What actually happened when our blog went AI-first: A manager's account

When our blog squad made the toggle to an AI-first penning example in June 2025, my reply was split. There was excitement on one side: I could see the inefficiencies in our procedure that AI power really fix. On the other, anxiety. And during plentifulness of that anxiety was concerning the innovation itself (how fine these tools could really write, how much we could rely them), my chief involvement was my team.

I’ve managed HubSpot’s blog penning squad for three years, and before that was a HubSpot Blog author myself. I comprehend what it method to have ownership complete your process. Many of the writers I oversee got into this activity since they affection writing. They bask the accomplishment of finding the correct declaration and earning a reader’s attention. Traffic numbers and conversions are secondary. During this AI transition, I felt the importance of what it could average for those whose complete method to their accomplishment was changing fast.

So much of the community conversation concerning AI satisfied gets the narrative wrong. There’s an assumption that since AI can compose passably well, the difficult part is over. Just use the identical low-effort polish you’d use to an AI-drafted email or memo to your blog post, afterward container it.

But I can inform you from experience: it doesn’t activity that way. A blog article is meant to vie against dozens (or hundreds) of another posts on the identical topic. In this game, generic, low-effort penning volition not win, and that’s fine news for your satisfied team. It method that equal although the penning procedure has changed, the accomplishment remains important.

So, how do you balance craftsmanship alongside the effectiveness gains AI promises? And how do you build a procedure that takes advantage of what AI is really fine at without falling into its traps? We spent the final twelvemonth figuring that out. Here, I’ll portion what we learned: anywhere period was saved, anywhere it got added, and what it takes to create AI activity without compromising the norm standards you’ve spent years building.

Table of Contents

• Before AI, penning was personal.
• What We Learned From Letting Writers Lead
• Comparing Pre-AI vs. Post-AI Writing
• AI satisfied at measure is easy. AI satisfied value study is hard.
• What This Means for People Managers
• My Takeaways as a Content Team Manager

Before AI, penning was personal.

Before AI, our penning procedure was, well, a penning process. While we had editorial standards and guardrails to justify accordant quality, we didn’t enforce a particular method for how posts have to be written. We hired writers and editors for their capability to craft, and we trusted them to use that ability. We understood that penning is additional than fair a specialized skill. Giving our writers the area to create their own idiosyncratic workflows yielded satisfied alongside character and distinctive insights.

When training new writers, we taught a simple, elastic example to assistance folks commencement to construction their time:

• Outlining and planning took about 30 minutes to an hour.
• Research could obtain one to two hours, depending on the complexity of the topic and the writer’s familiarity alongside it. This included sourcing quotes, finding first data, and penning from individual experience.
• Drafting, the part anywhere person really puts words to the page, took about five to seven hours on average for a archetypal draft.
• Editing, QA, formatting, and finalizing in our CMS added another approximately two hours on top of that.

For a norm blog post, you’re effortlessly looking at a complete day’s work, normally longer. The bulk of that period lived in the example phase — penning the words, finding the correct structure, and figuring out how to explain item plainly was the center of the job.

What We Learned From Letting Writers Lead

When we formally adopted AI-assisted penning in June 2025, my method was deliberate: don’t over-prescribe. Allowing our writers to create their own processes had served us fine for complete a decade, and we wanted to keep that awareness of authority for our penning squad as much as possible, equal as we added AI to the mix.

We gave writers admission to AI tools (specifically Claude and ChatGPT, alongside alongside NotebookLM) and area to fig out how to activity them into their process. Even alongside AI, I believed penning is personal, and I wasn’t going to inform a squad of informed writers exactly how to do their jobs. Our goal was experimentation first, construction later.

What followed was not the spotless effectiveness win I expected. AI solved several problems and introduced new ones. But by operating through those problems alongside our writers, we built a procedure that we cognize works.

Faster example doesn’t average faster content.

At first, the signs were good. Writers were putting words on the leaf faster than before, and I assumed that would translate immediately into faster posts overall.

It didn’t.

