Back to all posts
Insights on Software, AI & Operations
One AI gateway for three products, one agentic workflow for every support ticket
Two AI tools built inside one Circeus portfolio company. A gateway that lets any of its products add an AI feature without rebuilding the foundations, and an agentic workflow that reads the live product and drafts the support reply before a person opens the ticket.
506 is a Circeus portfolio company, a small team building three commerce products, GetSale, EasyScan and EasyGift, for online stores. Two problems kept coming back, and the team ended up solving both with AI, in two very different ways.
The first sat inside the products. The businesses using the products did not need a grand AI feature, but small ones kept making sense: a countdown timer written for a specific sale, a plain-language summary of a store's analytics, a fix to the CSS of an app embed without touching code. Each one, built properly, meant integrating a model provider, managing the context the model sees, adding guardrails and logging what came back, then doing all of it again for the next product. Three products, three copies of the same plumbing.
The second sat inside the company. 506's own support team answers for all three products, and in a busy week those products ship over 100 updates. Nobody can hold that much in their head, so the quality of a first reply often depended on how recently the person answering had seen that part of the product. The knowledge existed in the codebase and the documentation. It just was not where the support team needed it, at the moment they needed it.
These are different problems, and the team built two separate AI tools to solve them. SCOUT is an AI gateway: it does the model integration once, so any of the three products can add a small AI feature without rebuilding the foundations. IRIS is an agentic workflow for the support team: it reads the live state of the products and drafts a reply before a person has opened the ticket. One puts AI in front of the customer. The other puts it behind the people who serve them.
SCOUT puts a model behind any feature in one call
The clearest example is a sale countdown timer. A customer using GetSale, one of 506’s products, picks a tone and an occasion: luxury, a product launch, Black Friday, whatever fits the campaign. The product sends that request through SCOUT, which assembles the context, calls the model and returns the result. The person running the store sees the generated copy and styling and accepts, regenerates or adjusts it.
The same pattern runs through the rest of the products: AI-written analytics reviews in all three, inventory log review in EasyScan, and an admin tool in EasyGift that turns a plain-English request into a minor CSS change to an embed. Each is a different feature. Each is the same call through the same gateway.

Customers are not asking for these features in volume, and the team is comfortable with that. The value is on the engineering side. One gateway means guardrails are enforced once rather than per product, and prompt injection and misuse are handled at a single point. Caching removes the obvious waste, so a repeat query from the same store does not trigger a repeat model call. Per-user limits cap what any single user can push through the gateway and block attempts at malicious use.
The gateway also decides what runs where. Because every feature calls the same place, the team chooses which model handles which feature, and can switch models or LLM providers without touching the products. That matters as much to the cost as the per-call price does, and the per-call price is small: a generated timer costs a fraction of a cent.
Every prompt that goes through the gateway is logged in a shared activity view, with what it cost and what came back. Usage and cost per feature are visible across all three products, so when a feature would run fine on a cheaper model, or better on a stronger one, the team can see it and switch. The model behind a feature becomes a choice the team revisits, not a decision baked into the code. The gateway currently lives inside 506’s products.
IRIS drafts the reply before anyone opens the ticket
SCOUT is about customer-facing features. IRIS is a separate tool with a different job: an agentic workflow that helps 506’s own support team keep up with three products that change constantly. The name comes from Greek mythology, where Iris carried messages between the gods and mortals. IRIS does the same between the codebase and the support team.
IRIS works alongside the team’s support platform. When a ticket comes in, the agent reads the live state of the relevant product, its current behaviour and its help documentation, works out what the customer is actually asking, then drafts a reply. IRIS does not send anything. A person on the support team reads the draft, and either sends it as written, edits it, or ignores it. In most runs, they agree with it and send it almost exactly as written. The AI has done the research; the person makes the decision.
Draft-first design matters because of scale. With updates, improvements and new features shipping across three products, sometimes more than 100 in a single week, no one on the support team, however good, can hold all of it in their head at once. IRIS does not have that problem: it reads the current version of things every time rather than memorising a snapshot that goes stale. Drafting rather than sending also made it quick to get into production, three days from the idea to draft replies landing in the support team’s hands.

The agent also remembers. Short-, medium-, and long-term memory means a question it has already answered well does not need the same research done again, so repeat questions are faster and cheaper to handle, and the answers stay consistent. The team can set up playbooks too: a defined way of handling a particular kind of ticket, so the model follows the approach the team actually wants rather than whatever seems plausible at the time.
One more thing it can do. If a pattern in incoming tickets looks like a product bug rather than a one-off question, IRIS checks the live codebase itself and raises both the issue and a proposed fix for an engineer to review. That pushes a real problem up the priority list instead of leaving it invisible until enough customers complain loudly enough. An agent built to draft replies ends up as an early warning system for the product.