How to Find and Automate Repetitive Business Work With AI

Most businesses do not need to start by asking, “Where can we use AI?”

Start with a much simpler question for employees:

List the most tedious computer-based work in the job that consumes a meaningful amount of time.

That produces a real list of problems instead of a list of AI ideas looking for a problem.

From there, focus first on the task with the highest chance of actually being automated: something AI is likely to handle successfully and where the required systems, data, or tools are relatively easy to give it access to.

That does not always mean choosing the task that costs the most money or consumes the most hours. A huge problem that requires access to ten disconnected systems, sensitive data, and destructive permissions may be a terrible first AI project. A smaller task with clean inputs and easy read access can be a much better place to start.

OpenAI recommends a similar workflow-first approach: look for repetitive low-value work, skill bottlenecks, and other recurring workplace problems, then prioritize the use cases that can create meaningful impact without unnecessary implementation effort. See OpenAI's guide to identifying and scaling AI use cases.

Start With Work People Already Hate Doing

Do not brainstorm AI features in a conference room first. Ask the people actually doing the work.

Look for computer-based tasks that are tedious, repetitive, time-consuming, easy to forget, difficult to keep up with at scale, or dependent on gathering information from many places.

Good candidates include:

1. Summarizing meetings and turning them into action items.
2. Reviewing customer feedback, support tickets, reviews, or account activity to find recurring problems.
3. Generating recurring reports from analytics, sales, financial, or operational data.
4. Updating CRM records from emails, meetings, forms, and other activity.
5. Researching and qualifying inbound leads or prospects.
6. Handling common customer-support requests and escalating unusual cases.
7. Analyzing spreadsheets or datasets to reconcile information, detect patterns, and summarize findings.
8. Monitoring projects for missing deliverables, overdue work, risks, or things that need attention.
9. Handling repetitive IT support, troubleshooting, access requests, and ticket routing.
10. Performing recurring research such as competitor monitoring, market research, vendor research, or weekly intelligence reports.

The important part is that these candidates should come from actual work inside the company. The objective is not to automate something because AI can do it. The objective is to remove work that is already consuming people's time.

Choose the Most Automatable Task First

Once employees provide their lists, do not automatically choose the biggest or most expensive problem.

Choose the one with the highest chance of automatability.

Ask:

• Can AI reliably understand the task?
• Can the necessary information be made available?
• Is that information already available digitally?
• Does the software involved have an API, connector, export, database, or other usable interface?
• Can the workflow start with read-only access?
• Can a human clearly determine whether the result is correct?

A workflow becomes much easier to automate when its inputs and success criteria are observable.

The Subject Matter Expert Should Direct the AI

One of the biggest mistakes in business automation is treating AI like a substitute for knowing how the job should be done.

The person directing the AI should ideally be a subject matter expert who could perform the task without AI.

That person knows what information matters, what a correct result looks like, what can go wrong, what edge cases exist, and which actions are dangerous.

That knowledge needs to make its way into the instructions.

For any serious AI workflow, the initial prompt or instructions should define:

• The goal.
• The completion criteria.
• What a good job looks like.
• What a bad job looks like.
• What the AI should do.
• What the AI should not do.

Account for the ways the instructions could be misinterpreted before the AI encounters them. The goal is to triangulate a successful task: constrain the workflow until the model has a very clear path to doing exactly what was intended and very little room to do something that was not intended.

OpenAI makes the same general point in its practical guide to building agents: reliable agentic systems depend on clear instructions, appropriately scoped tools, guardrails, and testing.

Give the AI Context Before Giving It Control

Start with read-only access whenever possible.

The primary reason is context.

AI should not make changes to something it does not fully understand. Let it gather the context first. Give it a clear picture of what exists, what happened, what the objective is, and why something would need to change.

Only give the permissions absolutely required for the task.

If the job is analysis, read-only access may be all it ever needs.

If the workflow later needs to create records, modify data, assign work, send messages, or perform other write operations, add narrowly scoped write access only when there is a real reason for it.

For important write operations, Astra's default rule is stricter: no write operation unless the specific action has been explicitly approved twice. The AI should state the write action it wants to perform and receive approval before performing it, then obtain final confirmation before execution.

Even with those controls, if an AI system has permission to write or delete data, something can go wrong. The operator is responsible for the permissions granted to the system and for having appropriate backups, versioning, review, and recovery procedures in place.

Protect Private and Proprietary Data

Before giving an AI workflow more data, check what is actually in that data.

Do not blindly expose proprietary information, private customer information, credentials, or other data that should be protected.

The fact that a system technically can access something does not mean it should.

Only expose the information required for the task, use approved systems, and keep permissions as narrow as possible.

Test With a Small Sample First

Do not point a new workflow at the entire company on day one.

Give it a small, representative sample and inspect what happens.

Check whether it follows the instructions, reaches the right conclusion, ignores irrelevant information, avoids inventing facts, respects its permissions, and does not modify anything outside the requested scope.

Once the sample performs correctly, expand gradually.

That is especially important when the AI will eventually be able to take actions instead of simply analyzing information.

