How do I find which workflows to automate with AI?

It depends on what your people spend their time on. If much of the week goes to reading, summarizing, drafting or moving information between systems, those tasks are usually the best starting points. If the work is mostly judgment calls, relationships or physical tasks, AI will help less, and you should look for the supporting admin around that work instead. The reliable method is the same either way: list real tasks, score them on volume, pain and risk, then test the top few on real examples before buying or building anything.

Who this is for

COOs, operations leads, managing partners and team heads who know AI could help but are not sure where to point it first.

The short version

Start from tasks, not tools.
Score each task on how often it happens, how painful it is, how standard it is and how risky a mistake would be.
Test the top three on real work for two weeks before committing.
Pick tasks where a person still reviews the output, at least at first.

Why most firms pick the wrong workflows

Firms often start with whatever a vendor demo showed or whatever a senior person finds annoying that week. That leads to projects that look impressive but touch a task that happens twice a quarter, or projects that aim at high-stakes work where one mistake is costly. The result is a stalled pilot and a team that decides AI is not ready. Our guide to running an AI pilot that does not stall covers what happens next, but the choice of workflow is where most of the outcome is decided.

Step 1: Build a task inventory

Ask each team to list the recurring tasks they do in a typical week or month. Keep it concrete. “Client reporting” is too broad. “Pull last month’s numbers from three systems and write a two-page summary” is useful.

For each task, capture:

Who does it and how often.
Roughly how long it takes each time.
What goes in (documents, emails, spreadsheets, systems) and what comes out.
What a mistake would cost.

A shared spreadsheet is enough. Thirty to sixty tasks is a healthy first list for a small or mid-size firm.

Step 2: Score each task

Rate each task from 1 to 5 on five questions:

Volume. How often does it happen across the firm?
Pain. How much time or frustration does it cause?
Pattern. Does it follow a similar shape each time, or is it different every time?
Text and data heavy. Is most of the work reading, writing, summarizing or moving information?
Low risk if wrong. Can a person catch a mistake before it matters?

Add the scores. The highest totals are your candidates. A task that scores well on volume and pain but low on risk safety is still a candidate, but it needs a careful review step built in.

Step 3: Sort the candidates into three buckets

Use what you have. Tasks that a general AI assistant can handle with good prompts and training. Most firms find many tasks land here.
Connect your tools. Tasks that need information from your CRM, email or file storage. See how AI agents connect to your CRM and email.
Build something. Repeated, multi-step work where a custom assistant or agent would save real time. See build vs buy AI tools before deciding.

Starting in the first bucket is usually cheapest and fastest.

Step 4: Test the top three on real work

Pick the three highest-scoring tasks. For each, ask one or two people who do the task to try AI on it for two weeks, using real examples and following your AI acceptable use policy.

Track simple things:

Time taken with and without AI.
How much editing the output needed.
Any errors that a reviewer caught.
Whether the person wants to keep using it.

Step 5: Decide and roll out

After two weeks, keep what worked, drop what did not and write up the winning approach as a short guide or shared prompt. Then train everyone who does that task. Our playbook on creating prompting standards helps here.

Revisit the inventory every quarter. Tools change, and a task that was a poor fit six months ago may be a good one now.

Signs a workflow is a good fit

People describe it as tedious rather than difficult.
The inputs are already digital.
There is a clear template or standard for the output.
Someone already reviews the result before it goes out.

Signs it is a poor fit, for now

The task depends on information nobody has written down.
Every instance is unique.
A single error would cause serious legal, financial or client harm with no review step.
The task is rare.

Common mistakes

Starting with the hardest problem. Early wins build trust. Save the ambitious projects for later.
Skipping the people who do the work. They know where the time actually goes.
Measuring nothing. Without before-and-after numbers, you cannot tell what worked. See measuring AI training results.

When you do not need us

If you have an operations lead with time to run the inventory and scoring, and a team willing to test, you can follow this playbook yourselves. Outside help is useful when you want a faster, independent view across teams, or when the best candidates need connected or custom tools.

How Ruby helps

AI strategy: running the task inventory and scoring with your teams and turning it into a clear roadmap.
AI training: teaching people to handle the first-bucket tasks well with the tools you already pay for.
AI engineering: building the connected and custom tools for the workflows that need them.

If you want help finding where AI fits in your firm, book a call with Ruby.