Dirty data has been coming up a lot in our conversations lately. It keeps surfacing when clients ask why their new AI features aren’t delivering what they expected. The answer is almost always the same, and it’s typically not the AI itself.
I wanted to bring the topic back because it’s worth understanding before you flip the switch on anything AI-powered in HubSpot.
Here’s the thing that stuck with me. When a person looks at a CRM record with an obvious mistake, they catch it and move on. AI doesn’t do that. It treats whatever it finds as a pattern to learn from, and then it repeats that pattern at scale. A contact record split into three duplicates becomes three separate signals. A lead source field that’s blank on a third of your contacts teaches the model to ignore that field entirely. Nothing flags it. The reports still run, the dashboards still show numbers, but you just can’t trust them.
There are five types of dirty data, and each one degrades AI output through a different mechanism, which means each one needs a different fix. That distinction matters a lot if you’re getting ready for an AI rollout, or if you’ve already turned on predictive scoring and forecasting and found the results underwhelming.
The short version: reliable AI starts with reliable data. If the foundation is messy, the model doesn’t rise above it. It learns the mess.
-Sage Levene, VP of Marketing, Hypha HubSpot Development
Open Mic
The Space Between Deliverables
By Jon Sasala, President
One of the things we’ve always tried to do at Hypha is make it easy for our team to do their best work.
Every client engagement requires thoughtful planning, clear expectations, and a well-defined scope. Those details matter because they create alignment before the work begins. But our goal has never been to anticipate every possible scenario or write the most restrictive statement of work. It’s to create enough clarity that everyone can stop thinking about the agreement and focus on solving problems.
That philosophy shapes how we deliver for our clients.
Our project managers aren’t successful because they follow a checklist. They’re successful because they’re trusted to lead. We expect them to think strategically, adapt when priorities shift, and make decisions that move a client’s business forward. The less time they spend worrying about whether something falls inside or outside the contract, the more time they can spend creating value.
That’s the kind of flexibility we want every client to experience.
Like any good partnership, though, it only works when the trust goes both ways.
The strongest client relationships aren’t transactional. Both sides are invested in the outcome. We look for opportunities to help because we believe it’s the right thing to do, not because a contract tells us to. The best clients approach the relationship the same way. They trust the expertise they hired, communicate openly when priorities change, and understand that great work often requires thoughtful decisions that couldn’t have been predicted on day one.
That doesn’t mean boundaries disappear. Healthy partnerships protect both organizations because both are investing talented people, time, and expertise into the work. Flexibility should create momentum, not become an expectation that one side carries more of the burden than the other.
When trust replaces scorekeeping, the contract fades into the background, and the work takes center stage.
Hot new strategies and hacks may come and go, but this is a good reminder that abuse of a platform never wins out.
“‘If you’ve been abusing Reddit as a platform to promote yourself, a drop in citations may make it appear less enticing,’ Glen Allsopp, head of marketing strategy and research at the AI marketing platform Ahrefs, told me. ‘Nothing should really change for those who view it as a place to legitimately interact with their audience.’”
“Meta has agreed to change the ways teens use its platforms and pay states up to $17.1 billion to settle claims its products are addictive and harmful to kids. It’s the latest sign that long-running public frustration with tech companies is turning into concrete changes for the industry.
“The settlement announced this week with 47 states, the District of Columbia, and three U.S. territories aims to be about more than just Meta. It’s structured as an attempt to set industry standards for how platforms treat young users.”
Hypha Highlights
Salesforce Free Suite is a solid starting point for a two-person team focused primarily on structured sales and customer service. HubSpot’s free CRM is generally the stronger fit for businesses that want lead capture, marketing, sales, and service tools connected in one system. But the best choice is not necessarily the platform offering the most free features today. It is the one whose limitations, data structure, and paid path best match how your business expects to grow.
Z.ai, the Chinese lab behind the GLM series, has been revealed as the maker of Ox Alpha, a mysterious open-weight model that launched anonymously on OpenRouter and quickly topped benchmarks against leading models. Described as a reasoning model built for coding and sustained agentic work, Ox Alpha adds to the growing wave of cheap, capable Chinese models challenging pricier frontier labs.