A brace of months in, following reviewing period tracking data and having conversations alongside my team, a additional complex image emerged. The amount of posts stayed concerning the same, but posts were getting longer, additional detailed, and additional demanding to produce. Writers were spending additional period evaluation software, sourcing expert quotes, reviewing tools in-depth, and going deeper on topics. The period AI saved in example was being absorbed by everything alternatively a rivalrous article requires.

What made it harder was that several writers felt the force of the two expectations at once: compose faster and compose more. When I understood that this was happening, I had to recalibrate. AI hadn’t decreased the workload. In many ways, it had risen the workload.

So my archetypal instruction was: If you’re measuring AI’s effect on your squad purely by article count, you’re going to be disappointed. The win is in what those posts become, not how accelerated they get there.

Research was our biggest surprise.

I additionally assumed AI would create investigation faster. After all, alternatively of spending an hr getting up to speed on a topic, a author could ask Claude and move on.

For broad topic knowledge, that’s really true. On well-documented subjects alongside plentifulness of training data to diagram from, AI can explain concepts clearly, answer follow-up questions, and get a author up to speed faster than navigating a dozen browser tabs.

The shortcomings, however, were flagged by writers nearly immediately.

The archetypal period it really surfaced was on a listicle covering competitor products. A author came rear to me flagging that the pricing data AI had surfaced was wrong. Not slightly off, but completely outdated. AI didn’t cognize the pricing had changed, and it didn’t emblem its own uncertainty.

I likely should have seen that one coming. After all, a tool trained on data from months ago evidently can’t cognize that a competitor changed their pricing. But it took a author caring adequate to emblem it before I accounted for it.

From this, we learned that for item requiring accuracy concerning competitor pricing or features, AI investigation falls apart. This makes awareness whenever you think concerning how these tools work: An LLM’s training data has a cutoff date, definition it’s improbable to indicate the latest pricing changes, characteristic updates, or merchandise rebranding. And whenever AI tools do attempt a live web search, they’re lone as fine as the pages they can parse. Product pages are frequently organized in ways that are difficult for an LLM to interpret, which makes equal real-time lookups prone to inaccuracies. The outcome is data that appears authoritative but may be months or years out of date.

We create a lot of competitor listicles and product-focused content, and for those, investigation motionless requires important manual effort. So during AI has sped up general-topic research, competitor- and product-focused investigation takes approximately the identical amount of period as it continually has.

The Hallucination Problem

The investigation issue ran much deeper than merchandise pricing alone. Beyond outdated competitor data, we were seeing fabricated statistics, misattributed quotes, and incorrect terminology, all delivered alongside the identical confident, authoritative tone. AI doesn’t cognize what it doesn’t know.

We location a elevated premium on accuracy, and these so-called AI hallucinations turned fact-checking from a ray final stage into among the most time-consuming parts of the process. In several cases, AI example took longer than manual example since writers had to expend period verifying what the AI got wrong.

What struck me was how quickly my squad caught it. That dedication to publishing lone truthful, verified data is exactly what separates a fine author from a satisfied generator. It reinforced item I already believed: the author matters additional than the tools they use.

To location hallucinations, we introduced a new ceremonial stage in our process: dedicated fact-checking. Pre-AI, writers were liable for the accuracy of their own research, and a final editorial continue would capture item obvious. With AI in the mix, fact-checking became its own distinct phase. For a representative post, I evaluation it adds about 30 to 60 minutes to the process. It’s longer for research-heavy pieces or item involving stats, quotes, or merchandise information.

Sure, you could skip fact-checking and assume AI gets everything right. But is that really the norm you desire to clasp your satisfied to? Publishing misinformation, equal unintentionally, damages rely in a way that’s difficult to regain from.

Quality doesn’t constabulary itself.

QA period grew too, for reasons beyond what fact-checking was already catching. Fact-checking is the procedure of verifying that the data in a article is correct. QA is concerning everything else: Does this article audio akin a individual wrote it? Does it encounter our editorial standards? Is it organized in a way that hunt engines and AI tools can parse?

