Which Business Jobs Are Actually Worth Automating?
Ben Crowther is the founder of Crow Flies Growth Systems and co-founder of Aiden Solutions. He helps businesses replace static websites and fragmented marketing with AI-powered growth systems built to convert, support and scale.
The best first automation is usually a repeatable job with clear rules, a visible result and a way to recover when it goes wrong. The impressive demo can wait. The first task you automate probably shouldn’t be the one that looks best in a demonstration. Start with something your team does repeatedly and understands properly, where they can tell when it’s gone wrong.
That might be acknowledging an enquiry, assigning it to the right person or checking whether a record is missing important information. There’s little glamour in any of it. There’s also little glamour in paying someone to chase the same missing details every Tuesday.
Ask staff what they spend time correcting, chasing or entering twice. The answer may be a forgotten form field, a missing attachment or a routine status update. If that work happens every day and the rules are stable, there may be a useful starting point. If the team only encounters it twice a year, the business case will look rather different.
The temptation is to start with the impressive part: an artificial intelligence (AI) system making decisions, generating proposals or running the customer conversation. Sometimes that’s appropriate. Sometimes the business has simply skipped the less exciting job of agreeing how the process should work.
Use the simplest tool that works
If the job is “when this form arrives, create a record and notify the right person”, it may need an ordinary workflow based on rules. It does not need a language model to develop an opinion about who should receive the email. Clear rules are easier to test, cheaper to maintain and less likely to surprise you.
AI can help when the input is less structured, such as summarising an enquiry or finding an approved answer in a large collection of documents. Even then, decide what the system is allowed to do with the result. Preparing a draft for a person is a different level of permission from sending that draft to a customer or committing the business to a price.
I would start with the part you can explain. If you cannot describe the decision, test the result or see when it has failed, adding another layer of AI is unlikely to make it easier to manage.
Find a task with a clear answer
I’d look at how often a task happens, how much time it takes, the cost of getting it wrong and how clearly you can write its rules. Check whether the information it needs is available and whether you can reverse an incorrect action.
Ask what the task costs today, including the interruptions it causes. A job that takes two minutes may still matter if it happens a hundred times a week. A job that takes twenty minutes once a quarter might not justify an integration that needs somebody to maintain it. The calculation of savings needs to account for checking, corrections and support, not just how fast the demonstration runs.
An automatic reminder for a confirmed appointment can be fairly straightforward. A recommendation that depends on an expert interpreting someone’s circumstances is a different job. Both can be described as automation, but that doesn’t give them the same level of risk.
Think about the consequences of a mistake. A duplicate internal reminder is annoying. An incorrect payment instruction, public technical claim or sensitive recommendation can do real damage. Those jobs need different safeguards, and some should remain with a qualified person even when software can help prepare the information.
If the task changes depending on who you ask, settle that first. Software can’t resolve a disagreement nobody has acknowledged. It’ll just follow whichever version happened to make it into the build.
Write down what should happen
Before choosing a tool, write down what starts the process, what information it needs and the rules for the next action. Name the person responsible. Explain how you’ll know it succeeded and what should happen when it fails.
Take a hypothetical enquiry workflow. A form arrives, the system checks the service and location, creates a record and assigns an owner. What happens if the location is missing? What if it’s outside the service area? What if the record system is unavailable? These questions matter more than how quickly it runs when everything goes to plan.
Now add the uncomfortable cases. If the form is incomplete, does the customer get asked for the missing detail, or does the record quietly fail? If it is outside the service area, who decides whether an exception is possible? If a required system is down, where is the information held until the connection recovers? The answers belong in the design, not in a frantic conversation after the first failure.
Keep a unique identifier for each run, record when it starts and finishes, and prevent duplicate actions. Some temporary failures can be retried safely. That doesn’t mean sending the customer the same message with every attempt. Preserve the input and give someone a way to recover the work manually.
Give the person taking over enough context to act. An error notice saying “something went wrong” is not a handover. It should identify the affected record, what the system attempted and what is safe to retry. Test a duplicate submission, a partial failure and a manual restart before expanding the workflow.
Keep customer information under control
A workflow may move names, contact details, purchase questions and internal notes between systems. Before connecting them, check what information is actually required and who should be able to access it. The Office of the Australian Information Commissioner (OAIC) guidance on collecting personal information is a useful starting point for Australian businesses that need to establish their obligations.
Do not give an automated assistant access to an entire customer database because it needs one field. Keep permissions narrow, make the records of its actions available for review, and decide what happens if information is wrong or a person asks for it to be corrected. Whether the law applies in a particular way depends on the business and the information involved, the sensible design principle is to collect and expose no more than the job requires.
Earn wider permission
I’d keep human approval for proposals, public performance claims, sensitive advice and first contact with a prospect. Preparation can often be automated while the decision stays with a person. That can still remove a lot of repetitive work.
For work assisted by AI, the National Institute of Standards and Technology (NIST) AI Risk Management Framework describes the importance of defined human oversight, testing and monitoring during operation. You do not need a committee to send an appointment reminder. You do need to know who can stop a system and who is accountable when it gets a consequential decision wrong.
For a publishing workflow with limited risk, I’d suggest starting with four clean weeks of reviewed output before considering automatic publication of a preapproved evergreen template. That’s a practical checkpoint, not an industry standard or proof of safety. Volume, subject matter and the consequences of an error may call for a longer review or keeping human approval permanently.
Likewise, three repeated failures can be a useful trigger to pause the affected action and create a recovery task. Serious failures may mean stopping immediately. Keep the records and alert an owner, don’t let a dashboard turn missing data into a reassuring zero.
A useful alert should tell someone what needs attention. It should not generate fifty new alerts because the same connection has failed fifty times. Distinguish the original customer request from repeated attempts, pause only the affected action where practical, and make recovery part of the normal work.
Build one narrow version and test duplicates, missing information, unavailable services and a manual restart. Keep sampling the output after launch. Useful automation is something your team can trust enough to stop checking constantly, while still knowing how to step in.
Measure whether automating the task saves work
Record the starting point. How long does the task take? How often is information missing? How many corrections does somebody make? Then compare a small, supervised run with the old process. Look at completed tasks, time saved, errors caught and how often a person still has to intervene.
Include the work required to keep the automation running. If it saves one hour a week and creates two hours of checking, debugging and explaining strange results, that is a fairly poor exchange. Equally, a workflow that saves little time but prevents expensive omissions may still be worth keeping. Decide what benefit you are actually testing.
Keep reviewing it after launch. Products change, staff move roles and customer questions change. A workflow that was correct in November may be wrong in February if nobody owns its rules or information. Make someone responsible for those changes.
My recommendation is to pick one repeated job, write down the steps and run the failure cases before buying anything. If it works and people can recover it when it does not, expand from there. Automation should leave the business with less work to chase, not another system to babysit.
Read more from Ben Crowther
Ben Crowther, Founder of Crow Flies Growth Systems & Co-Founder of Aiden Solutions
Ben Crowther is an Australian entrepreneur, marketing strategist and artificial intelligence (AI) advocate focused on how technology is changing the way businesses attract, convert and support customers. He is the founder of Crow Flies Growth Systems and co-founder of Aiden Solutions, a knowledge and customer support platform powered by AI. His work combines digital strategy, search engine optimisation (SEO), answer engine optimisation (AEO), generative engine optimisation (GEO), automation and conversational AI to turn traditional websites into active business systems.
Ben is particularly interested in the shift from static online brochures toward intelligent digital experiences that can sell, answer questions and support customers around the clock. Through his writing, he explores AI, entrepreneurship, marketing and the future of digital business.
