Instead of pursuing an all-encompassing, high-visibility AI procurement, there are likely small, low-profile AI enhancements you can immediately pilot without a significant investment that are less risky and can provide immediate benefits. To put it in California terms: if you want to save money on gas, try replacing your car tires instead of buying a whole new car. We call this “Adding AI the boring way” and offer this checklist to identify simpler and quicker ways to implement AI without significant investment:
1. What is the unmet need that AI or an LLM can solve?
As always, start with a problem statement rather than the solution. What specific process or workflow needs to be improved? And for the first few AI projects, are there any internal-facing workflows that AI could improve first, rather than adding AI functionality to public-facing or similar high-visibility processes - like building an AI-driven chatbot to provide compliance information? For example, what back-office processes need to be expedited or require significant resources (e.g., staffing, budget) right now?
2. Is an AI/LLM tool the only way to solve the unmet need/problem?
Consider whether there are simpler tried-and-true options. By evaluating and eliminating those alternatives, you’ll be better able to defend whatever decision you make.
3. Research which AI tools are already accessible and cost-effective.
As there may already be a list of approved AI tools (e.g., Copilot for Microsoft environments) with the required licenses, first find out which are currently authorized (and not authorized) for use by your organization. Leveraging current AI models and tools that can easily be integrated or work with your existing IT environments and systems helps reduce cost and lower the risk of failure. It also greatly simplifies compliance with existing security or AI-use policies easier and allows for faster deployment.
Second, given the multiple options out there, try out different AI tools and LLMs to identify which combination is best suited to address the specific unmet need or problem. For example, look for Generative AI tools when you need to create new content or Discriminative AI tools if you need to identify or categorize existing content. In addition, test out different versions of the common AI tools (e.g., Chat GPT, Copilot, and Gemini), including a free versus paid version, to see how results may differ.
4. Start with the smallest, lowest-risk idea
Just as with any other IT decision, pick the most boring option. That is, line up the combination of tools and use cases that produces the least uncertainty for development time, budget, and data protection.
How A1M Solutions Added AI the Boring Way
By using the checklist above, A1M Solutions was able to add AI and Machine Learning (ML) functionality to an existing policy knowledge management solution we built and maintain for a health and human services agency without any additional funding.
Evaluating what AI could help us solve: We started by reviewing past user research and our backlog to identify where AI could help us. Since Generative AI can quickly summarize existing content, we first considered implementing AI to provide our users summaries of the lengthy and complex policy documents they relied on to make decisions. However, we quickly ruled that out because our users valued and trusted our product as a source of truth for decisions that affect public services and benefits and prioritized accuracy over simplicity. Any inaccuracies from AI-generated summaries may not only have led to incorrect conclusions or wrong policy decisions, but also a loss of trust in our product or fear of using AI altogether.
Therefore we investigated where AI could help improve our back-end processes and free up our developers’ time to tackle more complex backlog projects. We decided to explore using AI to improve search results with documents that were not in easily searchable formats.
Was AI necessary? By all metrics, our search indexing worked well and our developers had automated as much as possible. However, if we could use AI to further improve search results and also free our developers from routine and repetitive coding tasks, they would have additional capacity to work on more complex problems or new features.
Evaluating AI tools: We knew any AI tools had to meet the agency’s security requirements and work with our users’ IT environment. However, investigating low-cost AI tools to improve search was more complicated. No single LLM, even ones dedicated to better information retrieval, did everything we needed.
Our small, low-risk options: So, we incorporated AI as discrete components rather than implementing a single, unified solution. We found and implemented: 1) Google Magika (a free, open-source file-classification tool) to detect file types and route documents to the appropriate text extraction tool and 2) Amazon Titan Text Embeddings plus AWS RDS Aurora which got us beyond simple keyword matching to truly semantic search.
Our results: Working methodically, we added better search, better file processing, and development assistance by integrating readily available services. Starting with small, low-risk updates let us experiment with and deploy features that would have required more significant engineering time and budget just a few years ago.
The Benefits of Adding AI the Boring Way
It can be hard to see past the hype about AI. What we found is that there is rarely an all-encompassing AI solution to every possible problem. Moreover, investment in a high-profile, high-stakes “AI solution” may be costly and risky – to the organization as well as the users. Instead of looking for the perfect AI solution, try experimenting with adding different AI or ML tools to solve some parts of some problems of your existing systems right now.
Want to have an interesting talk about boring AI? Get in touch.
