When our blog team made the switch to an AI-first writing model in June 2025, my reaction was split. There was excitement on one side: I could see the inefficiencies in our process that AI might actually fix. On the other, anxiety. And while plenty of that anxiety was about the technology itself (how well these tools could actually write, how much we could trust them), my primary concern was my team.<\/p>\n
I\u2019ve managed HubSpot\u2019s blog writing team for three years, and before that was a HubSpot Blog writer myself. I understand what it means to have ownership over your process. Many of the writers I manage got into this work because they love writing. They enjoy the craft of finding the right sentence and earning a reader\u2019s attention. Traffic numbers and conversions are secondary. During this AI transition, I felt the weight of what it could mean for those whose entire approach to their craft was changing fast.<\/p>\n
<\/a><\/p>\n So much of the public conversation about AI content gets the story wrong. There\u2019s an assumption that because AI can write passably well, the hard part is over. Just apply the same low-effort polish you\u2019d apply to an AI-drafted email or memo to your blog post, then ship it.<\/p>\n But I can tell you from experience: it doesn\u2019t work that way. A blog post is meant to compete against dozens (or hundreds) of other posts on the same topic. In this game, generic, low-effort writing will not win, and that\u2019s good news for your content team. It means that even though the writing process has changed, the craft remains important.<\/p>\n So, how do you balance craftsmanship with the efficiency gains AI promises? And how do you build a process that takes advantage of what AI is actually good at without falling into its traps? We spent the last year figuring that out. Here, I\u2019ll share what we learned: where time was saved, where it got added, and what it takes to make AI work without compromising the quality standards you\u2019ve spent years building.<\/p>\n Table of Contents<\/span><\/strong><\/p>\n Before AI, writing was personal.<\/a> <\/a> <\/p>\n Before AI, our writing process was, well, a writing process. While we had editorial standards and guardrails to ensure consistent quality, we didn\u2019t enforce a specific method for how posts should be written. We hired writers and editors for their ability to craft, and we trusted them to use that ability. We understood that writing is more than just a technical skill. Giving our writers the space to develop their own individual workflows yielded content with personality and unique insights.<\/p>\n When training new writers, we taught a simple, flexible framework to help folks start to structure their time:<\/p>\n Outlining and planning<\/strong> took around 30 minutes to an hour. <\/p>\n Research<\/strong> could take one to two hours, depending on the complexity of the topic and the writer\u2019s familiarity with it. This included sourcing quotes, finding original data, and writing from personal experience. <\/p>\n Drafting<\/strong>, the part where someone actually puts words to the page, took around five to seven hours on average for a first draft. <\/p>\n Editing, QA, formatting<\/strong>, and finalizing in our CMS<\/strong> added another roughly two hours on top of that. <\/p>\n For a standard blog post, you\u2019re easily looking at a full day\u2019s work, usually longer. The bulk of that time lived in the drafting phase \u2014 writing the words, finding the right structure, and figuring out how to explain something clearly was the core of the job.<\/p>\n <\/a> <\/p>\n When we formally adopted AI-assisted writing in June 2025, my approach was deliberate: don\u2019t over-prescribe. Allowing our writers to develop their own processes had served us well for over a decade, and we wanted to maintain that sense of control for our writing team as much as possible, even as we added AI to the mix.<\/p>\n We gave writers access to AI tools<\/a> (specifically Claude and ChatGPT<\/a>, along with NotebookLM) and space to figure out how to work them into their process. Even with AI, I believed writing is personal, and I wasn\u2019t going to tell a team of experienced writers exactly how to do their jobs. Our goal was experimentation first, structure later.<\/p>\n What followed was not the clean efficiency win I expected. AI solved some problems and introduced new ones. But by working through those problems with our writers, we built a process that we know works.<\/p>\n At first, the signs were good. Writers were putting words on the page faster than before, and I assumed that would translate directly into faster posts overall.<\/p>\n It didn\u2019t.<\/p>\n A couple of months in, after reviewing time tracking data and having conversations with my team, a more complicated picture emerged. The quantity of posts stayed about the same, but posts were getting longer, more detailed, and more demanding to produce. Writers were spending more time testing software, sourcing expert quotes, reviewing tools in-depth, and going deeper on topics. The time AI saved in drafting was being absorbed by everything else a competitive post requires.<\/p>\n What made it harder was that some writers felt the pressure of both expectations at once: write faster and write more. When I understood that this was happening, I had to recalibrate. AI hadn\u2019t decreased the workload. In many ways, it had increased the workload.