The short answer.

AI workflows connect to your accounting system, CRM or help desk in three levels, and you start at the first. At level one a person exports a file from the system and gives it to the workflow, so no connection is needed, and most first workflows work this way. At level two the workflow reads from the system through a connection your company has approved, and at level three it writes back, drafting a change that a person approves before anything is saved.

So you do not need a connection to start. Letting AI near your systems is safe when you decide reading and changing separately, give a workflow only what its work needs, and keep a named person approving every change. Each level is set out below, with a sheet to copy for any workflow and any system.

Level one: a person exports a file.

At level one, a person exports a report or a list from the business system and gives the file to the workflow in your company’s own ChatGPT, Claude, Copilot or Gemini. The workflow prepares the work from that file: a follow-up, a summary, a checked list, a draft reply. Nothing connects to the system, so the workflow cannot see or change anything in it.

Level one is where most first workflows start, for three reasons.

  • It needs no approval from anyone who runs the system.The person already allowed to export the report is the one who exports it.
  • It proves the workflow first.Your team learns whether the workflow prepares the work well before anyone decides on access.
  • It shows exactly what the workflow needs.The columns the workflow uses in the export become the list of what a connection would have to read later.

We have a plan. How do we build the first AI workflow? walks through building one from an export.

Level two: the workflow reads through an approved connection.

At level two, the workflow reads from the business system directly, through a connection your company has approved, so nobody exports a file. The workflow pulls the records it needs when a person runs it, such as open invoices from the accounting system or this week’s tickets from the help desk.

A connection at level two is read-only. The workflow sees what it was given access to and changes nothing. The person who runs the system approves the connection, and the approval names which records the workflow may read. Move a workflow to level two once it is in weekly use at level one and the export has become the slow part.

The work that gains most from level two crosses more than one system. The work that matters crosses systems shows how to map those handoffs.

Level three: the workflow writes back, and a person approves.

At level three, the workflow writes back to the business system: it drafts a record, a note or an update, and a person approves the draft before it is saved. A workflow at level three can draft a follow-up note on a customer record in the CRM, or a reply on a ticket in the help desk, and leave it for the person who owns that work.

A workflow can be designed to pull from the systems your company already runs and push data back, with guardrails and a person approving. The guardrails are written into the workflow and into the connection: which kinds of record the workflow may draft, which it may never touch, and who approves each change. Anything that sends, pays or changes a record waits for that person.

Reading and changing are separate permissions.

Reading a system and changing a system are two different permissions, and your company decides each one on its own. A workflow that reads open invoices has no right to edit them. Changing anything is a second decision, made later, by the person who runs the system.

  • Read only what the work needs.If the workflow prepares overdue-invoice follow-ups, it reads open invoices and customer contacts, and nothing from payroll.
  • Change only through drafts.At level three, the workflow drafts and a named person approves. The workflow never saves on its own.
  • Know how to switch it off.Every connection has one person who can remove it, and everyone who uses the workflow knows who that is.

Your rules about which assistant is approved and what may go into it apply to connections too. Does a small company need an AI policy? covers those rules.

A one-page connection sheet to copy.

Fill in one sheet for each pair of one workflow and one system. Keep the sheet with the workflow’s instructions.

  1. The workflow.Its name, and the person who owns it.
  2. The system, by kind.Such as the accounting system, the CRM or the help desk.
  3. The level.One (a person exports a file), two (the workflow reads) or three (the workflow drafts changes).
  4. What the workflow reads.The records and fields, and nothing else.
  5. What, if anything, the workflow may change.The kinds of record it may draft, and the kinds it may never touch. Write “nothing” at levels one and two.
  6. Who approves access.The person who runs the system, by name.
  7. Who approves each change.The person who checks every draft before it is saved, by name.
  8. How it is switched off.Who removes the connection, and how.

One workflow, three levels.

Illustrative example.

Take a 70-person logistics company whose customer service team answers delivery questions every day. The first workflow on its plan prepares a reply to each open ticket, and the customer service lead owns the workflow.

At level one, the customer service lead exports the open tickets from the help desk and the week’s delivery list from the dispatch system, and gives both files to the workflow in the company’s own assistant. The workflow drafts a reply for each ticket. A colleague on the team reads each draft, edits it and sends it from the help desk as usual.

Once the team runs the workflow every day, the operations manager, who runs both systems, approves read-only connections for open tickets and delivery status. Nobody exports any more. That is level two.

At level three, the operations manager approves one change: the workflow may save its draft reply on a ticket as an unsent draft. It may not send a reply, close a ticket or edit a delivery. The person who works the ticket approves and sends every reply. The connection sheet names the operations manager as the person who can switch both connections off.

What is different in a company of under 200 people.

Most published advice on connecting AI to business systems assumes a large company with an integration team and a platform to buy first. A company of under 200 people works differently, and level one is the reason.

  • You start without an integration project.A first workflow runs on exports from the systems you already run, in the assistant you already pay for.
  • The person who approves access is down the hall.The operations lead or the owner who runs each system makes the decision about a connection.
  • You connect only what has earned it.A connection is approved for a workflow your team already uses every week, so every connection has a reason on paper.

Your systems and seats, pointed at named work one level at a time: that is 10X the ROI of your existing AI.

Where to start.

Start every workflow at level one, and move up when the work asks for it. Connectors: how a workflow works across your business systems follows one routine across three systems. Workflows that connect your business systems explains how FullyPowered AI chooses a connection when a workflow needs one. The Connectors Directory lists business systems by kind and the public connectors known for each. Once a workflow is in use, How to deploy AI workflows to a whole team, and keep them working covers the rest.

Which workflows come first, and which systems each one draws on, is part of the plan. How organizations plan their use of AI sets out the three steps. FullyPowered Planner is where the plan lives: you set out your teams, their roles and the work between them, and choose what to build first.