David Robertson is Director of Enterprise Architecture, Software Engineering, Applications at Exeter Finance. With deep expertise in software strategy, digital transformation and enterprise-scale AI deployment, David focuses on aligning emerging technologies like generative AI with real-world business outcomes. His work blends architecture, governance and team development to create scalable, secure and user-focused technology solutions.
Unlocking AI Value with Strategy and Skill
In a previous article, I wrote about how to integrate generative AI into enterprises1. Interestingly enough, many of those same principles still apply today. Why is that so? Hasn’t AI adapted and changed a lot in months rather than years? Or is it that the software extrapolation of Moore’s law is somehow irrelevant in today’s world of technology? Read on to see what has stayed the same and what has changed recently?
Retrieval augmented generation (RAG) is now assumed, allowing enterprises to ground AI using their proprietary data, making it more meaningful for their purposes. Testing and validating AI outputs often still requires a “human in the loop” to either make the final decision or review the outcomes. Given large language models (LLMs) are stochastic, few tools are designed for such a thing, much less in an automated way. Agents can use tools to perform automated tasks on behalf of a user, allowing for automation, similar to traditional bespoke software development or robotic process automation (RPA).
Recent Advancements and New Capabilities
For automated testing of LLMs, new tools are coming from specialized vendors that allow teams to validate AI using existing people and skills; some LLM models even advertise the ability to “self-correct” over time, a capability similar to traditional machine learning training.
Risk, governance and security are big concerns, especially if using RAG. You do not want your proprietary data “leaked” for LLM training or other unauthorized access. Also, an AI agent should have the same permissions to tools as the user on whose behalf it operates. Certain non-public information (NPI) should be masked in responses according to an enterprise’s operating policy. As a result, many low-code or no-code AI platforms offer these things built in, providing these features as configuration rather than something you would have to build yourself when creating it all from scratch; they continue to add and update those features as regulations and needs change over time which can save you time and money while mitigating business risk.
“Enterprises must go beyond the AI hype. Finding value requires the right people, solid governance and tools that evolve with your data, risks and goals. Success isn’t about adopting AI fast but about adopting it smart”
AI agents can function in isolation, or as a cohort of collective agents, each with its specialty. New standards are advancing that allow agents to communicate effectively with an LLM, a collection of LLMs, or a collection of other agents. Such complex automation and behavior can allow for large parts of bespoke software development to be optimized with minimal traditional development. It also increases the need for automated testing in such scenarios.
Assuming you have done the needful according to my earlier guidance, customer-facing AI apps are now easier than ever to develop and deploy; just be aware of the increased risk due to changing regulations, some of it at the international level.
A Way Forward to Harness the Promise of Generative AI
What does it take to employ AI using Rag and agents, all grounded in my company data? The short answer is that it is complicated and the details are constantly changing. One must be ready to change over a short period and have a team who can adapt to the technology. Generative AI inherits characteristics and tools from traditional data science and much of the value in AI requires lots of quality data. At the same time, infrastructure using existing software development techniques is required for any bespoke solution. If one chooses a low-code solution—configuration versus custom development—the implementers still need good UI/UX skills paired with some foundational software development. Do you just want to buy generative AI as a turn-key solution? That’s fine too and while those are often “baked” into existing tools and platforms, what happens to your data and how it is secured and retained is still something to be aware of and governed according to your company standards, not those of the vendor.
The generative AI landscape is more complicated today than it was a year ago and those patterns do not look to belay any time soon. Enterprises need skilled, cross-functional personnel who can do the homework to evaluate the proper tools and platforms to fit the business needs and the business people, be that IT or on the run side of the company. Those people are difficult to find in today’s market and it will take some time for the skills to catch up to the hiring needs. Be it build or buy, low-code or bespoke, you need the right people to not just sell something new and costly but to also ensure it adds the right value. Hire them or upskill them and then trust them with the challenge. You will be glad that you did and your business will thrive accordingly.
