Asking IT leaders to be hands-on with AI in some way could be a stretch, but what they should at least have is an informed perspective and strategy for this change in the external environment.
This post by Levie on topics leaders are thinking about when deploying AI Agents, was great in framing 8 of the major areas to consider when thinking about how to approach AI as an enterprise, especially if you don’t have a huge budget to compete on Compute or Algorithms.
I’ll summarize his concerns, and share my current position from exploration over the last few years.
- Agent interoperability:
- Concern: Enterprises don’t have their data in just one environment. We need to design more interaction patterns for how AI Agents talk to each other and exchange data in the future.
- Ajared: A lot exists on this front. Middle ware drivers like ODBC and JDBC and now MCP can help loosely couple applications. Apps from the web 2.0 era have broken interoperability, the rise of agents give another reason to solve it. https://www.linkandth.ink/p/apps-break-data
- Over-permissioned systems.
- Concern: Liability is increased as AI can instantly access information if permission and access controls are not secure.
- Ajared: https://x.com/kidehen/status/1904729170405437758
- Preparing enterprise data for AI
- Concern: Systems have important data, but not in environments that AI agents can talk to. Retrofitting will most certainly not work.
- Ajared: Modernization and cloud computing do not on their own solve the problem of having enterprise data AI ready. We have seen limitations of approaches like Datawarehouse/lake/ any similar solutions to bring all data into ‘one place’. We should be doing the opposite – connecting data across silos – using proven distributed technology of the web to integrate enterprise data at scale. Organizing your enterprise Data using a knowledge graph is another paradigm for focus. https://www.knowledge-graph-guys.com/blog/why-use-a-knowledge-graph
- Compliance:
- Concerns: Industries don’t have an official stance on where AI can provide suggestions or make decisions, some might need regulatory clarity.
- Ajared: Let’s think about “agents” more as bundled capabilities, than sentient entities. While we wait, even regulated industries can work on semantics to encode decision making and constraints for their specific use cases, and share and standardize where needed. I suspect working groups already exist for a lot of industries, an example is https://idea.rpi.edu/research/projects/heals
- AI Agents doing the work
- Concerns: Human in the loop is increasingly impractical for all workflow steps. Companies need policies for hand-off to agents.
- Ajared: Bullish on human in the loop, and techniques to augment agents with representations of human knowledge. Even as agents learn and become ‘smarter’, why learn what you already know? There’s also a chance for humans to get smarter about handling the non-agentic steps of their workflow. Policies will adapt to the work on ground. https://pod.co/the-connected-data-podcast/systems-that-learn-and-reason-frank-van-harmelen
- AI model quality:
- Concerns: Model quality is still risk for high value workflows. We’re not 100% there for production.
- Ajared: There’s already use cases in production, and a healthy middle ground where instead of vibe coding, professionals can use AI as another tool. We have other variables to improve outputs. Retrieval Augmented Generation is yielding many benefits.
- It’s safe to assume there’ll be continuous improvements in model quality for the foreseeable future. Unlike most elephant fights where the grass suffers, smaller companies have everything to gain from the ongoing breakthroughs with models. Models are increasingly being commoditized, and we seem to be trending to a buyers market.
- Change Management:
- Concerns: Will take the longest. Expectations of radical transformation can’t happen overnight. Companies are still working on roadmaps for deploying AI for the largest impact-change ratio.
- Ajared: Nothing jumps out about current AI trends that breaks the expected change management model for disruptive technologies, but need to think about this some more.
- Business Models of AI Agents
- Concerns: Still early in getting to ultimate pricing model for agents. Need stability.
- Ajared: The approach should be to loosely couple agents together to reduce costs instead of getting locked into any vendor solution. Open source will play a huge role here. While I still pay for some models, i’m always on the look out for open source equivalents. Just like hobbyists assembling computers in the early days, IT teams should take a hobbyist approach and mix building and buying to get to production on a budget.