As AI becomes a core part of digital products, founders are learning what truly works — and what doesn’t. In this roundtable, startup leaders share firsthand insights into how they've embedded AI into their product features to enhance functionality and user experience. The discussion also explores where AI has yet to live up to expectations, and the technical, design, and trust-related challenges that come with AI integration. This session offers a grounded look at the current state of AI in product development — what’s delivering value, what still needs work, and where to go next.
Our experts:
Ajay Singh, Founder at UnleashX
Andrew Miracle, Head of Research at Tecmie
Arpan Singh, SDE-2 at Atlassian
Camilo Chacón Sartori, PhD Student in AI at IIIA-CSIC & UAB
Can Kisi, AI Engineer and Consultant at Neural River
Lucrezia Noli, AI Strategy and Sales for Life Sciences at Google
Natallia Zhur, Compliance AI Champion
Os Hourani, Marketing Manager at Woolpert
Syed Bilal Ahmed, CO Founder and CTO at AstrumAI
Zaid Ahmad, GTM Specialist at Qapita
Ajay Singh, Founder at UnleashX
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
At UnleashX, we’ve deeply integrated AI into our core product features, transforming how users automate and scale operations.
Some key AI-driven features include:
Autonomous Agents: Users can create or deploy pre-built agents like AI SDRs, CRM updaters, or research assistants that operate independently — handling outreach, follow-ups, data updates, and insights without human intervention.
AI-powered Feedback Loops: We analyze customer feedback (NPS, CSAT, etc.) in real time, trigger automated follow-ups, and surface actionable insights — drastically reducing churn and enhancing product iteration.
Voice AI Agents: Our AI handles real-time outbound calls with voice cloning, multilingual support, and contextual understanding — improving conversion rates while saving human bandwidth.
These features have led to a 75%+ reduction in manual work, faster response times, and a significant boost in productivity for our users. AI isn’t just a layer on top — it’s baked into the foundation of how our users get things done.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
One major gap is in context-aware decision-making across complex workflows. AI often struggles when tasks require deep business context, long-term memory, or nuanced judgment — especially across multiple systems with inconsistent or siloed data.
Another area still falling short is autonomous reasoning in real-time scenarios, like voice interactions. While voice agents are improving, maintaining coherent, empathetic, and goal-driven conversations — especially in multilingual or noisy environments — remains a challenge.
Also, AI explainability is still lagging. Many users and businesses are hesitant to fully trust AI outputs when they can’t understand the "why" behind them — which limits adoption in critical domains like healthcare, finance, and legal ops.
Closing these gaps will require not just better models, but better integration with human workflows, real-world feedback loops, and hybrid systems that combine automation with human oversight.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
User Trust & Transparency: One of the biggest hurdles is getting users to trust autonomous agents. Users often ask: “What will it do next?” or “Can I control or override it?” We've had to build explainability into the UI — making AI decisions transparent and reversible, especially in CRM or voice automation use cases.
Performance vs. Latency Trade-offs: Real-time applications (like AI voice agents) demand ultra-low latency. Balancing model accuracy, context retention, and speed has been a constant engineering challenge — especially when deploying in environments with network or hardware constraints.
UX for Non-technical Users: Many users expect AI to “just work.” Designing intuitive interfaces that abstract AI complexity, allow customization, and handle edge cases gracefully has been an ongoing design effort. We learned the hard way that even the best AI needs thoughtful scaffolding in the UI to drive adoption.
Andrew Miracle, Head of Research at Tecmie
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
We've integrated AI into multiple product verticals within Tecmie, notably in Wootiv, our AI-powered conversation automation platform. One standout feature is its real-time intent detection and routing, which allows sales and support teams to prioritize high-intent conversations automatically — leading to a 60% increase in lead conversion velocity.
Another project, applybetter.co, uses AI to auto-match job seekers with opportunities based on a deep semantic analysis of their profile and behavior — not just their resumes. This dramatically improves discovery and application relevance, reducing time-to-hire.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
Trust and explainability remain serious gaps. While AI can automate complex decisions, especially in hiring (ApplyBetter) and sales qualification (Wootiv), users often want visibility into why certain actions were taken. We're still battling the “black box” problem.
Another weak point is generalization. AI models trained on startup or niche datasets often overfit. For example, in Wootiv, sentiment detection trained on Western corp-speak didn’t perform well in African startup contexts — requiring significant fine-tuning.