<\/p>\n So my first lesson was: If you\u2019re measuring AI\u2019s impact on your team purely by post count, you\u2019re going to be disappointed. The win is in what those posts become, not how fast they get there.<\/p>\n I also assumed AI would make research faster. After all, instead of spending an hour getting up to speed on a topic, a writer could ask Claude and move on.<\/p>\n For general topic knowledge, that\u2019s actually true. On well-documented subjects with plenty of training data to draw from, AI can explain concepts clearly, answer follow-up questions, and get a writer up to speed faster than navigating a dozen browser tabs.<\/p>\n The shortcomings, however, were flagged by writers almost immediately.<\/p>\n The first time it really surfaced was on a listicle covering competitor products. A writer came back to me flagging that the pricing information AI had surfaced was wrong. Not slightly off, but fully outdated. AI didn\u2019t know the pricing had changed, and it didn\u2019t flag its own uncertainty.<\/p>\n I probably should have seen that one coming. After all, a tool trained on data from months ago obviously can\u2019t know that a competitor changed their pricing. But it took a writer caring enough to flag it before I accounted for it.<\/p>\n From this, we learned that for anything requiring accuracy about competitor pricing or features, AI research falls apart. This makes sense when you think about how these tools work: An LLM\u2019s training data has a cutoff date, meaning it\u2019s unlikely to reflect the latest pricing changes, feature updates, or product rebranding. And when AI tools do attempt a live web search, they\u2019re only as good as the pages they can parse. Product pages are often structured in ways that are difficult for an LLM to interpret, which makes even real-time lookups prone to inaccuracies. The result is information that sounds authoritative but may be months or years out of date.<\/p>\n We produce a lot of competitor listicles and product-focused content, and for those, research still requires significant manual effort. So while AI has sped up general-topic research, competitor- and product-focused research takes roughly the same amount of time as it always has.<\/p>\n The research problem ran much deeper than product pricing alone. Beyond outdated competitor data, we were seeing fabricated statistics, misattributed quotes, and incorrect terminology, all delivered with the same confident, authoritative tone. AI doesn\u2019t know what it doesn\u2019t know.<\/p>\n We place a high premium on accuracy, and these so-called AI hallucinations<\/a> turned fact-checking from a light final step into one of the most time-consuming parts of the process. In some cases, AI drafting took longer than manual drafting because writers had to spend time verifying what the AI got wrong.<\/p>\n What struck me was how quickly my team caught it. That commitment to publishing only truthful, verified information is exactly what separates a good writer from a content generator. It reinforced something I already believed: the writer matters more than the tools they use.<\/p>\n To address hallucinations, we introduced a new formal step in our process: dedicated fact-checking. Pre-AI, writers were responsible for the accuracy of their own research, and a final editorial pass would catch anything obvious. With AI in the mix, fact-checking became its own distinct phase. For a typical post, I estimate it adds around 30 to 60 minutes to the process. It\u2019s longer for research-heavy pieces or anything involving stats, quotes, or product information.<\/p>\n Sure, you could skip fact-checking and assume AI gets everything right. But is that really the standard you want to hold your content to? Publishing misinformation, even unintentionally, damages trust in a way that\u2019s hard to recover from.<\/p>\n QA time grew too, for reasons beyond what fact-checking was already catching. Fact-checking is the process of verifying that the information in a post is correct. QA is about everything else: Does this post sound like a person wrote it? Does it meet our editorial standards? Is it structured in a way that search engines and AI tools can parse?<\/p>\n Without a structured quality layer, AI-assisted drafts introduced problems writers wouldn\u2019t necessarily catch on their own, such as:<\/p>\n AI clich\u00e9s and hedge-word-heavy phrasing (\u201cit\u2019s worth noting,\u201d \u201cleverage,\u201d \u201cit\u2019s important to understand,\u201d etc) To address this, I built a custom QA checker directly into Google Docs that flags these problems and more. It\u2019s become a required step in our workflow.<\/p>\n Through all the changes above, some things didn\u2019t move. The E-E-A-T requirements<\/a> we had before AI \u2014 original quotes, firsthand perspectives, solid data sourcing \u2014 stayed intact. Editorial judgment stayed intact. The strategic thinking that goes into deciding what a post needs to say, and who it needs to say it to, stayed intact.<\/p>\n AI handles much of the execution now: drafting, scoping, generating topic angles, and explaining concepts. But the layer of work that makes a piece worth reading (the human perspective, the original insight, the rigorous sourcing) is still on the writer. If anything, it matters more now because it\u2019s what separates your content from everyone else\u2019s AI output.