HubSpot's AI strategy, "Breeze," will be deeply integrated across Marketing Hub from 2025 to 2026. From content generation and lead prediction to customer agents and the Copilot Assistant, incorporating AI into the design of your marketing operations, rather than using it as a "nice bonus," can dramatically improve your team's productivity. A systematic explanation of the overall picture of Breeze and how to use each function.
HubSpot at 2024 INBOUND Conference Breeze , and unified the previously dispersed AI functions under the "Breeze" brand. In January 2026, the underlying model will transition to GPT-5, significantly improving content generation accuracy, context understanding ability, and multilingual support.
Breeze's architecture consists of three layers.Top layer “Agents”performs specific tasks autonomously,Middle-tier “Copilot”acts as a daily operational assistant,Foundation layer “Intelligence”will be responsible for data enrichment and predictive analysis.
With the transition to GPT-5 in January 2026, the following improvements have been confirmed across Breeze.①Improving the accuracy of Japanese content generation(appropriate adjustment of natural Japanese expressions and honorific language level),②Improving the ability to retain long text contexts(enables consistent generation of long blog articles and white papers),③Improvement of reasoning ability(Improving the accuracy of complex decisions such as ``Which measure is best for this segment?'') However, it should be noted that due to the increase in generation costs, some functions have begun to transition to pay-as-you-go billing based on usage.
Breeze Agents is a group of functions in which ``AI autonomously executes routine to semi-routine tasks that were previously performed by humans.'' However, it is not fully automatic,"Human-in-the-Loop" design where agents perform work and humans approve and modify itis the basis (detailed in 10-6).
The true value of Content Agent is not “generating one article”"Batch generation of multiple articles based on content pillars"It's in For example, for a pillar page called "Introduction to HubSpot MA," once you set up a pipeline that generates outlines for 10 related cluster articles → generates the body of each article → adds metadata → generates SNS posts, mass production of content becomes dramatically more efficient. Maximum effectiveness can be achieved by combining it with the SEO topic cluster strategy explained in Chapter 2.
Breeze Copilot is an AI assistant that resides on the right side of every HubSpot screen. Just say in natural language, ``Do this,'' ``Look up this,'' or ``Create a report on that,'' and HubSpot will perform the operations on your behalf.The more you use HubSpot, the closer you will get to an environment where you don't need to know how to use HubSpot.——This is the essence of Copilot.
① Convey the target, purpose, and conditions specifically——You'll get better results if you say, ``Write a subject line and body text that eliminates concerns about MA implementation costs for purchasing managers in the manufacturing industry,'' rather than ``Write an email.''② Tell “why”——Adding background information can improve the quality of your product.③ Polish by repeating——Brush up the initial product by giving additional instructions such as “shorter,” “more casual,” and “add examples.”
Breeze Intelligence is Breeze's "data foundation layer." Combining an external database (formerly Clearbit data source) with an AI predictive model,Auto-completion of CRM data, detection of purchase intention, prediction of contract probabilityWe provide three values: The "accuracy of decisions" made by upper-layer agents and Copilot all depend on the quality of this data.
| rule-based scoring | Predictive Lead Scoring | |
|---|---|---|
| structure | Scores are added or subtracted according to rules determined by humans (such as "Pricing page visit = +15 points") | AI automatically learns patterns from past sales/loss data and calculates scores. |
| setup cost | Sales and marketing need to agree and design and adjust rules. | Almost no settings required. Accuracy automatically increases as data accumulates |
| transparency | ◎ I can explain why I got that score. | △ Black box-like. It's hard to explain why that score |
| accuracy | △ Depends on the designer's hypothesis. There is an oversight | ◎ The richer the data, the more accurate it is than human hypothesis. |
| Recommended usage | When introducing a new product, when there is little data, and as a consensus building tool with sales. | Used in combination after contract data has been accumulated (more than 50 to 100 contract results) |
Attempting to deploy all Breeze features at once fails.There are stages of maturity when using AI.. It is important to accurately understand which phase your company is currently in and move forward gradually to the next stage.
Increasingly, companies are introducing Journey Automation or Customer Agent before the data infrastructure is in place due to the expectation that AI will do everything.AI accuracy is proportional to data quality and learning amount. In the end, the fastest path is to prepare the data in stages 1 and 2, verify the accuracy of the AI in stage 3, and then proceed to stage 4.
| stage | Priority KPI | Guidelines for “proceeding to the next stage” |
|---|---|---|
| 1: Assist | Reduction rate of content production man-hours/quality score of generated content | The main workflow (MQL promotion/nurturing) was activated and 3 months of data was accumulated. |
| 2: Automation | Reduced number of workflow operations, open rate of automatically delivered emails, and MQL promotion lead time | More than 50 contract data have been accumulated, and attribution by channel can be stably measured. |
| 3: Optimization | Predictive Score Order rate of top leads/AI email CTOR improvement rate | AI prediction accuracy begins to outperform human rule-based scoring |
| 4: Autonomy | Number of MQLs managed per marketer/Ratio of order revenue due to AI | The adoption rate of AI-based policy proposals has reached over 60%. |
Each function of Breeze is ``autonomously executed by AI → approved and corrected by humans.'' Human-in-the-Loop（HITL） It is based on the design concept of The final key to achieving results from AI utilization is to properly design the balance between ``leaving it to AI'' and ``human control.''
Routine/repetitive work (data completion, score calculation, posting scheduling), processing of large amounts of data, pattern discovery, draft generation/first draft creation, calculation of optimal time/distribution. Work that requires speed and volume.
Making brand values and ethical judgments, determining strategic direction, building emotional relationships with customers, final confirmation of expressions that pose legal risks, and setting a vision for why we are implementing certain measures.
AI-generated content will have a flow of ``save draft → human review → publish after approval'' rather than ``publish immediately''. Especially essential for legal, medical, and financial-related content, reports for management, and crisis response communications.
Monthly monitoring of Predictive Scoring prediction accuracy, AI email CTOR, and Customer Agent response accuracy rate. If the accuracy has decreased, review the learning data and reset the rules. AI is not a ``set it up and be done'' thing.
Agents (autonomous execution) → Copilot (dialogue assistant) → Intelligence (data infrastructure). The three layers work together. The data quality of the base layer determines the accuracy of the upper layer.
Rather than generating a single article, the greatest productivity improvement can be achieved by using it to build a pipeline that generates multiple articles at once based on SEO topic clusters.
Convey the target, purpose, conditions, and background in detail. The most efficient way to use it is to brush up the initial product by adding additional instructions.
AI predictive scoring requires training data. Before that, we will operate with rule-based scoring, and once data is accumulated, we will introduce Predictive Scoring in parallel.
Progress through the four stages of assistance, automation, optimization, and autonomy in order. If you rush to stage 4 without preparing the data infrastructure, accuracy will not be achieved and you will fail.
All AI generation goes through a “draft → human review → approval” flow. The key to long-term quality maintenance is to clearly separate the work left to AI and the decisions made by humans.