A Real Example: Finding Unhappy Clients From Project Data

Astra has used AI to analyze project-management activity in bulk to identify clients who may be unhappy and work the team may have failed to deliver.

Before AI, this was extremely difficult to do across an entire company because each employee naturally has visibility into different clients, projects, conversations, and responsibilities.

AI changes that because it can review the same accumulated context across many projects at once.

In one implementation, the project-management system was Basecamp. The workflow could consider the service agreement, meeting transcripts with the client, the client's posts and activity, and the work the team had actually completed.

Those pieces could then be cross-referenced against one another.

For example:

• What did the client buy?
• What did the team say it would deliver?
• What has actually been delivered?
• What has the client complained about?
• What expectations were created during meetings?
• What is still missing?

That matters when a client buys several complex services. The work cannot always be reduced to a copy-and-paste checklist. It often requires a prescription on a per-client basis. Without AI, keeping track of every small commitment becomes difficult as activity grows.

The output can be very specific: identify where the client is currently unhappy and why, identify where delivery is falling short and why, and explain exactly what needs to be done to make it right, with specific examples.

Ideally, turn those findings into a proposed to-do list for human review.

The responsible team then reviews the findings, discusses them, and agrees on what should actually happen before any changes are made.

Basecamp documents AI-agent options through its official API project, including Basecamp Skills and its CLI. See the official Basecamp API documentation.

What Scale Changes

The value becomes much more obvious when this analysis is applied across a large amount of activity.

For example, AI can review roughly three months of team and client activity across dozens of clients: comments, posts, messages, promises, completed work, and other project context. It can identify clients showing signs of dissatisfaction, commitments that may not have been fulfilled, and work that needs attention, then sort the findings by priority.

In practical use, that type of AI review can take minutes rather than the amount of distributed human effort required to inspect every client project individually. Exact timing varies by data volume and tooling, but the difference in scale is the point.

The AI can also prepare proposed to-dos based on surrounding context: who normally performs the work, what the deliverable should contain, who it should be assigned to, and when it should be completed.

That is where analysis starts turning into automation.

Keep those stages separate at first. Let the AI analyze, review what it found, then decide whether it should be allowed to create anything.

Another Example: Recurring SEO and Analytics Analysis

The same pattern applies outside project management.

SEO and analytics work routinely involve gathering information from several systems, reviewing large amounts of search and website data, identifying what changed, finding opportunities, and turning the analysis into priorities.

AI can dramatically reduce the repetitive part of that process when it has access to the correct source data and clear instructions about what constitutes a useful finding.

For example, a workflow can combine search-performance data with SEO-platform data, summarize important changes, surface pages or queries that need attention, and prepare a recurring report for human review.

The important part is not “use AI for SEO.” The important part is that a previously repetitive workflow has identifiable inputs, a repeatable analysis process, and an output whose quality can be checked by somebody who understands SEO.

For a detailed analytics workflow, see How to Use AI to Analyze Website Traffic Trends and Find What Changed.

Be Careful With Transformative Data Work

Some workflows look easy to automate but deserve extra caution.

CSV transformations are a good example.

AI can sometimes change things that were not part of the instruction. If the file matters, avoid a loose instruction such as “clean this CSV.”

A safer instruction is closer to:

Recreate this CSV verbatim. Do not change anything else. Make only the following change. Everything else must remain exactly as provided.

Then verify it.

Compare the original and the result. Make sure the intended fields changed and nothing else did. If the workflow involves code or important files, use version control such as Git or another rollback mechanism.

This applies beyond CSV files. If an AI system is transforming important business data, “the output looks okay” is not sufficient verification.

How to Know the Automation Worked

Verification should be designed before the workflow runs, not invented afterward.

A successful task should meet the completion criteria in the original instructions, do everything it was asked to do, and avoid doing anything it was not asked to do.

For analytical work, compare findings against the underlying source material and have the responsible subject matter expert review important conclusions.

For transformations, compare the before and after states.

For workflows that modify systems, maintain logs, backups, version history, or another way to determine exactly what changed and reverse it if necessary.

For operational recommendations, put the results in front of the people responsible for the work before blindly acting on them.

AI output should be reviewable.

The Actual Automation Process

1. Ask employees which tedious computer-based tasks consume the most time.
2. Find the tasks that have the highest chance of being automated successfully.
3. Map what information a human needs to perform the task.
4. Give AI the minimum access required to obtain that context, preferably read-only at first.
5. Define the goal, completion criteria, good output, bad output, allowed actions, and prohibited actions.
6. Test the workflow on a small sample.
7. Review the result with a subject matter expert.
8. Fix the instructions and edge cases.
9. Expand only after the workflow performs correctly.
10. Add write permissions only when they are actually necessary, and put stricter approval and recovery controls around them.
11. Schedule recurring workflows only after the manual process has proven reliable; see How to Schedule Personal and Business Automations to Run Without Manual Work.

The goal is not to automate everything.

The goal is to systematically find expensive, tedious computer work that people are already doing and determine whether AI can perform the repetitive portion safely and correctly.

When it can, that is where useful automation starts.