The new architecture for ecommerce AI
Customer Service that learns on its own
5 min read
Most conversations about AI in customer experience still come down to a false choice. On one side, a single large model that tries to answer everything and has to be constantly tuned. On the other, rigid rule-based workflows that break the moment a customer says something unexpected. We have never accepted that trade-off.
At DigitalGenius we look at things differently. We take a hybrid approach: AI for judgment, approved workflows for the parts of the job that have to be exact. That decision looks more correct with every month that passes, and it is the foundation for everything we are building next.
I want to walk through where the platform is today and what we are shipping over the next year, which I am really excited about. The two are of course connected. You cannot get to automations that improve themselves without first getting the architecture right.
The architecture: AI for judgment, workflows for certainty
Our platform runs on a multi-agent architecture. An orchestrator sits on top, and beneath it are specialist agents and a library of approved workflows. When a customer gets in touch, the orchestrator reads the request, investigates before it acts, and routes the work to the right specialist agent and the right workflow.
A WISMO question goes one way, a returns question another, an order change request goes yet another way again. If a customer asks about three things in one message, the orchestrator can resolve all three. If the request is ambiguous, it can ask a clarifying question rather than guessing.
This is a step up from what many of you might recognise as intent detection. The orchestrator still routes, but it does far more, and it is fully configurable through a combination of prompting and flows. Based on our own testing we anticipate the orchestrator will lift ticket resolution by roughly 10%, with the largest jump being on email and helpdesk channels, where resolution has historically been the most difficult to achieve.
The reason a system like this works is that each agent has limited responsibility. It’s a team of specialists working hand-in-hand with a project manager (the orchestrator) versus one generalist trying to do everything.
That is also what lets the platform scale in two directions at once. Horizontally, we are adding new use cases and new channels, where there are often quick wins. Vertically, we are able to resolve more scenarios within the use cases you already run, handling the unhappy paths and the edge cases.
The specialists carry out that work through a capability we call AI Activity, which is live in the DigitalGenius platform today.
The important concept here is Flows as Tools. In an agentic tool-calling loop, a specialist agent calls the flows and external tools it needs to work through a task, handling all of the complexity behind the scenes.
Trust is a system, not a feature
A lot of vendors answer the trust question with the words “guardrails” and “safety.” That language is defensive, and it is bolted on after the fact. We think about trust differently, and we build it into the platform as three things that work together.
The first is the quality of the journeys themselves. Our Genius Flow Library is drawn from millions of real ecommerce conversations and a decade of experience. Brands working with DigitalGenius are starting from proven flows rather than a blank canvas, whether you are migrating from another system or introducing a new capability.
The second is evaluation. We now run simulation and QA directly in the platform, so customer interactions are fully tested before they ever reach a live customer, and we monitor and score conversations continuously once they are live.
The third is decision traces. Every automation is fully inspectable. You can see a high-level view of what happened, dig into every decision, every integration call, and every piece of data that produced a given outcome so your AI is fully traceable; it’s easy to see historic actions and the reasoning behind them.
The path to self-learning
Here is the part I am most excited about. Today, improving an automation process is human work. We build a change, test it, evaluate it through an A/B test or controlled rollout, put it live, monitor it, and look for the next opportunity. Then the cycle starts again. It’s a good loop, and it works, but it is entirely dependent on people doing a lot of careful, repetitive analysis.
We are closing that loop so the system can optimize on its own. Several pieces of that puzzle are already live. Ask Genie gives you analytics on demand: it has access to all of your platform data, from tickets and executions to flows, and you can use it inside the dashboard or through your own LLM via MCP. Ask what your customers’ top complaints are, why CSAT dipped yesterday, or where a specific prompt lives, and it answers from your data. Simulation and evaluations, as I described above, cover the test and monitor steps.
The next two pieces are what make the loop self-sustaining. Automation Insights identifies areas that need addressing and then recommends a solution and shows you the evidence. Because we have a decade of ecommerce data across our customer base, it can tell you that brands with a profile like yours did X to achieve Y, and then show you the impact that change would have on your own conversations before you commit to it. Imagine being able to, at the click of a button, run complex analysis and suggest tried and tested recommendations to your leadership team for CX initiatives.
And the Architect is the agent that can actually implement the change: editing flows, adjusting prompts, adding specialist agents, updating the rules for closing tickets, etc..
I want to be clear about the role of people in this. The point is not to remove them. The point is to take the optimisation, maintenance, and bug-fixing work off their plates so that people can spend their time on the more complex problems, the genuinely difficult decisions about how to introduce a new use case or where a business policy should land.
The Architect is powerful and we are working on honing it so it’s as useful as possible to both CX and technical teams. Every step ships one at a time so you can use it and shape it, and our plan is to have self-learning fully live within twelve months. This does not really exist in the world yet. We intend to be the ones who build it.
