Beyond GPT wrappers: Making AI work for your business

5 strategies that make AI applications deliver real impact

Summary

As we spent the last year experimenting with AI applications, we've noticed a fascinating pattern: while most businesses struggle to extract real value from ChatGPT, some are quietly transforming their operations with AI. The difference is not about having your own foundational model - it's about knowing how to bridge the gap between raw AI capabilities and real business needs.

In this post, we'll talk about five tested approaches that turn basic GPT into powerful business applications. Whether you're leading a startup or running a Fortune 500 business, these practical strategies will help you move beyond the "GPT wrapper" debate and start delivering actual results with AI.

---

"That's just a GPT wrapper!"

If you've followed AI startups lately, you've heard this dismissive response. The criticism suggests these companies add no real value beyond using OpenAI's API and adding an interface on top. Critics argue you need your own foundation model or at least a fine-tuned one to build a viable AI business.

Having spent the last year experimenting with AI applications, we've found the reality more nuanced. While there's been a flood of basic ChatGPT repackaging, successful AI applications solve real business problems that off-the-shelf ChatGPT can't: having an intuitive UI, integrating business context, enabling system actions, maintaining conversation memory, and embedding into existing workflows.

Let's break down what’s a GPT wrapper and what are its limitations, how to move beyond basic implementations, and how to evaluate AI solutions for your business.

What’s a GPT wrapper and what are its limitations? 

A GPT wrapper is an application that uses a pre-defined prompt, makes an API call to GPT and packages the response into a user interface. You can try this out with OpenAI's custom GPT feature, where you can write a prompt to set up a wrapper. 

Custom GPTs at OpenAI’s GPT Store

While working on AI applications, we've identified five core limitations when using ChatGPT out of the box:

• Cumbersome interfaces: Requires extensive prompt engineering and typing

• No business context: The model speaks generally but can't access your company's specific data

• Can’t take actions: Generates text but can't take actions in your system

• Memory gaps: Forgets crucial context between conversations

• Siloed operation: Doesn't integrate with your existing workflows

There are proven solutions to each of these without needing your own model.

How to move beyond basic implementations?

Cumbersome interfaces → Intuitive UI/UX 

A common pain point with using ChatGPT is having to type out long prompts. Instead of requiring users to master prompt engineering, well-designed AI applications handle this complexity behind the scenes. Below is an example of how UI/UX can deliver value:

Example: AI Writing Assistant with Intuitive UI/UX

Step 1: Offer buttons or dropdowns for common use cases like "Write an email," "Create a proposal," or "Draft a social post for LinkedIn" 

Step 2: Break down the task into simple fields - audience, tone, key points, constraints 

Step 3: System automatically converts user inputs into an optimized prompt 

Step 4: Generates the output step by step and allows the user to provide feedback, say through preset options and toggles (concise vs. detailed, professional vs. conversational, technical vs. layman language, etc)

Some common use cases for intuitive UI/UX are:

• Content creation: Marketing materials, social media posts, email campaigns

• Document generation: Complex templates like legal contracts or technical documentation

• Data analysis: Report generation and insights extraction

The key benefits of intuitive UI/UX are:

• Ease of use: Makes AI accessible and quick to use to non-technical users who are not familiar with advanced prompt engineering

• Consistency: Ensures reliable outputs by standardizing how users interact with the AI

No business context → RAG

RAG (Retrieval-Augmented Generation) is a great technique to provide GPT your business context. Instead of expensive model fine-tuning, RAG lets you feed company-specific information to GPT on the fly, which GPT uses as additional context to generate its response. Swedish fintech Klarna saved $40M in costs by automating customer support and using RAG as one of the key levers!

Here’s how RAG could work in an e-commerce scenario:

Example: E-commerce Customer Service Chatbot with RAG

Step 1: Build your knowledge base: Upload your latest company policies and documentation 

Step 2: Retrieve relevant context: When a customer asks about returns, the system pulls your specific return policies 

Step 3: Augment GPT's knowledge: Append or add this context to the customer's query 

Step 4: Generate a response: GPT answers the customer’s question by using the additional context, not the generic knowledge it has been trained on

Some common use cases for RAG are: 

• Customer support: Responding to customer queries based on your policies

• Employee assistance: Responding to employee queries based on your HR policies

RAG has three key benefits:

• Factual accuracy: Responses are grounded in your current data, not outdated training data

• Competitive edge: If you’re building an AI application, this proprietary data becomes your moat

• Compliance: Stays aligned with your official policies and regulations

Can’t take actions → Function calling

Function calling is what can turn GPT from just a chatbot into a doer. It allows the AI to trigger real actions in your business systems. Back to our ecommerce example: 

