An annoying task does not automatically justify automation. Before investing in AI, weigh the time saved and errors avoided against the cost of setting up and running a solution, including the data needed to test it and the work of handling mistakes. Sometimes delegating the job to a person is the better answer.
Ask your employees which repetitive tasks consume their week or cause corrections further down the process, such as copying information between systems. Assess the process before choosing an existing AI tool, an AI agent that carries out steps on your behalf, or custom software. The questions below can help you decide which ideas to investigate and how difficult they would be to automate. That difficulty affects the cost and results you can expect.
Would automation save enough to justify its upkeep?
Could someone do this job for a few hours a week, with little supervision? A task can be boring without being expensive enough to justify automation. An existing tool may keep costs down, while custom development adds work before you see any savings. Include subscriptions or hosting costs alongside setup, integration, human review, and maintenance. Investigating failures takes time too, even when routine operation is cheap.
Repeated data-entry mistakes can make colleagues distrust the output and check everything again. Preventing those errors can be valuable, but you need to measure whether automation reduces the work of checking and correcting results before assuming it solves the problem.
Does this process need AI?
AI has expanded what software can automate, including interpreting free-text requests and extracting information from varied documents. Moving structured records between systems may only need an existing integration, and you can use AI for the part that needs interpretation while keeping predictable work in predictable software.
What feels difficult to a human does not always match what is difficult for a model. Producing a fluent translation can be easier than writing text in your company’s voice, even though fluency alone does not establish correctness. An engineer familiar with current capabilities can help assess the task.
Do you have enough representative data to test the idea?
Suppose you want to extract information from reports. Can you provide a representative sample of real reports and explain what the correct output should contain? Include ordinary work alongside unusual layouts or missing information, with conflicting values and other errors in the source data. Cases that require human review belong in the sample too.
Some examples may already have checked answers, such as values employees entered and verified. Others still need employees to review and label the expected output. That preparation takes time and is part of the cost of testing the idea.
Volume matters too. A few selected examples give little evidence about how often failures occur, while many nearly identical examples can miss an entire category of mistakes. You need enough examples across the kinds of work you expect, with extra attention to rare errors that would have serious consequences.
If you have little historical data, collecting examples can be the first experiment. You can also create cases to test specific failures, as long as you keep those separate when estimating everyday performance. A convincing demo tells you little about your own process without representative data behind it.
What restrictions apply to your data?
Personal information and trade secrets do not make automation impossible, but they change how you can build and test it. Hosted models can make early experiments easier when you know which information leaves your systems and under what terms. Check processing locations, retention, access, and whether a provider may use your inputs for training. The agreement also needs to cover what engineers may use during development, including in logs and debugging tools.
Running models yourself may fit some restrictions, but adds operational work and can limit the models available. Removing sensitive fields is another option where the remaining data still represents the task. Account for the cost of the hosting and data preparation your restrictions require when assessing the likely savings. Make those decisions before uploading real documents to an experimental system, with specialist support when the consequences of mishandling data are serious.
How costly are mistakes, and can you reverse them?
Preparing invoice details for a person to review is different from authorising a payment, just as drafting a customer reply is different from sending it. An incorrect action has consequences beyond an error in a draft, even when both use the same information.
Tasks are easier to automate when errors are cheap to find and fix. When mistakes could lose money or expose information, ask who would notice a failure and whether you can reverse the action. Raise the standard for those tasks, including ones where damage to a customer relationship would be difficult to repair.
Human approval helps only if the person has enough information and time to check the result, without spending as long on review as the original task took. Higher-risk actions can require more checks and review, reducing the savings. Preparing a draft without taking action can be a better first test.
Can someone or something check the output?
A clear way to check correctness makes a task a stronger candidate. Software can compare an extracted invoice total with checked source data or check whether a customer number exists. It can reject impossible dates too. Judging whether a summary leaves out an important qualification may need a domain expert.
When different answers could be valid, define a rubric that explains what a good result must contain and which mistakes are unacceptable. For a summary, it might require preserving conditions that affect a customer’s decision. Another AI model can act as a judge and score outputs against that rubric, provided you check its agreement with human reviewers before relying on its scores. Include the effort of defining the rubric and checking the judge in your assessment.
This feedback is one reason AI can be useful in software engineering and mathematics. With automated feedback from tests and proof checkers, an AI system can revise its work and run the checks again until they pass.
For your process, identify which checks can run automatically and which need a person, including how the system should respond to failures. An unreadable amount should prompt a request for review instead of a plausible guess. If checking each result takes as long as doing the original task, automation may offer little saving even when the output looks convincing.
Decide what to test next
Before testing a tool or commissioning a prototype, write down:
- The process you want to improve and who benefits.
- The time, cost, and errors in the current approach.
- Simpler alternatives, including delegating the work.
- Available examples, the work of preparing checked answers, and restrictions on using them.
- What a good output looks like and how you would check it.
- Consequences of mistakes and the review or recovery needed.
Use these answers to compare the likely benefit with setup, data preparation, review, and ongoing costs. A process that is technically possible to automate may still cost more than the work it saves. Missing answers identify what to investigate first, whether you need more data or a narrower task. You might find that existing software solves the problem, or that the costs rule out automation.
If the idea looks promising, our next article explains how to test your idea with an AI proof of concept before investing, starting with success criteria and a simple prototype tested against real examples.
Frequently asked questions
Should we automate every repetitive task?
No. The savings must justify development, review, and ongoing upkeep, and a person doing a low-volume task may remain the simplest option.
Do we need clean data before starting?
You need examples you can interpret and check, including messy inputs if those occur in daily work. Cleaning every test example can hide the problem the system needs to handle.
Does sensitive data rule out AI automation?
No. It constrains data access, hosting, and development methods, which you need to assess before choosing tools or sharing data.