Without a organized norm layer, AI-assisted drafts introduced problems writers wouldn’t necessarily capture on their own, specified as:

• AI clichés and hedge-word-heavy phrasing (“it’s value noting,” “leverage,” “it’s crucial to understand,” etc)
• Redundant sentences.
• Plagiarism from competing satisfied (not intentional, but from models quoting their origin matter immediately in the output)
• Broken or faulty links
• Missing depiction alt text
• Incorrect heading orders
• Sentences, paragraphs, and sections that aren’t optimized for LLM chunkability.

To location this, I built a tradition QA checker immediately into Google Docs that flags these problems and more. It’s rotate into a required stage in our workflow.

What AI Didn’t Touch

Through all the changes above, several things didn’t move. The E-E-A-T requirements we had before AI — first quotes, firsthand perspectives, firm data sourcing — stayed intact. Editorial judgement stayed intact. The strategic thinking that goes into deciding what a article needs to say, and who it needs to say it to, stayed intact.

AI handles much of the implementation now: drafting, scoping, generating topic angles, and explaining concepts. But the tier of activity that makes a part value study (the individual perspective, the first insight, the rigorous sourcing) is motionless on the writer. If anything, it matters additional now since it’s what separates your satisfied from everyone else’s AI output.

Comparing Pre-AI vs. Post-AI Writing

To comprehend what changed, it helps to see it. Here’s a coarse breakdown of how period was distributed throughout a representative blog article before AI, and how that allocation shifted following we adopted it:

Note: Time estimates are averages according to HubSpot Blog Team inner tracking. Individual posts change according to topic complexity, length, and satisfied type.

Of course, the actuality varies additional than these charts suggest. A author power outline for an hour, acknowledge they need additional research, go rear to drafting, capture item that needs fact-checking, and come back to example again. A genuine workflow looks additional akin this:

Still, the period funds are evident.

AI satisfied at measure is easy. AI satisfied value study is hard.

If AI can create a blog article in minutes, why not fair publish and move on? More satisfied method additional chances to rank, additional exterior area for leads, and additional output for small effort, right?

Early in the AI satisfied boom, we all noticed a certain scheme emerging in the SEO space: Publish as much AI-generated satisfied as possible, as accelerated as possible, and grasp market portion before competitors caught on. Flood the zone. It operates on the condemnation that quantity solitary would be adequate to win.

Except, it isn’t, and the consequences caught up.

Slop at Scale

The satisfied that comes from specified a fast-and-loose scheme is recognizable to anyone who’s spent period in this industry.

A blog article that strings together sentences without saying much. A listicle that could have been written concerning any business in any industry. No first data. No genuine item of view. No particular examples that couldn’t have been fabricated. And if you run it through a copying checker, you’ll equal discover satisfied that was lifted from competing sources (not intentionally, but as a byproduct of how LLMs work).

In short, it’s generic. You’ve likely heard the term “AI slop” to depict this benevolent of satisfied as well. Experienced readers and, additional importantly, hunt algorithms can place slop from a mile away, and it hurts far additional than it helps.

Marketers are catching on. According to HubSpot’s 2026 State of Marketing Report, 52% of marketers accept AI has made satisfied so uncomplicated to create that it’s rotate into small productive overall.

Need additional proof? In 2026, SEO expert Lily Ray tracked additional than 220 websites that identified themselves as customers of AI satisfied innovation platforms and established a accordant pattern: fast growth in integrated traffic, followed by a steep decrease that, in many cases, erased the gains entirely. Across her dataset, additional than fractional of the sites misplaced at smallest 30% of their highest integrated traffic, and nearly a fourth misplaced 75% or more.

Here’s an example of what that looks akin for one specified website.

Source

I comprehend why teams took to this method at first. The force to create is real: 83% of marketers say they’re expected to create additional satisfied than always since of AI.

But this is exactly what foremost lone alongside speed looks akin in practice. If norm becomes item you try to retrofit following the fact, it rarely works.

Start alongside standards, afterward add AI.

The teams doing AI penning fine commencement from the contrary end. They define what “good” looks akin — voice, depth, accuracy, originality — and afterward activity backward to decide anywhere AI can assistance them get there faster without compromising on those standards. All penning is anchored in norm from the beginning.

