LLM products need interfaces that help users refine imperfect outputs instead of hiding the model's uncertainty.
2
Prompting works better when products combine free-form input with structured controls, templates, suggestions, and model-generated prompt construction.
3
Production UX depends on accuracy feedback, human review, monitoring, fallback providers, caching, and smaller models for lower latency.
Summary
This panel examines how people should interact with products built on large language models. Artem Harutyunyan describes Bardeen's approach of showing users a generated automation in preview mode, keeping the input box available for refinement, and using suggestions to help people formulate requests. Misty Free explains how Jasper moved some workflows from chat toward buttons and dynamic templates after finding that long conversations created too much work. Dina Yerlan discusses Firefly's use of prompt assistance, predefined styles, variations, and user feedback for text-to-image generation. The panelists agree that users need help judging and correcting probabilistic outputs. Jasper uses Google search in Jasper Chat, while Bardeen keeps people in the review loop and collects telemetry. The final discussion covers latency. Jasper monitors spikes and switches providers, while Bardeen uses caching and is interested in smaller models that can run on devices.
Misty Free says marketers are facing fatigue from the number of AI tools and interaction styles available. Users can try chat, browser extensions, prompts, and many other paths, but their initial excitement is fading. Jasper therefore tests new features with existing customers and the wider market before adding them. The aim is to solve content-creation problems efficiently without making customers learn another feature that they may not need. This concern also appears in Artem Harutyunyan's comments about prototypes: a working demo can be built in an afternoon, but users quickly lose patience when a product works only three times out of ten.
Refinement is a normal part of using an unreliable model
Artem Harutyunyan compares LLM interaction with early web search. A generated automation opens in preview mode, while the original input box remains visible so the user can revise the request. Bardeen also suggests relevant ways to continue, using recommender-system techniques alongside the LLM. This gives users a path forward when the first result is close but wrong. Artem argues that the interface should make this back-and-forth feel natural rather than pretending that the first model response is final.
Jasper has tested free-form chat, a command bar inside its document editor, and button-based workflows. The team removed the command bar after finding that it became clunky and obstructed the editor. Jasper Campaigns also moved from a conversational flow toward buttons because users found it burdensome to read and write enough text before reaching content generation. Misty Free says there is no reliable rule for choosing an interface in advance. Dynamic templates have worked well because users describe their goal once, then receive fields tailored to the information needed for that task.
Prompting can be split between user input and product controls
Dina Yerlan says expert prompt writers get much better image results than people who enter a short request such as a picture of a dog. Firefly therefore considers prompt autofill, predefined styles, resolution controls, and other selections that keep users from writing every detail themselves. Variations let people learn from generated results and provide feedback for better suggestions. Artem Harutyunyan describes a similar middle ground for Bardeen: users select the dimensions that matter for a known task, while researchers construct and evaluate the final prompt behind the scenes.
Jasper Chat can enable Google search so users can check generated writing against web resources. Misty Free says users sometimes ask Jasper to include reference URLs, but generated citations can still be fabricated, so people must fact-check the result. She gives the example of Jasper producing a convincing legal defense of robot marriage with historical Supreme Court cases that never existed. Firefly collects thumbs-up and thumbs-down feedback about image quality, intent, and safety. Dina Yerlan connects these errors to training data, prompt handling, post-generation processing, and content moderation.
Human review and telemetry guide product improvement
Bardeen defines automations in a domain-specific language with properties that make them easier to verify. Artem Harutyunyan says the company keeps people in the loop by showing the proposed automation and asking whether it matches their intent. It records thumbs-up and thumbs-down responses, the number of attempts needed, later usage, and whether users appear satisfied. These signals feed back into the creation process. The panel treats accuracy as something measured through actual use rather than as a property that can be assumed from the model alone.
Latency requires operational fallbacks as well as model improvements
Misty Free says Jasper monitors latency and error spikes with Datadog and posts dashboard messages when users are likely to experience long waits or failures. Jasper also has fallback strategies that switch to another model or provider. Dina Yerlan discusses scaling Firefly across Adobe products and the use of caching for repeated requests. Artem Harutyunyan says Bardeen caches equivalent automation requests and expects smaller, specialised models that run on the edge or on a device to reduce latency and serving costs.
"The biggest challenge is just how to make this sequence seamless so it's very natural to the human being who is in front of the computer using your product."Artem Harutyunyan08:11
Who should watch
You are designing a product where users must correct, refine, or approve model-generated work.
Your team is deciding between chat, buttons, templates, suggestions, or a more specialised workflow.
You need practical ideas for measuring model quality and handling latency, provider failures, and scaling.
