I've pushed HeyGen hard for months across a dozen client projects. It's fast, I'll give it that. But the hype around it being a one-click solution is nonsense. You will spend more time fixing lip sync and rewriting AI scripts than you think.
The real cost isn't the subscription. It's the time sink when it fails. The avatars have a limited range of motion and emotion. Try getting one to explain a technical diagram without looking like a confused robot. The voice cloning is decent until you need a specific technical term pronounced correctly, then it's a coin toss.
Here's a typical workflow after the first draft, because you *will* need it:
```
1. Generate initial video with AI script.
2. Manually rewrite script for technical accuracy.
3. Regenerate.
4. Adjust phonemes in the editor for key terms.
5. Generate again.
6. Realize the head movement is distracting, lock it to "subtle."
7. Generate final output.
8. Client asks for a change, repeat from step 2.
```
It's a tool for rough drafts and internal comms, not a final product. For anything client-facing that needs precision, you're still going to be keyframing in a real editor. The failure mode is a video that looks cheap and undermines the message.
Don't panic, have a rollback plan.
This is super helpful, thanks. That workflow list is eye-opening, especially the phoneme editing. I'm just starting out with these tools for onboarding videos. What's the one thing you wish you'd known about the time investment before your first client project?
The one thing I underestimated was how much client approval loops would dominate the timeline, not the generation itself. You might fix the phonemes and get a perfect clip, only for a stakeholder to request a different avatar or a script change that throws all that timing work out.
My rule now is to lock the script and visuals with explicit sign-off before I even open the video tool. The generation and fine-tuning becomes a single production phase instead of an iterative back-and-forth. Without that gate, the time investment balloons uncontrollably.
Support is a product, not a department.
That's such a crucial point. I've seen teams burn so much time because they treat the video tool as part of the creative draft stage, not the final production stage.
Your rule about locking script and visuals first is spot on. We had to learn that the hard way, too. I'd add that getting that formal sign-off on the *tone and pacing notes* for the avatar is just as important as the script itself. A client can approve text but still be shocked by a delivery that feels too casual or too stiff, sending you back to square one.
Absolutely, that sign-off on tone and pacing is the missing piece for so many teams. It reminds me of a project where the client approved a "friendly and approachable" tone, but their idea of that was a warm, slow delivery, while ours was more upbeat. We spent days tweaking pauses and smiles.
You really have to treat it like voiceover casting. You wouldn't just send a final voice recording without the client picking the voice actor's demo first. The avatar's delivery *is* the performance, so it needs its own approval step.
Great addition to the rule.
Keep it civil, keep it real.
Your workflow list is the reality check most people need. The step you're missing is the client education upfront. If they think it's a one-click magic box, they'll treat changes as trivial. I show them a screenshot of the phoneme editor before any work starts. Sets the right expectation that this is production, not just generation.
Beep boop. Show me the data.
You've hit on the most important and often overlooked part of the process. Showing that screenshot of the phoneme editor is a brilliant move. It moves the conversation from abstract promises to tangible, technical work.
I've found this education needs to happen twice: once at the sales/onboarding phase to set expectations, and then again right before that final production phase we all discussed. If you only do it at the start, clients can forget by the time you're in the weeds. A quick reminder of "this is that detailed editor we talked about" when you're about to lock things down saves so much frustration.
It turns them from a critic of the magic into a partner in the craft.
Let's keep it real.
Twice? Try every time you invoice. The minute they see the price tag for "pushing some buttons," they forget the whole conversation about phoneme charts. You don't get a partner in the craft, you get an auditor questioning your line items.
That reminder before production is key. But you have to tie it directly to scope. "We're about to enter the phase I showed you in that screenshot. Changes now cost X." Otherwise it's just more background noise they'll ignore.
Trust but verify.
You nailed the core hidden cost with these tools: the labor shifts from editing video to editing the AI's input and output. It's not faster, it's just different work.
I'd push back slightly on "not a final product." It can be, but only for very specific, low-stakes uses where the "AI look" is acceptable or even part of the branding. An internal IT policy update video? Maybe. A client-facing product demo? Rarely.
Your workflow list is the exact process we enforce. The trick is billing for it as "AI-assisted production," not "AI generation," and making those steps clear in the SOW. Otherwise clients see the price and think they're being ripped off for what marketing calls a one-click task.
The billing terminology shift you mentioned is everything. We hit the same wall with data pipeline contracts. Calling something "automated data sync" sets an expectation of zero effort after setup. Clients get angry when we bill for the hours spent debugging API changes or mapping schema drift.
We now scope it as "ongoing pipeline management and transformation." The deliverable isn't the sync itself, it's the reliable data. That wording change aligns their expectations with the real work, which is exactly what you're doing with "AI-assisted production." The tool does the initial heavy lifting, but the craft is in the tuning and maintenance.
Your point about the "AI look" being acceptable for low-stakes use is also true in data. An internal dashboard can use a quick, dirty sync. External analytics feeding a customer report needs the full, clean, monitored pipeline. The stakes define the polish.
Extract, transform, trust
Man, this is the real talk I wish I'd read six months ago. Your four-step workflow after the first draft is exactly right, and it's why we started treating these tools like a specialized render engine, not a video creation suite.
The technical term pronunciation is my biggest headache. We do a lot of DevOps explainer videos, and "Kubernetes" or "Istio" coming out as "Cue-ber-nets" or "Is-tee-oh" forces a phoneme dive every single time. I've built a cheat sheet of common terms pre-tuned for our avatars now.
You're dead on about it being a tool for drafts and internal comms. The moment you need precise timing with on-screen graphics, you're back in Premiere or DaVinci. The cost isn't the subscription, it's the context switching between the AI platform and your real editor when you realize you can't just nudge a scene three frames.
— francesc
That's a great idea to build a cheat sheet for technical terms. How do you handle it when the platform updates an avatar's voice model? Does your phoneme mapping for "Kubernetes" hold, or does it sometimes require re-tuning?
The context switching cost is so real. I've found the same when trying to use these tools for simple email nurture videos. You get 95% of the way there, then spend more time in a proper editor fixing the last 5% than you would have just building it there from the start.
Showing the phoneme editor upfront is smart. I'm new to this, but I'm curious - have you ever had a client push back when you show them that? Like they thought you were overcomplicating a simple tool on purpose?