<\/p>\n <\/a> <\/p>\n To understand what changed, it helps to see it. Here\u2019s a rough breakdown of how time was distributed across a typical blog post before AI, and how that distribution shifted after we adopted it:<\/p>\n Note: Time estimates are averages based on HubSpot Blog Team internal tracking. Individual posts vary based on topic complexity, length, and content type.<\/em><\/p>\n\n Of course, the reality varies more than these charts suggest. A writer might draft for an hour, realize they need more research, go back to drafting, catch something that needs fact-checking, and return to drafting again. A real workflow looks more like this:<\/p>\n\n Still, the time savings are evident.<\/p>\n <\/a> <\/p>\n If AI can generate a blog post in minutes, why not just publish and move on? More content means more chances to rank, more surface area for leads, and more output for less effort, right?<\/p>\n Early in the AI content boom, we all noticed a certain strategy emerging in the SEO space: Publish as much AI-generated content as possible, as fast as possible, and capture market share before competitors caught on. Flood the zone. It operates on the belief that volume alone would be enough to win.<\/p>\n Except, it isn\u2019t, and the consequences caught up.<\/p>\n The content that comes from such a fast-and-loose strategy is recognizable to anyone<\/a> who\u2019s spent time in this industry.<\/p>\n A blog post that strings together sentences without saying much. A listicle that could have been written about any company in any industry. No original data. No real point of view. No specific examples that couldn\u2019t have been fabricated. And if you run it through a plagiarism checker, you\u2019ll even find content that was lifted from competing sources (not intentionally, but as a byproduct of how LLMs work).<\/p>\n In short, it\u2019s generic. You\u2019ve probably heard the term \u201cAI slop\u201d to describe this kind of content as well. Experienced readers and, more importantly, search algorithms can spot slop from a mile away, and it hurts far more than it helps.<\/p>\n Marketers are catching on. According to HubSpot\u2019s 2026 State of Marketing Report<\/a>, 52% of marketers believe AI has made content so easy to create that it\u2019s become less effective overall.<\/p>\n Need more proof? In 2026, SEO expert Lily Ray tracked more than 220 websites<\/a> that identified themselves as customers of AI content creation platforms and found a consistent pattern: rapid growth in organic traffic, followed by a steep decline that, in many cases, erased the gains entirely. Across her dataset, more than half of the sites lost at least 30% of their peak organic traffic, and nearly a quarter lost 75% or more.<\/p>\n Here\u2019s an example of what that looks like for one such website.<\/p>\n\n Source<\/em><\/a><\/p>\n I understand why teams took to this approach at first. The pressure to produce is real: 83% of marketers say they\u2019re expected to create more content than ever because of AI<\/a>.<\/p>\n But this is exactly what leading only with speed looks like in practice. If quality becomes something you try to retrofit after the fact, it rarely works.<\/p>\n The teams doing AI writing well start from the opposite end. They define what \u201cgood\u201d looks like \u2014 voice, depth, accuracy, originality \u2014 and then work backward to determine where AI can help them get there faster without compromising on those standards. All writing is anchored in quality from the beginning.<\/p>\n On the HubSpot Blog, that means keeping our E-E-A-T requirements intact. We build quality guardrails into our workflow and are honest about which content types benefit from AI involvement and which don\u2019t. For example, a how-to guide and a piece of original research will have different requirements, and treating them the same way will result in generic content that goes unnoticed.<\/p>\n As former Director of the HubSpot Blog Karla Hesterberg\u00a0puts it: \u201cThe only truly effortless process produces slop.\u201d Teams that resist the slop, keep good writers in the driver\u2019s seat, and use AI as a lever rather than a replacement are the ones producing work that converts.<\/p>\n For a more thorough dive into how we maintain quality standards in an AI-assisted editorial process, check out this post from HubSpot Blog writer Amy Rigby on how HubSpot creates quality content in the age of AI<\/a>.<\/p>\n
\n What We Learned From Letting Writers Lead<\/a>
\n Comparing Pre-AI vs. Post-AI Writing<\/a>
\n AI content at scale is easy. AI content worth reading is hard.<\/a>
\n What This Means for People Managers<\/a>
\n My Takeaways as a Content Team Manager<\/a> <\/p>\nBefore AI, writing was personal.<\/h2>\n
What We Learned From Letting Writers Lead<\/h2>\n
Faster drafting doesn\u2019t mean faster content.<\/h3>\n
Research was our biggest surprise.<\/h3>\n
The Hallucination Problem<\/h3>\n
Quality doesn\u2019t police itself.<\/h3>\n
\n Redundant sentences.
\n Plagiarism from competing content (not intentional, but from models quoting their source material directly in the output)
\n Broken or faulty links
\n Missing image alt text
\n Incorrect heading orders
\n Sentences, paragraphs, and sections that aren\u2019t optimized for LLM chunkability<\/a>. <\/p>\nWhat AI Didn\u2019t Touch<\/h3>\n
Comparing Pre-AI vs. Post-AI Writing<\/h2>\n
AI content at scale is easy. AI content worth reading is hard.<\/h2>\n
Slop at Scale<\/h3>\n
Start with standards, then add AI.<\/h3>\n