Finally, in edge-case handling (especially in identity attestations), AI often makes overly confident but incorrect decisions that could affect real-world outcomes like user verification or access to assets.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
User Trust: Users are wary of decisions made by opaque systems — especially in identity, hiring, and automation. We’ve had to build human-in-the-loop processes and visibility layers (like feedback loops and model audits) to mitigate this.
Latency vs Intelligence: In products like Wootiv, delivering real-time AI responses at scale without sacrificing performance was a challenge. We adopted hybrid architectures — fast keyword-based routing first, followed by deferred deep inference — to bridge the gap.
UX Alignment: Designing UIs that adapt based on AI outputs without overwhelming or confusing users has been tough. For instance, surfacing “AI-suggested edits” in ApplyBetter needed thoughtful design to avoid user fatigue or distrust.
Data Drift & Feedback Loops: Over time, the models degrade if not constantly retrained. We’ve had to engineer robust feedback capture mechanisms across our products to maintain relevance.v
Arpan Singh, SDE-2 at Atlassian
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
At Atlassian, we’ve integrated AI into Rovo, our new intelligent work assistant, to improve how teams search, collaborate, and automate tasks across tools. One key feature is AI-powered natural language search, which allows users to query information from Jira, Confluence, and even third-party tools like Slack, GitHub, or Google Drive. This has drastically reduced the time spent looking for information and context switching.
We’ve also built Rovo Agents — customizable AI bots that help with everything, from summarizing documents and onboarding new team members to triaging issues or cleaning backlogs. These agents automate repetitive tasks and ensure consistency at scale, letting teams focus on higher-value work.
By combining generative AI with our Teamwork Graph, Rovo doesn’t just understand content, it understands how your team works. The result is a more intuitive, responsive experience that saves time, reduces friction, and unlocks smarter decision-making within the flow of work.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
I think that generative AI has made content creation faster, but it often lacks true originality, emotional nuance, and long-form coherence. Whether it’s writing, design, or ideation, much of what it produces feels like it is pieced together from existing patterns rather than offering genuinely new perspectives. AI can mimic tone or structure, but it rarely captures the unique spark that comes from real experience. So, it works well as a brainstorming companion but still needs human direction to avoid generic or shallow outputs.
For example, AI tools can often generate stunning images, but they fail to keep some features consistent, like details on our face or the number of fingers in our hands. They sometimes blend inconsistent styles and confidently generate incorrect text, which can be improved by making them context-aware.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
A major challenge has been ensuring data privacy and security, especially when integrating AI into products that handle sensitive user or enterprise data. Users are increasingly cautious about how their data is used to train AI models. Striking the right balance between giving users a personalized experience and respecting their privacy is tricky but essential.
Camilo Chacón Sartori, PhD Student in AI at IIIA-CSIC & UAB
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
As a researcher working as an AI consultant in industry, I always start with the problem. The nature of the problem dictates the appropriate solution. While generative AI is currently in vogue, it’s not necessarily the right tool for every challenge. In my experience developing algorithms for underground mining companies, I had the opportunity to improve the accuracy of a system by replacing basic heuristics with an LSTM-based deep learning model, a classical neural network architecture. This significantly improved performance without requiring large-scale infrastructure or high costs. A small LSTM model proved to be both efficient and effective. The key lesson: don’t get carried away by trends; use the most suitable technology when the problem truly demands it.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
One of the main areas is the advanced design of algorithms. Today, algorithms are not just one component among many, they are a technology in their own right. Typically, this kind of work is carried out by R&D departments, where PhD-level researchers seek to design better algorithms that reduce computation time and improve results, ultimately helping companies increase profitability. However, recent developments such as Google DeepMind's AlphaEvolve suggest that generative AI can play a valuable role in this domain as well, assisting in the co-creation and implementation of advanced algorithms. We are entering an era in which software — especially non-trivial, high-level algorithmic software — will increasingly be written by both humans and machines working together.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
Trust is built through effectiveness and transparency. Effectiveness means demonstrating with clear data that an AI system outperforms the previous one, this is the primary argument for its adoption. However, many AI systems, particularly those based on deep learning or generative AI, operate as "black boxes." This lack of interpretability can raise concerns. That’s why it’s our responsibility to thoroughly document experiments and understand the path to the achieved results. As Ken Thompson once said: "You can't trust code that you did not totally create yourself." We must keep this in mind, especially when deploying AI systems that affect core business processes.
