Will AI Kill The UI?
Last week I was in a training session, and one of the facilitators floated the idea that AI will make UI obsolete. That statement made me think, while my brain associated it with a fairly old song titled “Video Killed the Radio Star”, which explores themes of technological change, nostalgia, and the impact of media on our lives.
The Holy Grail
Some of our customers are starting to say that all they want is to have a single interface for everything. That interface would be their LLM of choice: Microsoft Copilot, Anthropic’s Claude, or OpenAI’s ChatGPT. I understand very well the lure. Imagine writing the following prompts in a single interface and a single session:
- Build a journey that implements the cart abandonment use case. I want to send two emails: one 1h after the last add-to-cart event and, if the customer has not purchased after 24h, a second one with an offer. Build the needed audiences in RT-CDP. For both emails, use the corporate template. For the first email, send a generic text with an approved image. For the second email, include a 10% offer in the text.
- Send a proof of each email to myself and my manager.
- Review the email I have just received with the proof. Check that company guidelines are followed and that the personalization is accurate.
- Push the journey to production, at midnight.
- Build a report that shows the increase in conversion and average order value thanks to this campaign.
- Generate a daily PDF based on the report data, using the company’s brand guidelines, and send it to the distribution list
[email protected]. - Keep an eye on the reports and send me an email in case you detect any anomalies in the data.
I can think of the following benefits if this worked:
- No need to get trained on enterprise-grade tools like Adobe Experience Platform and its apps. LLM’s skills can do the job.
- No need to go to multiple interfaces to do each of the steps. You do not leave the comfort of your LLM’s interface.
- The LLM takes the heavy lifting and does all the work: creates segments, journeys and emails, sends proofs, reviews proofs, builds reports and evaluates reports.
- The LLM knows your company’s requirements and restrictions and automatically adds them to the prompts. You do not need to remember them.
I am sure there are more.
All That Glitters Is Not Gold
I have to admit that I am skeptical of these claims. Do not get me wrong, I am using AI to write Python scripts, summarize meetings, correct text, translate, and I am amazed at how good it is. I have even seen demos of Adobe CX Enterprise Coworker and it is very promising.
However, I see many reasons why the UI is needed and why we will have to use it, maybe not as often as now, but still indispensable.
The Limits of Prompts
Language is inherently ambiguous. If you are married, you know it all too well: how many times has your spouse misunderstood you, when you thought your question or request was crystal clear?
While natural language is fantastic for sparking ideas or high-level orchestration, it falls short when dealing with precise visual layouts, design adjustments, or complex setups. Again, think of how often you have to provide additional information to your LLM chats, as the responses were not what you were expecting. With pointing, clicking, dragging, and toggling you create exactly what you have in mind.
Training for heavy prompt engineering does not necessarily eliminate friction; it just shifts it. Crafting, debugging, and maintaining massive prompts to execute hyper-specific tasks can become just as cumbersome as learning a new application. And this is without considering that prompt syntax or model behaviors drift over time.
Accountability and Trust
When dealing with mission-critical systems, enterprise data, or production environments, would you trust “black box” automation? I know we are in the digital marketing world and nobody is going to die because of an agent going rogue, but thousands of dollars could be wasted.
- UIs show you the exact state of a system, asset libraries, or audience segment counts, rather than forcing you to trust a conversational summary.
- Fine-tuning pixel placement, adjusting permissions, or visually inspecting a rendered email template layout can be faster than through iterative conversational prompts.
- Enterprise software relies heavily on role-based access control (RBAC), adding another layer of complexity in setting up the right permissions to agents and other AI tools.
The Cost of Ignorance
I remember reading newspaper articles about how the younger generations were not more technically literate. Sure, the expectations were probably too high: some people thought that just because they used a tablet in school they could write code in seven languages. But what was also clear is that they did not have any additional skills that the previous generation lacked. They knew how to use tablets and phones earlier than us, but that was about it.
One of my personal concerns with current society is that we are willingly giving up responsibilities to others or tools, without understanding what is really going on under the hood. We tell ourselves that we do not need to know how an internal combustion engine works, how the financial system is built, or what is really behind the numbers our governments lie to us about. With AI, we will keep on lowering the bar and, if we trust LLMs to do our job, nobody will know what AEP is doing.
I am a big fan of first learning how to do something by hand, to only then use automation tools.
Only Time Will Tell
I think we still have to wait a year or two before we can conclude one way or the other. As I said, I am initially skeptical; I believe that we will still need to log in to the underlying tools to do some or most of the work.
What are your thoughts?
Photo by Eric Krull on Unsplash