On the HubSpot Blog, that method keeping our E-E-A-T requirements intact. We build norm guardrails into our workflow and are honest concerning which satisfied types advantage from AI involvement and which don’t. For example, a how-to guide and a part of first investigation volition have distinct requirements, and treating them the identical way volition outcome in generic satisfied that goes unnoticed.

As erstwhile Director of the HubSpot Blog Karla Hesterberg puts it: “The lone really effortless procedure produces slop.” Teams that oppose the slop, keep fine writers in the driver’s seat, and use AI as a lever fairly than a replacement are the ones producing activity that converts.

For a additional thorough dive into how we keep norm standards in an AI-assisted editorial process, inspect out this article from HubSpot Blog author Amy Rigby on how HubSpot creates norm satisfied in the age of AI.

What This Means for People Managers

Almost overnight, my function shifted in two directions at once. On one side, I became additional process-oriented than I’d always been. I had to think additional deeply concerning workflows, tooling, and how posts get assembled. On the other, group administration got harder. Not since my writers were difficult, but since I was asking them to alter how they work.

Process changes are comparatively straightforward to document, but maintaining motivation and morale through a passage that basically changed their jobs is a distinct issue entirely. You can build the most thoughtful AI penning procedure in the world, but if your writers aren’t bought in, none of it works.

Writers arrive to this job for distinct reasons, and AI lands alternatively depending on which logic that is.

Some writers are wired for outcomes. They enter satisfied promotion since they’re fine at penning and they akin seeing the endeavor impact. For them, AI is a greeted alter since it gets them to the outcome additional quickly.

Other writers came to this activity since they love the craft. The penning procedure is item they attention concerning deeply. For them, an AI-first example can awareness akin it’s taking item away. I don’t see this as a achievement matter or a nonaccomplishment to adapt. It’s an inner strain that have to be taken seriously.

As a manager, my job is to acknowledge anywhere all squad associate falls on that spectrum and activity alongside them accordingly. I create a area for honest feedback, including negative feedback concerning AI. I cognize that if writers don’t awareness comfortable telling me what isn’t working, I’ll build processes on bad information. The writers who most honestly pushed rear on our first AI workflows gave me the clearest roadmap for what needed fixing.

Steps for Managing Your Team Through the AI Transition

Here’s what I’d inform any director navigating AI writing:

• Don’t over-prescribe early. Give writers admission to tools and area to test before locking in rigid workflows.
• Mandate experimentation, not method. The teams that prescribed particular AI use from day one created unnecessary friction.
• Recognize that you’re asking writers to alter how they work. That takes period and trust.
• Create area for honest feedback, including constructive feedback. If writers don’t awareness harmless telling you what isn’t working, you’ll build processes on bad information.
• Understand anywhere all author falls on the spectrum between outcome-driven and craft-driven, and oversee them accordingly.

The goal hasn’t changed: a squad that produces activity value reading. The way to get there fair looks distinct from what it did in 2024.

My Takeaways as a Content Team Manager

A twelvemonth into this transition, I’m arrogant of anywhere we landed, equal although we’re motionless figuring several things out.

What made the difference wasn’t lone the tools, but the writers. I knew that going in, and this former twelvemonth confirmed it at all turn. The writers who pushed rear honestly gave me the roadmap, and the ones who struggled helped me comprehend what was really being asked of them.

What sticks alongside me most from this cognition is item writers have started saying recently: that they’re emotion akin genuine writers again. The balance between AI effectiveness and genuine accomplishment has started to awareness sustainable. That happened not since we established a ideal immediate or the correct tool, but since we never let go of what made our satisfied fine in the archetypal place.

If you’re an editorial guide navigating the AI change correct now, cognize this: What gets you through is having writers who attention adequate to be honest alongside you, and being a director who creates area for that honesty. Don’t underestimate what it method for a author to feel akin a writer. It shows up in the work.

AI is a tool. Keep your writers at the center, clasp the norm line, and the remainder follows.