Telling a client that they should rethink their plans

We talk to many prospective clients, and hear dozens of potential use-cases for artificial intelligence agents, machine learning models, and data platforms. This puts us in a unique position to be able to see patterns in how people want to use these tools. We also see common traps that people fall into when building new products and models.

In this case, a client came to us with an idea for a data platform that would serve as an integration hub for accounting companies. It would enable the users to pull data from different software applications that they use, and use an AI agent to perform work across these applications. Their reasoning was compelling, and some of their clients wanted exactly this product. The only problem was, we had heard this exact idea, for different industries, three times in the previous month. All four of these proposals had the same issue: there is no moat.

The Problem

When we look at it as a one-off data platform for industry X, it seems like a good idea. A lot of people across this industry use these three software applications, and they want to be able to integrate them into an AI agent and interact with all three in one interface.

Where the idea breaks down

What we noticed over the first three groups that brought us this idea was that almost all of these software applications had built, or were building, integrations to ChatGPT, Claude, or both. These companies are quite literally the perfect platform for integrating software applications into an AI agent, and they have no overhead. OpenAI and Anthropic will always be able to integrate with more software applications than a small startup, and they will be able to do it faster and cheaper. What we came to realize was that the age of integration hubs are essentially over, unless there is a very specific reason why it would be difficult or impossible for OpenAI or Anthropic to allow a given application to integrate with their platform.

Giving the bad news

As much as it pains us to tell our clients that their idea will probably not work in the long term, we have often been successful in conveying the bad news. In this case, we showed our client the integrations and model context protocol (MCP) servers that already exist for the large accounting software applications. We showed them how easy it was to integrate these applications with a ChatGPT or Anthropic account. By the time we were done explaining that these integrations already exist, our client understood what we were about to tell them.

New sparks of light

Our clients rarely just give us a two-sentence idea with no context. In this case, we were given multiple case studies of our client's customers, their workflows, and how our client wanted to improve them. The rejection of this broad integration platform for accounting services gave our clients more clarity on which parts of these workflows are more valuable, and which parts they can focus on to attract more customers. In the case of accounting services, there are numerous parts of workflows that still, and will for the foreseeable future, require a human in the loop to manage and judge the work of AI agents, as well as to interpret very specific rules and real-life situations.

Our client has pivoted to a more precise product that really does save their customers time in ways that ChatGPT or Claude will be unlikely to solve through integrations and LLM improvements. Human-in-the-loop processes and workflows ingrained with domain-specific accounting knowledge should keep their customers dug into their product for years to come.