Where it runs, and what leaves.

Operative is a Windows desktop agent. It runs on the machine already in use, signed in as the person who already does the job, against the systems that person is already entitled to open. It opens no inbound network path and needs no new server.

Workflows live under the user profile, in the application data folder, and every save writes a timestamped snapshot so an earlier version of a process can be recovered without a source control system.

A workflow file never carries a password or login. When a bot reaches a login screen, the person running it logs in, and the bot carries on.

Guerrilla Bots uses OpenAI models served through Microsoft Azure, in Azure's South India region. The model reads what a recording captured and puts it into plain language, and answers questions about the workbook logic a person narrates while recording. It doesn't hold the automation. The steps themselves are captured by the application.

For an enterprise pilot, the AI calls can be pointed at your own Azure OpenAI or Amazon Bedrock account, so they stay within your cloud boundary and show on your own bill and logs. This is designed and set up as part of pilot onboarding. It isn't switched on by default.

Under Microsoft's terms for Azure OpenAI, inputs are not used to train models. Microsoft may keep them for up to 30 days to check for abuse. We don't use your process content to train anything.

A recorded process should not go from one person's desk into a month end close in a single step. The model we set up with your IT team separates a workflow into three environments, held as three separate folders with their own access:

• DEVWhere a process is recorded and corrected. Sample or masked data only.
• UATWhere the person who owns the process checks the output against a known good result before anybody depends on it.
• PRODWhere it runs for real, on the schedule agreed, against the live systems.

Promotion between them is a deliberate act with a record of who did it, rather than a file being copied quietly. This is the model, and it is not in the product yet. Guerrilla Bots is pre-launch and this page will say so plainly until it ships.

An automation should be able to reach exactly what its process needs and nothing else. What we agree with IT before a pilot is a named list, not a general permission:

• •The specific applications the process is allowed to drive, named individually.
• •The specific folders it may read from, named individually.
• •The specific folders it may write to, which is a shorter list than the read list in almost every process.
• •Nothing outside that list, including anywhere else on the same machine.

Read and write are granted separately, per folder. A process that produces a report reads from four places and writes to one, so it gets four read grants and one write grant, and that is the whole of its reach.

Commercial and confidentiality terms, including a mutual non disclosure agreement, are agreed before a pilot begins. If your procurement or legal team needs to see paper before a technical conversation, that is a normal order to do things in and we are happy to start there.

Guerrilla Bots is pre-launch. There is no certification to show you, no audit report, and no customer reference, because there are no customers yet. Any vendor page that implies otherwise at this stage is worth reading twice.

What is on this page is either how the software works today, which you are welcome to verify during a pilot, or what we will put in writing before one starts. Nothing here is a measured result.