Anthropic plans to publicly unveil its IPO prospectus after Labor Day, with a potential listing in late September or early October. The company is reportedly considering steps that would set its offering apart from this year’s SpaceX and Cerebras debuts, including letting existing shareholders sell shares and weighing lockup periods longer than the customary 180 days, and is expected to surpass the $86 billion SpaceX raised in June.
OpenAI CEO Sam Altman said on a podcast that he worries AI could end up controlled by a handful of companies or people, leaving most consumers without a say, and warned that excessive fear of AI’s risks could lead to trading away too much liberty for safety.
Additionally, OpenAI is attempting to recover after losing its lead to rival Anthropic. Following a breach in which one of its unreleased models escaped a test sandbox and hacked Hugging Face, OpenAI says it is slowing development—pausing a major training run until new safeguards are in place—while pushing an aggressive product and infrastructure roadmap and claiming it is roughly 80% of the way to AGI.
In his first address to Cursor’s staff after SpaceX completed its $60 billion acquisition, Elon Musk said SpaceX’s AI unit had fallen behind competitors, praised Anthropic for leading the AI race, and urged employees to help Grok catch up while the company continues stockpiling compute. Cursor is now being folded into SpaceXAI teams with CEO Michael Truell reporting directly to Musk, even as roughly 45 people have left the startup since April amid the integration.
In an hour-long interview and a nearly 6,000-word essay, Bill Gates warned that AI poses grave risks to jobs and human life, arguing the tech industry is knowingly downplaying the threats because of the trillions of dollars at stake and that addressing them should be “the world’s top priority.”
Salesforce and Anthropic announced an expanded partnership called Claudeforce, whose centerpiece is Salesforce in Claude—a plugin for Claude CoWork that includes 37 pre-built sales skills and lets sellers query, update, and act on live CRM data without ever opening Salesforce itself, available to pilot customers now with an open beta in September.
More than 100 tech companies—including OpenAI, Anthropic, Google, Microsoft, and cyber firms like CrowdStrike, Okta, and Fortinet—signed an open letter urging the public and private sectors to collaborate on defenses against increasingly sophisticated AI-enabled cyberattacks, warning that critical infrastructure like hospitals and water treatment plants is at risk.
Nvidia has agreed to buy open-source AI platform Hugging Face for $12.9 billion, a deal that would expand the chipmaker’s reach across the AI stack and put one of the most widely used platforms for sharing open-source models under its ownership.
“HubSpot spent three years rebuilding how it attracts, engages, and delights customers around AI agents. Jon Dick, chief customer officer at HubSpot, walks through what that actually looked like — starting with AI fluency rather than tools, then support, then a full audit of the customer journey for constraints AI could break.
“Along the way: the AEO strategy built for a world where buyers search answer engines, the AI SDR handling more than 80% of website chats, a prospecting agent that booked 10,000 meetings last quarter, a seven-point lift in customer save rates, and the org change three months ago that moved every engineer and data person working on agentic go-to-market under one leader.”
How can we help you?
Case Study: HubSpot CMS Rebuild
Behaviorally’s platform helps shopper-insights teams read the market faster than competitors can react. The website wasn’t keeping pace—still running on its original layouts, rigid enough that making changes required engineering support, and mobile was an afterthought.
We rebuilt the theme from the ground up: drag-and-drop backend, separate mobile and desktop spacing controls, and 15+ custom modules including an animated homepage hero, interactive PPS Score Card with real animation, scroll-triggered product slider, and animated data counters. We rewrote core page copy to speak directly to shopper-insights professionals instead of explaining the platform to itself. And, we used Behaviorally’s brand pink as an accent, not a default—a deliberate call that keeps the identity from tipping into caricature.
Session time is up 15% in the 30 days since launch. The team now edits and rebuilds pages without engineering involvement.
Have a product that’s moved faster than the site representing it? Contact Hypha to talk through a CMS rebuild.
