Given the abundance of open-source and proprietary LLMs on the market, choosing the right one for enterprise deployment is still a make-or-break situation for many. The choice directly affects an organisation’s scalability, performance, and compliance.
As a result, selecting an LLM goes far beyond evaluating benchmark scores. It requires a holistic assessment of how well a model aligns with business objectives, integrates with existing technical infrastructure, and satisfies regulatory and compliance standards.
Some of the leading LLMs for enterprise use include OpenAI’s GPT family (GPT-4o, GPT-4.5, GPT-3.5 enterprise), Google’s Gemini 2.0 Pro and Gemini for Workspace, Anthropic’s Claude 3 series (Opus, Sonnet, and Haiku), Meta’s Llama 3 and the upcoming Llama 4.
However, scaling GenAI across an enterprise demands more than just picking a performant model. Organisations must carefully compare each model’s strengths, unique features, and pricing structures, while also weighing critical factors such as security, data privacy, operational control, and ease of integration into existing systems.
“So the question for us was how do we scale this (ChatGPT) technology and allow our team members to safely experiment with it and rapidly build applications that power our business value through its use,” Natarajan Ramamurthy, VP of data engineering at Target, said at DES 2025.
This discussion led to the creation of Target’s GenAI platform, Think Tank. Target’s small team of engineers hosted an open-source model, Llama, on Target’s data centre and built a simple chat application. Once it demonstrated real business value, the focus shifted to scaling the solution.
Learning to Scale
At the foundation of Target’s Gen AI platform was the model garden. It was designed to provide controlled access to a variety of LLMs and SLMs—both proprietary models hosted in the cloud and open-source models running on Target’s own GPU stack. Ramamurthy pointed out that this gave teams the flexibility to choose the most relevant model for their use case.
Target recognised that LLMs need context to generate meaningful outputs and, hence, developed grounding services. These services allow teams to inject, manage, and guide the use of contextual data within responses.
To ensure a consistent brand voice across use cases, prompt services were introduced to enrich prompts with tone and brand metadata, and optimise token usage—helping with prompt compression and token optimisation to manage costs, Ramamurthy mentioned.
Given the risks associated with GenAI, such as hallucinations, bias, or security threats, moderation and evaluation services were established to scrub harmful content, mask personally identifiable information (PII), and ensure safe, policy-aligned responses, in close collaboration with cybersecurity and responsible AI teams.
As many use cases are real-time and guest-facing, Target mentioned that performance management tools were also built in, offering tenants observability to track, optimise, and throttle usage as needed.
Cost is a major consideration, and quota management services provide detailed visibility into the cost of each LLM interaction. Ramamurthy added that they “provide visibility to developers, product managers, and business leaders alike”.
Moreover, to promote a culture of innovation, Ramamurthy revealed that Target launched the Think Tank AI Studio, a safe experimentation space with a monthly $10 token quota. This space enables developers, product managers, and business users to test ideas, validate them quickly, and scale the most impactful ones, ensuring “fearless experimentation” across the organisation.
Target’s AI Shopping Assistant
One of the first use cases built on Think Tank was the AI Shopping Assistant on Target.com, which helps guests make informed product choices.
“Initially, we only had the model garden. We experimented with various LLMs to find the most relevant responses, but had limited product data,” Ramamurthy stated.
Target’s team quickly realised that the key to improving the AI assistant’s relevance and accuracy lay in grounding. Their grounding services integrated structured data from product attributes, FAQs, and guest reviews. “This improved the assistant’s contextual awareness,” Ramamurthy said, adding that this enables it to deliver responses that are more aligned with guest expectations.
As development progressed, routing challenges surfaced, particularly when users posed questions like “Where is my order?” Initially, the assistant could not handle such queries. To solve this, the team implemented intent classification, allowing the system to route questions to the correct internal services.
Ramamurthy further mentioned that this was, in effect, the early signs of agentic abstraction, signalling the platform’s gradual evolution toward more autonomous, goal-oriented functionality.
Moderation also became a critical focus area, especially for managing inappropriate or off-topic prompts. This approach not only safeguarded the brand but also reduced costs by minimising unnecessary LLM calls.
Given that this was Target’s first guest-facing GenAI use case, the risk posture was deliberately stringent. The company’s evaluation services assess the assistant’s responses, using a combination of real customer queries and human-in-the-loop sampling.
Ramamurthy revealed that one of the most valuable outcomes of this iterative development process was the creation of model benchmarking capabilities. The platform gained the ability to compare multiple LLMs based on performance and fit, opening the door to dynamically selecting the best-performing model per use case.
What are Others Doing?
Meanwhile, Wayfair has launched a new AI tool called Muse, which helps customers visualise how furniture might look in a specific setting. Wayfair uses Gemini on Google Cloud to automatically categorise products across its 30 million product catalogue, cutting the time to curate new listings and update existing ones by 67%.
Fiona Tan, CTO of Wayfair, in an interview with AIM, said, “The company is currently experimenting with several LLMs, including those from OpenAI, Google, and Anthropic. Gemini performs better for catalogue enrichment, ChatGPT is more effective for customer responses, and Claude is preferred for coding tasks.”
In 2022, IKEA launched IKEA Kreativ, a tool that helps customers design and visualise their rooms in 3D on any device.
Last October, Walmart also shared its plans to use AI, AR, and immersive tech. This included Wallaby, a set of AI models built to improve customers’ interaction with its platform.
At the same time, Amazon has introduced its own AI shopping assistant, Rufus.
The post How Target is Building its AI Think Tank appeared first on Analytics India Magazine.