Can Kisi, AI Engineer and Consultant at Neural River
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
We built a RAG-powered chatbot for one of our major travel customers, pulling real-time info from their internal knowledge base to answer user questions. It made the bot way more accurate and useful, especially for complex itinerary or policy queries, and reduced the load on their support team.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
Creative originality is still limited:
While useful for drafts and ideation, AI fails to consistently deliver truly novel, emotionally rich, or context-aware creative work
Hallucinations and factual errors:
In areas demanding high accuracy, like legal, healthcare, or policy advice, AI often generates confident but incorrect or misleading output
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
One key challenge was keeping the AI fast and reliable, especially with RAG, where retrieval can slow things down. We also had to design around trust issues, adding fallback UX and clear indicators so users knew when to rely on the AI. Even with good performance, user skepticism and occasional hallucinations were hard to fully overcome.
Lucrezia Noli, AI Strategy and Sales for Life Sciences at Google
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
My focus is on applying AI to the Healthcare and Life Sciences domains. This includes fine-tuning models for scientific domains on Google infrastructure, leveraging the power of Gemini as part of an Agentic ecosystem to streamline processes like literature review and analysis, clinical trials comparison, identification of data gaps, and support with creating medical affair assets.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
Interacting with the external world and upstream/downstream applications, which is now the direction of travel with Agentic AI.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
Access to Data and access controls, change management within organizations.
Natallia Zhur, Compliance AI Champion
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
AI can help leveraging day-to-day tasks of a compliance officer by its integration into documents translation. This includes automated detection and classification of document types, context-aware translation engines, and terminology alignment with regulatory frameworks.
By leveraging AI-powered NLP models, it can significantly improve the speed, accuracy, and consistency of translations across multiple jurisdictions and languages. AI engine can be trained on domain-specific compliance language, which ensures that translated content retains regulatory integrity which is crucial in high-stakes environments like financial services.
For users, this means faster onboarding of global clients, reduced reliance on manual translations, and enhanced confidence in cross-border regulatory submissions. In practice, what used to take several days of human review can now be processed in near real time, dramatically accelerating workflows and reducing human error.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
While AI has made significant strides, there are still product areas where its impact hasn’t fully met expectations. One key area is nuanced judgment in complex regulatory interpretation. Despite advancements in natural language processing, AI models still struggle to understand the subtle context or intent behind legal language, especially in cross-border regulations where phrasing and enforcement priorities can vary.
Another limitation lies in document translation when dealing with low-resource languages or documents containing mixed legal and technical jargon. AI may produce fluent output, but not always with the jurisdiction-specific nuance required by compliance teams. This means human validation is still essential.
Additionally, AI models sometimes fall short in explainability which is critical for compliance-driven industries. If the decision-making logic behind AI-generated recommendations isn’t transparent, regulators and internal teams are reluctant to trust or adopt those solutions fully.
Finally, integrating AI into legacy systems and workflows remains a challenge. Without full end-to-end automation, the benefits of AI can be fragmented, requiring manual intervention at key points and limiting overall impact.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
Integrating AI into product development can be both transformative and challenging. One of the primary hurdles has been designing UX that strikes the right balance between automation and human control. Users often want the efficiency AI offers, but they also need transparency and the ability to override or audit AI-driven decisions.
User trust has also required deliberate attention. Many users are cautious when it comes to AI interpreting legal or compliance-sensitive documents. To address this, it is crucial to focus on explainability offering visibility into why an AI model made a certain suggestion, and building features that allow easy validation or rejection.
Lastly, from an engineering standpoint, integrating AI into existing systems, especially legacy infrastructures, might pose architectural challenges. AI models often require different data pipelines, real-time processing capabilities, and security protocols, which might be required to be rebuilt to ensure smooth deployment and scalability.
Os Hourani, Marketing Manager at Woolpert
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
In my consulting work with clients around the world, I've seen significant improvements in client engagement through the integration of AI tools such as Microsoft Copilot, AI-powered CRM systems, ChatGPT, Stable Diffusion, and Midjourney. Microsoft Copilot and custom GPT's have enabled teams to efficiently generate highly tailored communications, making each interaction more meaningful and impactful for clients.
Additionally, implementing AI-powered CRM has enhanced our ability to analyse customer data, predict client preferences accurately, and proactively shape customer journeys. Stable Diffusion and Midjourney have supported creative efforts by enabling high-quality, customised visual content, further enriching the client experience. These combined technologies help deliver personalised, timely interactions at each touchpoint, substantially enhancing overall client satisfaction.
