runnrr
A fixed automation follows steps you set in advance: when this happens, do that. An AI agent gets a goal and decides its own next steps, picking which tools to use as it goes. For most client admin in a small firm, start with a fixed automation. Add a single AI step where the input is messy, like reading a free-text email. Keep a person approving anything that goes to a client or touches money or deadlines.
An agent makes sense when the steps can't be set in advance and you can still check what it did.
The three options
Fixed automation | Fixed automation with one AI step | AI agent | |
|---|---|---|---|
How it works | Same steps every time, triggered by an event | Same steps, but one step reads, sorts or drafts something | Plans its own steps toward a goal and uses tools to get there |
Good at | Reminders, logging, moving data between tools | Reading emails and attachments, sorting replies, drafting messages | Open-ended jobs where the number of steps isn't known up front |
Weak at | Anything it wasn't set up for | Being right every time. It needs a review step | Predictability, cost per run, explaining what it did |
Checking it | Easy: you can see every rule | Check the AI step's output, the rest is fixed | Harder: you need a log of what it did and why |
The middle column is what OpenAI's business guide to agents calls LLM-powered steps: the workflow stays rule-based and the AI handles one step that needs interpretation, then hands control back. Anthropic's guidance on building agents says to find the simplest solution that works and notes that agents bring higher costs and the risk of errors that build on each other. Both are worth reading if you want the longer version.
Decision table
Answer each question for the task you have in mind.
Question | If yes | If no |
|---|---|---|
Does it start from something structured, like a form, a status change or a date? | Fixed automation | Consider an AI step to read the input |
Are the steps the same every time? | Fixed automation | Consider an agent, with a person checking |
Does it need to read free text, a PDF or a photo? | Add one AI step | Fixed automation is enough |
Would a wrong action be costly (wrong client, wrong amount, missed deadline)? | Keep a person approving that step | Can run without approval once tested |
Do you need to explain later exactly why it did something? | Fixed automation or an AI step with its output logged | Either |
Does it run often enough to be worth setting up? | Go ahead | Leave it manual for now |
Use the table one row at a time. Fixed steps can handle a date, status change or form submission. Unstructured input may need an AI step, while costly or unclear actions need a person’s approval.
Examples from client work
Task | Best fit | Where a person approves |
|---|---|---|
Reminding a client about a document that's overdue | Fixed automation | Not needed once the template is approved |
Logging a website enquiry into your client list | Fixed automation | Not needed |
Sorting client replies into "sent the document", "has a question" and "out of office" | Fixed automation with an AI step | A person checks anything the AI isn't sure about |
Drafting a reply to a client question | AI step drafts it | A person reads and sends it, every time |
Matching an emailed receipt to the right client and month | AI step suggests a match | A person confirms unclear ones |
Pulling a summary of where a matter stands from email, calendar and the case system | Agent or several AI steps | A person reads the summary before anything is sent or changed |
Anything that gives legal, tax or financial advice | Not automated | The professional decides |
Slack shows the same split. A Workflow Builder form with fixed fields (client, item, due date) is a fixed automation. It's easy to check. A message like "can someone sort out the Smith file before Friday" has to be read first. An AI step can turn it into a draft task for someone to confirm. It shouldn't assign or close anything on its own.
Where human approval belongs
Put a person between the automation and the outside world at these points:
- Before any message goes to a client, unless it's a pre-approved template with only names, dates and item names filled in.
- Before anything involving money, such as an invoice, a refund or a payment link.
- Before anything is filed, submitted or sent to a court, a tax office or a regulator.
- When the AI step isn't confident or the input doesn't match any rule.
- When a client replies about the outstanding request, pause that request’s sequence until the owner reads the reply. An unrelated reply must not stop a different request.
Approvals slow things down a little. In exchange you get a clear record and nothing goes out that nobody saw. You can loosen them later for steps that have run cleanly for a while.
Working around the software you already have
Neither option means replacing your practice management system, accounting software or inbox. There are four common ways to connect to what's already there:
Method | How it works | When it fits | Trade-off |
|---|---|---|---|
Official connection (API or built-in integration) | The automation reads and writes through the tool's own connection | The tool has one and your plan includes it | Best option when available. Check plan limits and permissions. |
Scheduled export | The tool emails or saves a report (often a CSV) on a schedule. The automation works from that. | No connection, but the tool can export | Data is only as fresh as the last export |
Connected inbox | A bot handles the conversation and starts the agreed work in connected tools | Requests arrive by email, WhatsApp or Slack | Requires the right channel access, permissions and tool connections |
Manual assist | A person does the one step the tool can't support, like pasting a reference number. The automation does the rest | Nothing else works | Someone has to remember the step |
Use the official connection where you can. Only use accounts and access the firm has authorised. Never share a staff member's password with an automation tool.
How to decide for one task
- Write down the task in one line: what starts it and what the result should be.
- Run it through the decision table.
- Pick the simplest option that works. If a fixed automation covers most cases, send the odd ones to a person instead of building an agent for them.
- Mark where a person approves.
- Choose the connection method for each tool involved.
- Test it on real past cases before it touches a live client.
Runnrr scopes automations like these around the software a firm already uses, with the approval points agreed before anything goes live. If you'd like to talk through one workflow, see how Runnrr sets up a workflow.
Frequently asked questions
What is the difference between an AI agent and automation?
A fixed automation runs the same steps every time something happens, such as sending a reminder when a document is overdue. An AI agent gets a goal and chooses its own steps and tools to reach it. Automation is predictable and easy to check. An agent handles more variety but costs more per run and is harder to check.
Should a small firm use an AI agent or a fixed automation?
Start with a fixed automation for most client admin. Add one AI step where you need to read free text or attachments. Use an agent only when the steps can't be set in advance and a person can review what it did.
Where should a human approve in an AI workflow?
Before any non-template message goes to a client, before anything involving money, before anything is filed or submitted, whenever the AI isn't sure and when a client replies about the request being followed up.
Can I automate client work without replacing my existing software?
Yes. Use the tool's official connection if it has one. If not, a scheduled export or an agreed manual step may cover the missing connection. Confirm what can be read and updated before relying on it.
Is an AI step in a workflow the same as an AI agent?
No. An AI step does one job inside fixed steps, like sorting an email or drafting a reply, then hands back to the workflow. An agent decides the sequence of steps itself.
