Accounting bots can own whole processes, from billing and billing support to cash applications and collections. They work when the setup is right: one process per bot, a process that already runs reliably, a manual the bot can't edit, permissions set by risk, and a staged ramp from watching to proposing to running.

What can accounting bots do?

Quite a lot, when each one owns a real job. xAI recently published how five Grok Bots run accounting work at SpaceXSI (https://x.ai/news/grok-bot-accounting): provisioning customer accounts, billing, billing support, cash applications, and collections. Each bot handles its process from start to finish, the way a person on the team would.

The figures xAI reports give a sense of scale. Billing Bot reduced the time required to issue an invoice from roughly 24 hours to minutes. By xAI's estimate, Billing Support Bot has freed up approximately eight hours a day across the accounting team. Cash applications that ordinarily take an accountant three to four hours finish in roughly 30 minutes. xAI estimates the same work would otherwise require more than doubling the team's headcount.

Those results come from a specific way of setting the work up. Here's how I'd implement it at a business.

Give each bot one process, end to end

Start with a job a person on your team owns today: invoicing, cash application, collections. Give the bot that whole process, not a pile of tasks to help with. If a bot doesn't have a clear job, you don't have an owner. You have a helper, and the process is still sitting on someone's desk.

Get the process ready before the bot

A bot won't fix a process people can't run reliably. If the workflow is poorly defined or lives in someone's head, that's the first project. Write down the steps, the inputs, and the exceptions. Put controls on which actions are allowed. Then hand it over. Define the process, then assign it.

Write a manual the bot can't edit

Each bot works from an operating manual: how to verify a contract, which adjustments need a person, how the job gets done. In xAI's setup, the bots consult their manuals each night and can't change them. Only the people responsible for the process can. That keeps the people who know the job in charge of how it gets done.

Set permissions by how hard a mistake is to undo

Routine, low-risk work can run on its own. Sensitive actions, like credits or refunds, wait for a person to approve them. Decide who can authorize those actions, too. Anyone with the right access can ask a question about an invoice. Only designated people should be able to change billing records. That's a control environment, not a chatbot with system access.

Roll out in stages

Ramp a new bot the way you'd ramp a new hire. xAI's typical onboarding now takes about three weeks. In week one, the bot observes. In week two, it proposes actions and a person approves them. In week three, it runs routine work and escalates exceptions. Customer-facing work takes longer: xAI spent approximately three months on its Collections Bot. The sequence matters more than the calendar: watch, then propose, then run.

How 20/20 Solutions sets this up

I run my own work this way: about ten specialist bots, each owning one job, working from a shared written manual, with approval set by how hard a mistake is to undo and a staged ramp-up for new bots. At 20/20 Solutions, the work is finance operations, accounting modernization, and AI agent deployments. It starts with a real conversation, then a working prototype instead of a proposal, built in production shape, deployed, then extended.

If you have an accounting process that eats hours every week, that's where I'd start. Get in touch at hello@2020solutions.ai.

Book a conversation →