AI agents can reason across documents, write software, and coordinate increasingly long sequences of work. Inside a company, however, the final step still tends to look much less advanced: a person copies a value into an ERP, opens a desktop application, works through an internal portal, or checks that a record changed in the system that actually owns it.
The constraint is no longer only intelligence. It is access to the operational reality of the company. Most of that reality lives in software built for people to operate one screen at a time, not for agents to call as dependable tools.
Today, we are introducing Graft. Graft is the interface layer between AI agents and the software a business already depends on. It observes real workflows, models how work moves through them, and turns those workflows into stable, permissioned, auditable tools. The result is a living map of company operations that agents can act through without a replacement project underneath it.
For clarity, Graft AI is a standalone Axcelner product launched in 2026. It is not affiliated with the earlier company named Graft whose team and technology were acquired by You.com in 2025.
The missing layer is operational
A company does not run on a clean diagram of its systems. It runs on the accumulated detail inside ERP installations, mainframes, web portals, virtual desktops, internal applications, files, exports, approvals, and exceptions. Some of these systems have APIs. Many have incomplete APIs, inaccessible APIs, or no useful machine interface at all.
An agent can be shown how to navigate those interfaces from scratch. That can be useful for exploration, but production work asks for something stricter. The same action should accept the same inputs, respect the same permissions, handle the same failure states, and produce evidence that the intended business effect occurred. The agent should not have to rediscover the workflow every time it runs.
Traditional integration work approaches this by replacing the application, waiting for an API, or writing a brittle path around the interface. Graft starts with the workflow as it exists.
From interface to living map
Graft watches a real workflow and maps its screens, inputs, transitions, rules, and side effects. It uses those observations to build a capability graph: not a recording of where someone clicked, but a model of what the system can do and what each action means.
From that graph, Graft compiles a typed tool contract with bounded inputs, explicit policy, recovery behavior, and conformance tests. Agents call the contract—including through MCP where appropriate— instead of operating the underlying interface directly. If the application changes, the adapter can be isolated and recertified while the tool presented to the agent remains stable.
This changes the unit of integration. Instead of giving an agent broad access to an application, a company can expose a precise operational capability: create this kind of invoice, update this field, reconcile this report, resolve this class of request. Each capability can carry its own identity, role, approval, and audit boundaries.
Action requires proof
A completed click path is not the same as a completed operation. A confirmation screen may be optimistic. A timeout may happen after a write succeeds. A retry may create the same order twice. For agents acting in systems of record, a plausible success message is not a sufficient definition of done.
Graft separates execution from verification. The adapter performs the bounded action; an independent witness checks the resulting state in the source system. The output is not simply that an agent reached the end of a workflow. It is evidence of the effect, tied to the request and the policy under which it ran.
Graft is designed to run in the customer’s environment, with least-privilege access and an audit trail for every action. Existing systems remain systems of record. Graft gives agents a controlled way to use them.
Start with one workflow
Graft is now in private beta. We are starting with a deliberately concrete question: what is one valuable workflow your agents cannot reliably reach today?
Bring us the software you cannot replace and the operation that still ends in a human handoff. We will begin there. Learn more and join the private beta at graft.axcelner.com.
— Axcelner