Example: E-commerce Customer Service Chatbot with Function Calling

Step 1: Define Available Actions: Create a menu of functions GPT can use (e.g., check payment status, check order status, initiate refund, update shipping address) 

Step 2: Analyze User Intent: When a customer asks "Where's my order?", GPT recognizes this requires an order status check 

Step 3: Execute the Right Function: GPT calls your order tracking system with the relevant order ID 

Step 4: Generate the Result: Customer gets the real-time order status, not just a generic response

Some example use cases for function calling are:

• Financial analyst Assistant: Execute actions like fetching updated market data

• Customer service: Execute actions like processing refunds or updating orders

• Sales operations: Update the CRM, schedule meetings, or generate quotes 

Function calling has two main benefits:

• Action-oriented: Makes GPT move beyond answering simple questions to completing tasks

• Scalable: Seamlessly integrates the AI into your existing business processes

Memory gaps → Memory

If you use ChatGPT, you’d have experienced how ChatGPT uses memory. Memory helps GPT remember important context from previous conversations. 

You can implement various types of memories depending on your use case: short-term (recent conversations), long-term (user preferences), episodic (specific past interactions), or semantic (domain knowledge) memory. 

Consider this Sales Assistant example to understand how to use memory + RAG: 

Example: AI Sales Assistant with RAG + Memory

Step 1: Build Knowledge Base (RAG): Create a database of things that don’t change often, such as your clients’ profiles, your product documentation, pricing sheets, historical deal data, market data (competitor analysis, industry reports) 

Step 2: Set Up Memory: Add any dynamic knowledge here, such as recent meeting notes, key discussion points, client’s communication style, personal rapport, etc

Step 3: Context Retrieval: When preparing for a client meeting, the AI pulls both RAG and Memory to understand the full context

Step 4: Personalized outreach: The AI combines these to create relevant outreach - "Given your security concerns from our last chat, here's how our latest release addresses them.."

Some common use cases of memory are:

• AI sales coach: Remembers the sales rep’s strengths, common mistakes and personalizes the guidance

• AI tutor: Leverages memory to tailor the content to the learner’s current proficiency, remembers their weak areas and provides more practice in those areas

The main benefits of memory are:

• Context: Helps retain key information like past decisions, user preferences, etc.

• Personalization: Personalizes the response to deliver high value to the user

Siloed operation → Combining LLMs with deterministic logic

LLMs can deliver high value when you integrate them with deterministic business logic (rules). This guardrails the output, making it predictable and scalable. Let’s see this through an example:

Example: B2B Lead Qualification Using LLM Plus Deterministic Logic

Step 1: Lead Qualification (Rules): System checks each inbound lead against your fixed criteria - company size, industry fit, budget signals, and decision maker level

Step 2: Personalization (AI): For qualified leads, GPT analyzes their website, LinkedIn, and news to identify pain points and craft personalized outreach.

Step 3: Email Creation (Rules + AI): GPT drafts email using approved templates while the system verifies all compliance rules and checks against the do-not-contact database

Step 4: Automated Delivery (Rules): System schedules emails following your exact rules - business hours in lead's timezone, frequency caps, etc

Some common use cases of LLM + deterministic logic are:

• Lead qualification and routing for sales teams

• Support ticket escalation for customer support teams

The key benefits of combining AI with deterministic logic are:

• Scale with control: Automate high-volume processes without sacrificing compliance or quality standards

• Ease of audit: Clear visibility into every decision since the system follows documented rules

Now that we've covered the key solutions, let's look at how to evaluate AI applications for your business.

How to evaluate AI applications?

When evaluating AI solutions for your business, look beyond the "wrapper" debate. The real question isn't whether an application is built on top of GPT, but how effectively it solves your business challenges. The most valuable AI applications will:

• Have intuitive interfaces that remove the pain of lengthy prompt engineering for the user

• Build knowledge assets specific to your business through RAG and usage data

• Integrate seamlessly with your existing workflows and systems

• Remain flexible to swap between different foundational models as the technology evolves

• Adapt to improvements in underlying models e.g., start with human in the loop and eventually move to full automation as the models improve

Final thoughts

As Sam Altman noted, successful AI applications will be built assuming models will keep getting better. The key is choosing solutions that focus on your business problems while leveraging – not competing with – rapid advances in foundation models.

Start by identifying which of the above approaches could most impact your business goals, then build from there.

In our next post, we'll explore AI agents and what they mean for business operations. Until next time!