Utilising these AI tools to automate routine tasks allows marketing and customer relationship teams to prioritise strategic planning, relationship-building, and innovative storytelling. Ultimately, adopting these advanced AI solutions has driven greater efficiency, more effective marketing strategies, and significantly improved engagement outcomes.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
AI agents and generative AI solutions are areas where AI has not yet fully delivered on expectations, despite considerable hype. AI agents, while promising, often fall short when nuanced human judgment or complex decision-making is required. Their capabilities in autonomous decision-making and contextual understanding still lag behind expectations.
Similarly, generative AI solutions, while highly effective in general-purpose scenarios, frequently struggle with delivering consistently accurate and high-quality outputs in niche or highly technical industries. The need for precision, domain-specific expertise, and context-sensitive content means these AI tools still require significant human input, review, and oversight to achieve truly useful outcomes. Consequently, the real-world application of AI in these specific contexts has yet to match the excitement and potential often advertised.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
Working with high-profile clients has highlighted significant challenges around maintaining user trust through robust AI policies and clear guardrails to protect client data and ensure privacy compliance. Establishing these measures has been critical for building and sustaining user confidence.
A key area of focus has been creating transparent, natural, and intuitive user experiences. It is essential for users to clearly understand whether they're interacting with AI or human agents. With rapidly advancing AI capabilities, maintaining this transparency is increasingly complex yet remains crucial for ethical user engagement and trust.
Additionally, addressing occasional AI inaccuracies or inconsistencies has required continues monitoring and refinement. Ensuring AI reliability through rigorous testing and quality assurance processes is essential to maintain user confidence and deliver consistent results.
Syed Bilal Ahmed, Co-Founder and CTO at AstrumAI
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
I integrated AI into Microsoft plugins (e.g., NLP for auto-tagging), predictive analytics models, and NLP/computer vision systems. These improved automation, reduced manual effort, and enhanced decision-making for users across productivity, planning, and data processing workflows.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
AI still falls short in areas requiring deep contextual understanding, such as nuanced human conversation, ethical decision-making, and generalising across domains. It also struggles with real-time edge deployment in low-resource environments due to computational demands.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
Key challenges included balancing model accuracy with real-time performance, designing intuitive UX around complex AI behaviour, and building user trust, especially when AI decisions lacked transparency or explainability.
Zaid Ahmad, GTM Specialist at Qapita
In what specific product features have you integrated AI, and how has it improved the user experience or core functionality?
I helped build an AI-driven outbound engine that activated real-time, personalized outreach based on high-intent buyer signals. We used a multi-tool integration of Clay, RB2B, Factors.ai, and Instantly to bridge the gap between intent, qualification, and outreach:
Factors.ai gave us company-level website visitor intelligence, surfacing firmographic details (e.g., industry, headcount, location) and behavior patterns (pages visited, scroll depth, repeat sessions). This helped identify which companies were showing buying intent, even if individuals were anonymous.
RB2B layered on person-level identity resolution, matching visitors to decision-makers within those companies. This was critical for mapping engagement from account to persona and segmenting based on ICP criteria (role, seniority, function).
Clay served as the orchestration hub — enriching contacts in real time, qualifying them against our ICP filters (e.g., Head of Finance at a US-based SaaS company with >50 employees), and dynamically generating personalized email copy using OpenAI and contact-level data points (job changes, tech stack, etc.).
Finally, Instantly automated the outreach — launching trigger-based campaigns the moment high-intent signals fired (like demo views, 409A valuation pages, or repeat visits from a single account).
This AI-enabled system transformed our lead flow from passive form fills to proactive, hyper-targeted engagement, significantly improving response rates and pipeline velocity. It felt like productizing GTM intuition — but at scale.
What are some product areas where AI has not yet delivered the expected impact — or where it's still falling short?
While LLM-based personalization is powerful, it sometimes outputs overly generic or verbose copy if the input signals aren’t sharp. Fine-tuning for crispness and relevance at scale remains a challenge.
What challenges have you faced when integrating AI into product development — whether in terms of UX, performance, or user trust?
The biggest challenge has been balancing automation with human-like messaging. While AI helps scale outreach, maintaining trust and context in messaging — without sounding robotic — took experimentation.