What does an AI agent integrated into a SaaS product look like?
Your users see it in your screen, under your name and colours. They ask it to do a piece of work: prepare a quote, find a record, answer a question about figures. The agent calls your software’s functions to do it, under the user’s rights, and hands over what it does not know.
Behind the scenes, nothing in your product has been rewritten. The agent goes through your existing APIs. That is what a headless platform allows: it provides the engine, you keep the head, meaning the screen, the brand and the relationship with your users.
What are the steps?
Six steps, in this order. On a known activity with API access, allow six to eight weeks from the first to the fifth.
1. The workshop: choose the activity and write the criteria
We start from an activity your users perform often in the product, whose result can be checked. With your business experts, we write down what a good answer is, and the threshold to reach before going live. The evaluation set starts here, before any code.
In parallel, we read your software screen by screen to derive its navigation graph. It is redone with every release.
2. The APIs and the MCP gateway
The agent does not access your database, it goes through your API. We set up a single entry point, an MCP gateway, which exposes the useful functions as typed tools, under your roles and rights exactly as they are.
An API with several hundred endpoints usually yields a handful of tools that are genuinely useful for one activity. A narrow, well-described scope beats a broad one.
3. The workflow graph
The steps of the work, in order, with what the agent checks at each one, the tools it is allowed to call and the point where a human takes over. The agent follows that path instead of improvising it. An agent then comes down to four objects: a system prompt, a set of APIs, a workflow, a test set.
4. Evaluation
The agent is measured on a separate test set, before production and after. Tool-call accuracy, faithfulness to sources, escalation rate, cost per action: the page on measuring an AI agent sets out each indicator. It only goes live at the threshold agreed during the workshop.
5. Go-live
The agent ships as Docker images, on your servers or your cloud, with the model of your choice. It starts in proposal mode on sensitive actions, and gains autonomy when the measurements allow. The protections are described in securing an AI agent.
6. Billing and handover
Every successful action is counted. You set the credit price and bill it to your customers under your brand, with the per-customer counter and prepaid packs provided by the platform. The credit model explains the two ways to set the price.
Your team then becomes autonomous on configuring its agents. The second agent costs a fraction of the first: the gateway, evaluation and console are already in place.
Can one agent serve several brands?
Yes. Comet Software Group runs three ERP products, each with its own documentation and users. The agent answers the users of each ERP from its brand’s documentation, and collects their feedback to improve its answers. The need was intelligent support per brand, without tripling costs. The details are in the case study.
Build in-house or start from a platform?
Your team can build an agent. What takes time is the building blocks nobody sees: the graph, the map of the software, per-role rights, evaluation, the credit counter. Building them took us twelve months of R&D. The platform provides them, and your team focuses on what is specific to you: your users’ business.
More than 30 software companies already have agents in production with us. How we run these six steps with you is described in how we work, and the pages by type of software show what it looks like on an ERP, an ATS or a CRM.
