Adobe’s AI agent delivers design intern-level results

Most AI image tools feel like they’re designed to cut the user out of the creative loop entirely. Type a prompt, get a result, and hope for the best. That’s why Adobe’s latest experiment with its Firefly AI Assistant caught my attention. Instead of replacing the creator, this bot is built to handle the grunt work while leaving the artistic decisions where they belong, in your hands.

Currently in beta, the Firefly AI Assistant doesn’t generate images from scratch in the same way as other tools. Think of it more as a highly capable middleman that operates Adobe’s design apps on your behalf. According to Adobe, you can simply “tell Firefly AI Assistant (beta) what you need, and it will use tools from apps like Photoshop, Illustrator, and more to complete multistep projects in moments.”

The interface is familiar, a standard chatbot window with a text box and an upload button for media files. But instead of running the Adobe apps on your machine, it taps into their core capabilities, including masking, object detection, and image generation. The conversational design means you can ask it to “make this photo more colorful,” and it will comply while walking you through its reasoning.

I tested it with a challenging photo of myself, one with unusual lighting. The results were convincing at a glance. It changed my hair color, swapped out backgrounds, and adjusted lighting. Was it perfect? No. Some colors came out too vivid, and a few edits weren’t seamlessly blended. But to the average viewer, the final images wouldn’t scream “AI manipulation.” They’d just look like the work of a designer who’s still learning the ropes.

What truly sets the Firefly AI Assistant apart is its transparency. I uploaded a picture of my cat by a window and asked for a cloudless, sunny sky. Instead of silently delivering the result, the chatbot described the scene in detail, correctly identifying my cat as a Maine Coon even though the photo mostly showed his backside. Then, it explained exactly how it would achieve the requested edit, referencing specific Photoshop and Lightroom tools with proper terminology. You don’t see the edits happen in real time, but you get a step-by-step narration of the process.

The assistant is also refreshingly honest about its own limits. When I asked it to separate objects in a JPG into individual layers, it admitted it couldn’t do that directly. Instead, it offered two alternative approaches, complete with the pros and cons of each. After I chose one, it described its progress, even acknowledging when a technique wasn’t working. “I notice the gaussian blur approach isn’t giving me true transparent cutouts,” it wrote, before pivoting to use masks and Adobe’s cropping tools instead.

Adding new objects to images, similar to Photoshop’s Generative Fill or Google’s Magic Editor, is also possible. It didn’t hesitate to add cigars or even guns to my photos, but it drew the line at anything outright illegal. I could make a fake album cover of myself pointing a gun at the camera, but not one where I appeared to be shooting someone. The quality of these generative edits was noticeably lower than simpler adjustments like lighting changes, but they weren’t terrible. It also refused to alter my body shape or put me in revealing clothing, a restraint that other AI tools could learn from.

Most chatbots feel like overly enthusiastic theme park mascots. The Firefly Assistant isn’t immune to that, occasionally offering unnecessary praise like calling my request a “great idea.” But for the most part, its commentary is genuinely useful. When it needs more information, it asks. I requested a photo of two cocktails be turned into an illustration for a bar’s social media ad. The bot then asked which platforms I was targeting, offering options like TikTok, Instagram, and Facebook, along with the standard pixel dimensions for each. Even if I hadn’t known that Instagram Square posts are 1080×1080 pixels, I would have learned it on the spot.

This is the most engaging interaction I’ve had with a creative chatbot. By doing the tasks I never bothered to learn, it reveals the gaps in my own knowledge, explaining its process as it goes. It’s less comprehensive than a Google or YouTube tutorial, but those won’t complete the job for you while you learn. I personally enjoy manual editing too much to hand it over to a bot, but for people who want quick, decent results without the time investment, I can see the appeal.

Canva recently launched a similar conversational design agent. It shares the same tendency for flowery language and patronizing praise, but it doesn’t explain its workflow. You just give it instructions and keep prompting until you’re satisfied. The results I saw from Canva didn’t match what Adobe’s Firefly Assistant produced, especially for those willing to learn the basics of design and editing along the way.

Adobe is positioning this assistant as a time-saver for creative professionals, a way to offload labor-intensive tasks. With my middling editing skills, I found it useful only if I was comfortable with subpar results. For seasoned designers, using this tool might feel more like supervising a clumsy intern than collaborating with a capable colleague. If the Firefly AI Assistant can eventually produce work indistinguishable from professional editing, it could become a valuable asset. For now, it’s still very much a novice.