers use AI, and we’ve noticed a 73% building up in strains of code written through our engineers.
We didn’t get right here in a single day. It took 3 stages, actual infrastructure funding, and a willingness to construct what didn’t exist but. Right here’s how we did it.
Contents
Section 1: Productiveness with Co-pilots (2023-2024)
In 2023, massive language fashions had simply crossed the edge of being really helpful in a coding context. The most efficient resolution for the use of AI in engineering used to be to begin with what used to be confirmed. At the moment, it used to be code of entirety: a human writes code, and AI copilots recommend what comes subsequent.
We rolled out a coding copilot and were given to 30% adoption briefly. Then we pulled the incident information, when compared groups the use of the copilot towards groups that weren’t, and proved AI adoption didn’t negatively have an effect on the reliability of the product.
With that information in hand, we got rid of the guardrails and gave everybody copilot get admission to. Adoption shot previous 50% in a single day. This taught us a lesson in how we make choices. Measure, turn out, then scale.
By means of the top of Section 1, 80% of engineers have been the use of AI gear. We noticed a 51% development in engineering speed, that means engineers have been transport operating code to manufacturing considerably sooner, and a 7% building up in strains of code up to date according to engineer. We proved AI may just make each and every engineer sooner with out compromising product reliability.
Section 2: Scaling with Coding Brokers (2024-Mid 2025)
Your next step used to be self sufficient coding with brokers. Our groups may just recommended the gear to finish end-to-end duties. The brokers may just learn context, write code, run checks, and connect mistakes, all whilst the engineer reviewed and urged. We felt strongly this used to be the way forward for engineering and dedicated absolutely.
The true constraint got here briefly. Off-the-shelf coding brokers may just no longer get admission to inner construct methods, our libraries, or test that code in truth labored in the environment. So, we constructed the ones agent integrations ourselves the use of MCP, a typical that permits AI brokers to connect with exterior gear and methods, and deployed them to each and every engineer. To pressure adoption, we arranged occasions to provide engineers devoted area to be informed, experiment, and construct self belief with new gear. Agent utilization went from 0 to 80% adoption in a month.
The following problem used to be scale. Engineers sought after a couple of brokers working in parallel, in a single day, with out supervision. So we constructed an agent execution platform on most sensible of our Kubernetes infrastructure. Each and every agent runs within an remoted container that replicates an actual HubSpot developer surroundings. Brokers bring together the code, run computerized checks, learn error outputs, and iterate on their very own till the entirety works. No human intervention required.
By means of the top of Section 2, 96% of engineers have been the use of AI gear, engineering speed used to be up 60%, and contours of code up to date according to engineer had larger 48%. We have been beginning to send higher merchandise sooner with brokers. However that used to be just the start.
Section 3: Scaling with our AI Platform (Mid 2025-Provide)
HubSpot’s platform technique to product construction has all the time been how we’ve created extra buyer price. Once we constructed reporting and automation on the platform stage, we didn’t simply send one characteristic; we shipped that capacity throughout each and every hub concurrently. That’s how innovation compounds.
We carried out that very same good judgment to our AI infrastructure in Section 3. As a substitute of creating each and every agent from scratch, we constructed the shared basis as soon as: how brokers get admission to information, what movements they are able to take, how they connect with the remainder of HubSpot. The entirety runs on most sensible of it.
The result’s that every one of our brokers are interoperable. They discuss the similar language, proportion the similar toolsets, and draw from the similar context. A buyer will get a constant revel in without reference to which agent they’re the use of as a result of, beneath, they’re all constructed at the identical infrastructure. And since they’re all attached, each and every new capacity we upload makes the entire machine extra treasured. That’s one thing a choice of level answers can not reflect.
And it used to be made imaginable through how we’ve scaled engineering with AI. Lately, 100% of our engineers use AI, strains of code up to date according to engineer are up 73%, and time-to-first-feedback on pull requests has dropped through 90%. That implies much less time ready and extra time transport issues consumers in truth use.
Why this issues: Compounding buyer price
Having the proper infrastructure hurries up the tempo of innovation. For HubSpot, each and every agent we construct makes the platform extra tough. Each and every piece of context we upload to the platform makes each and every agent simpler. For patrons, that suggests the product assists in keeping getting higher, sooner, and extra attached.
What used to take months now takes weeks, and the ones weeks translate at once into new features within the arms of entrepreneurs attempting to succeed in the proper target market, reps looking to shut offers, and Buyer Good fortune Managers looking to retain consumers. They don’t wish to take into consideration the platform beneath. They just get to revel in the outcome.
