## Legacy domains and systems of record A dominant trend that I see in the legacy space from digital service providers, they either launch numerous agents or an agent to position themselves as AI capable. These agents either come tightly coupled with their own product suite or narrowly scoped to limited intelligence. The dynamic nature of agents doesn’t fully materialise if there are artificial constraints. The AI model companies are focusing on building open standards in order to make interaction across agents interoperable. [Agentic AI foundation](https://aaif.io/) maintains open standards across many of the primitives required for agentic AI. Yet, this strategy limits the potential value a customer derives from offered Agentic AI solution. I wanted to propose a different approach that would allow digital service providers to present a complete solution and also address one of the undercurrents that require heavy enablement even for Agentic AI sessions. Before I get into it, I wanted to focus time and effort in explaining why customers in legacy domains require support to adopt new product. Take for example a Transport Management System(TMS), it is the most painful digital software implementation project that any transportation company undergoes. No one enjoys the process because it breaks most of the existing processes in place. Also, people operating these tools, don’t want to undergo the pain for future benefits. TMS is the system of record, i.e work done if not recorded in it will not be considered done. It has millions of integrations to other services that are critical to operations. Moving all of these functions to a new product or different provider is always a painful process. Yet, being a TMS provider for a customer in this space comes with an expectation of AI capabilities. In the current wave of Agentic AI, every customer would like to adopt AI. The demand from the customer in legacy spaces is to use the cheap intelligence to help their team improve outcomes. The other category of demand is workflow automation triggered based on situational awareness of the process. For example, a specific shipment is stuck in transit. What should be the next step of processes that need to be triggered and potential solutions to fix it. Customers would like agentic AI to help them do all such efficiency increments. Digital service providers are faced with a predicament, should they add AI capabilities by changing the fundamental interaction in their software or allow users to interact with their software using 3rd party agents. ## Alternate approach I would like to propose a 3rd option that I think addresses the underlying issue with systems of record. The user is using it for two purpose, co-ordination across upstream and downstream processes and documenting the transaction for record keeping. Multiple users are using it across all available modules but not everyone is accessing each module. Companies find it painful because training users to adopt to a system takes time. In the previous era we spent significant time in making the UX and UI accessible and intelligible for users to intuitively use these digital tools. What if human users no longer use it ? Computer use^[[Open AI](https://developers.openai.com/api/docs/guides/tools-computer-use), [Claude](https://platform.claude.com/docs/en/agents-and-tools/tool-use/computer-use-tool) and [Cursor](https://cursor.com/docs/cloud-agent) show some capabilities of how their models can enable this for their users] is now available in each of the model providers and the models themselves have sufficient primitives to navigate using any digital tool on a computer or more narrowly in a browser. I pick computer^[Excel is the both a boone and bane for success of a product in legacy domains. We punch it to make the product standpoint but we also loose to it when the product doens’t do what the customer wants. Computer use allows the VM to have its own excel and it can use it to bridge gaps] over browser because it then allows usage of tools that are provided by you or 3 rd party on the Virtual Machine that gets spun up. The goal for system of record is to help project manage and maintain transaction records to improve efficiency of their operations. What if the agent does both while the users are just texting with it. And what if that collaboration is similar to how they would with their fellow peer. ![[Agentic AI with Computer Use.png]] ## Computer Use is Agentic At the center of it is a Virtual machine(VM) with Agent SDK that allows the digital service provider to package agentic capabilities beyonds its own tooling. They can attach customer specific skills that document the System operating procedures into repeatable but dynamic execution artefacts which agents could follow. Workflow editor that you define the if else record updates within the system. A customer maps their current processes and then instructs the agent to follow the due course and every new occurrence is mapped iteratively. This also works if the customer is using more than one product or service from the provider. The VM could host both the products and agent is now capable to natively use them. The customer/user could just define the intent of their request and the agent handles the tool use on its own and they can verify its execution in their own leisure. The actual service or product needs no updation or modification to make it agentic for the customer. This setup allows customers to instal 3rd party software in the VM and ask the agent to execute multi-product workflows by building them in the editor and defining individual steps with skills. It could be as basic as opening the mail in the VM and then triggering workflow from a single email. I have said it [[Software was meant to be means to an end|before]] and I will say it again, with agents the ultimate goal is to use less software. ## Why turn the dial backward? Much of this looks like the era before cloud computing took off and Software as service became ubiquitous. Yet, I think this is how digital service providers can multi-hone and retain customers into the agentic era. This approach allows customers to launch in a limited scope without disrupting their current status flow. The VM and agent work on a sub-section of records and they are doing it in a remote environment that doesn’t come in the way of the customer. Both are connected to the same database but the agent is not part of the main product that humans are using. For example to check for its viability, run a test instance shadowing your production and see if the agent executes all the steps to complete a transaction. If it breaks, we can debug and release a patch. Even better, run it through an agentic loop to make it self sustaining the self correcting. I mentioned about the painful process of [[Understandings Systems at work|adopting]] such software in legacy domains. Most of it comes down to customers not confident to find time to train new users about the product and the process they follow with the product. This deters them from adopting the new and novel innovations that were released. Next, is the budget. Navigating both the constrains results in a big variation in support that needs to be provided for legacy products. This approach of using VM and AI models with computer use ensures that irrespective of gap from the most recent release, a customer is able to adopt Agentic AI for the same outcome as the most updated product customer. I am not getting into the mechanics and costs associated in making this happen because those come secondary to the approach. Once the configurability in this nature is evident, then investing in the infrastructure, model and tooling becomes significantly easy to unpack. And this is an emerging pattern the users in the consumer space are adopting with their personal agents like [Muse](https://ai.meta.com/muse/)^[This similar to how mobile as a platform was adopted in work space. Enterprises will have employees getting VMs which they access with their personal agents. In the VM, enterprise controls the software deployed and what an agent can do just as they prescribe for employees] *** `Tags:` #ai #agent #TMS #logistics #legacy-domains `Original Pulished date: ` 13-11-26 `Source:` Seed