By Aakriti Bhargava / 12 Jun 2026 / Topics: Artificial Intelligence (AI) , Application development , Modern infrastructure , Generative AI
Building AI agents for production reliability requires balancing three competing forces — security, cost, and reliability — and the right answer keeps shifting as models improve. Revionics, a 20-year AI pricing company serving enterprise retailers globally, found that integrating generative AI into their existing product was the easy part. The hard part was something their two decades of applied AI experience didn't fully prepare them for: an entirely new category of customer security concerns.
Aakriti Bhargava, VP of product engineering and AI at Revionics, walks through the architecture decisions her team faces daily when building agentic systems for production. The central tension is how much autonomy to give AI agents when security, cost, and reliability all pull in different directions. Her team is currently thinking in terms of a 70/30 hybrid model — build deterministically for the use cases you know will always occur, then let the LLM handle the edge cases. That approach increases speed to market, reduces cost compared to letting the LLM handle everything, and keeps the system reliable enough for enterprise customers making real pricing decisions. But Bhargava is clear this is evolving — what felt like a settled question at the end of last year has already shifted as the models improve.
The conversation also challenges a common assumption about the generative AI landscape. Revionics chose not to fine-tune foundation models, instead relying on Google's Gemini and focusing engineering effort on the agentic system architecture around it. The reasoning is straightforward — there's already a race at the foundational model level among Google, Anthropic, and OpenAI. The differentiation for applied AI companies lives in the engineering that makes those models reliable, secure, and cost-effective in production.
On the internal productivity side, generative AI has compressed RFP response timelines from three to four weeks down to days and helped address 20 years of accumulated documentation debt. But Bhargava is candid about the tradeoffs. Code generation has accelerated while code review has become a new bottleneck, and junior engineers risk building dependency on AI tools before developing the foundational problem-solving skills that make senior engineers effective. Her position is that productivity gains are real but uneven across experience levels, and organizations need to find the sweet spot for each individual rather than applying a blanket approach.
This is the third and final episode in a three-part series on building the infrastructure foundation that makes everything else possible. Where the first two episodes focused on migrations, this episode goes one layer up — into the product and engineering decisions that turn solid infrastructure into AI capabilities enterprise customers can trust.
If you liked this episode, share it with a colleague.
Have a topic you’d like us to discuss or question you want answered? Drop us a line at jillian.viner@insight.com
Aakriti Bhargava
VP of Product Engineering and AI, Revionics
Audio transcript:
We’re in the process of finalizing the transcript. Please check back soon for the completed version.
Marketing Manager, Insight
As marketing manager for the Insight brand campaign, Jillian is a versatile content creator and brand champion at her core. Developing both the strategy and the messaging, Jillian leans on 10 years of marketing experience to build brand awareness and affinity, and to position Insight as a true thought leader in the industry.
VP of Product Engineering and AI, Revionics
Subscribe to our podcast today to get automatic notifications for new episodes. You can find Insight On on Amazon Music, Apple Podcasts, Spotify and YouTube.
