A founder develops a list of features, hires a team and creates an application that includes a chatbot three months after the development began. The chatbot provides incorrect responses. It costs more money per user to provide service than each user pays. No one on the team knows why. Many times this scenario occurs in AI startup development. Often the underlying concept is solid.
However, the team simply did not possess the appropriate set of technical skills at the outset of the project. This article will discuss the technical skills required by an AI-ready MVP team and how to assess if potential team members possess those skills prior to hiring them.
An AI MVP is Not a Normal MVP With a Model AddedThe team building an AI-based MVP needs to understand that it is a very different kind of MVP than one which uses AI. A typical MVP will have its code do the exact same thing each time someone clicks on a button or performs some other action.
For example, if you click on a button in your MVP you should always receive the same response. An AI-enabled feature, however, will behave differently. It may provide different responses to the same question depending upon when the question was asked. Also, when you update your AI model it can drastically alter how your feature responds to questions.
Finally, each time your AI model is called it will cost real dollars.
Therefore, the team working on this type of MVP must be prepared to deal with incorrect responses, slow responses and expensive responses from day one. This preparedness to deal with potentially incorrect responses is what we refer to as being “AI-ready.”
The Skills That Matter
You have to select and test which models (e.g., OpenAI, Anthropic, Open Source) work best for your use case and run tests using your own data. The best-known model is not always the right choice. A smaller, less costly model may perform just fine.
- Prompt Creation & Structured Output
It takes time to develop good prompts that are specific, well-tested, and tracked in version control (like code). Teams should understand how to get a model to generate clean JSON outputs for use in other parts of their application.
Many AI applications rely on retrieving information from your data sources (e.g., help articles, contracts, product catalogs). To achieve this, you will need to chunk your documents, create embeddings, and store them in a system such as Postgres with Pinecone. Most of the "the AI made it up" complaints result from poor retrieval.
This is the skill many teams do not possess. Before launching your application, create a collection of 50 to 100 actual examples with correct answers. After creating your initial collection, every time you update a prompt or a model, run your evaluations again to determine if you improved or hurt your performance.
Although an AI feature can provide value to your customers, it still requires integration into your existing backend infrastructure. Therefore, strong programming skills in Python or Node.js, queueing mechanisms for long-running processes, and well-defined APIs are critical to the success of your project.
- Cost & Speed Optimization
Start monitoring token usage from day one. Implement caching for frequently requested answers. Utilize a low-cost model for simple requests. A feature that costs per user will quickly destroy a startup that charges less than per user.
Determine early on which types of data you intend to send to a third-party model. Customer records, health data, and payment information should be clearly defined prior to launch.
In-house or MVP Developers for Hire?
The best solution for most new companies would be to hire an outside development team. It takes a long time and a lot of money to assemble a full-fledged in-house AI team.
Finding engineers that are experienced with LLMs is difficult, and as a founder you will likely not know what to ask them in an interview to determine whether they would be a good fit. Many entrepreneurs choose to use MVP (Minimum Viable Product) developers.
These are teams that have successfully created other applications, and they can begin working on yours within weeks. There is a catch though; you need to clearly define the scope of the project and who owns the completed application so that when the project is over you retain both the source code and the knowledge gained during the development process.
In some cases, a combination of an outside development team building the initial version of your product and an in-house engineer joining them as soon as possible to take over once the product is complete has proven to be effective.
Questions to Ask Any MVP Development Services Provider
Before signing on the dotted line, you should be asking the following:
- What products have you created using AI features which are currently in production?
- When building those products, what was the biggest issue you encountered?
- How will you determine whether your AI model's responses are accurate enough?
- How will you manage the cost of maintaining your AI models?
- Whose property will be the prompts, code and data once you complete our project?
- If the model service provider raises its prices or discontinues a model, what will happen?
- If you get vague answers to these questions from potential providers, that is a red flag.
A team who has built products with AI features will speak in terms of specific tools, specific issues they faced and specific metrics related to their success.
A Realistic Eight-Week Plan
The following is a simple 8 week plan to build a Custom MVP using AI as its core function.
Choose one user problem. Obtain actual data and produce your first 50 test case examples. Consider testing two or three different types of model.
Construct the core workflow and the AI functionality. Each time you make changes to the system run all tests.
Include logging, cost tracking mechanisms and a method for end-users to indicate if they received an unsatisfactory response from the AI. Test this version of the product with between 10-20 actual users.
Address the most significant issues, implement usage limits and release the product to a limited number of users. It should be noted that there are no additional features included in this planning schedule.
Most AI MVPs ultimately fail because they try to accomplish too many things. One successful feature that works consistently outperforms five features that work sporadically.
What To Do Next?
In order to develop your AI product this quarter, begin by identifying one problem (e.g., answering customer service inquiries) and establish a single measurable objective (e.g., answer 80% of support questions correctly; reduce review time by 50%). Create the test set prior to developing the feature. Prior to scaling the solution, determine the cost per user.
If you would like a
MVPDevelopment team to help you accomplish this, the
WeblineGlobal team can provide AI MVP development for startup companies and product teams.