Being able to identify which workflows you should be automating is half the battle.
While anything technically could be automated with AI, there are types of tasks that are especially ripe for automation.
I'll be describing the patterns we've noticed among the work we've seen being automated with AI and the general rules you can use to spot it.
AI 🤝 Automation
Automating work has never been more accessible thanks to breakthroughs in AI.
Large Language Models (GPT, Claude, Llama etc.) have allowed logic and reasoning to be accessible at scale, anytime, anywhere for cheap. This means work that used to require basic human reasoning can be automated for the first time in history.
Work Before AI Existed
Before LLMs existed, anytime you had a task that required even the slightest amount of reasoning or logic you had 4 options.
1) Do it manually
When a workflow is critical, you'll often find companies "sucking it up" and wasting valuable working hours performing the task.
2) Outsource the work
Large corporations frequently outsource the work to a firm that can do it at scale for cheaper. This is a viable option but it's hard to govern the quality of the outputs and there is significant overhead of in getting an operation like this up and running.
3) Build custom software
Many organizations with an engineering team will opt for creating custom software in-house when work desperately requires automation.
This has been the most scalable approach historically, but it also poses a few issues. It's extremely expensive due to the cost of engineering hours, drawing the engineering talent away from the core product they were building. It also requires significant communication from both teams involved to specify requirements and plan the development.
These same custom software solutions also tended to involve proprietary ML models which added an additional cost to the automation project.
4) Don't do it
This one sounds ridiculous but it's probably the most common route. There's undoubtedly tons of work you wish you could do at scale but haven't yet because automating it didn't seem feasible.
What should you automate?
We like to break this down with two questions. If your answer for both of these questions (using the tricks below) is yes, start automating!
1. Is this worth automating?
2. Is this possible to automate?
Is this worth automating?
Automation Value = Repetition x Pain x Potential for Growth
Repetition:
This is the most common signal we see that something is worth automating. If it's being performed frequently enough across your organization to impact productivity it often should be a priority.
If you're considering automating a task for your business but it's being performed by one employee twice a week it might not be worth addressing. We wouldn't say this work absolutely shouldn't be automated, but this should be taken into account before diving into creating complex automations.
Pain:
How much time is it taking you and how painful is the work. Sometimes a task is performed 200 times a week but it only takes a few seconds. This sort of work would be more of a mild annoyance than a pain that absolutely needs addressing.
Potential to Grow:
Does this use case have the potential to become more common as you grow (ex: customer support tickets) or do you want to encourage it to become more common (ex: personalized sales outreach to prospects)
If you have a workflow that could scale in the future, then it may be worth setting the foundation and automating it early. At times, the sheer existence of the original automation can enable the workflow to scale because all friction has been removed.
Can it be automated?
This is the golden question. If it can't be automated, might as well stop before we start. There are a couple tricks we use to determine if something can be automated.
Step-by-step:
A rule of thumb we like to follow is
"if you can describe this task as a list of steps, like you would for a new intern, it can be automated"
Most automations are simply a series of carefully strung together steps. That's what most tasks are for a human if you break them down enough.
Building an automation means encoding each step into your creation so that it captures the use case exactly as you understand it. The more easily you can describe the work as a series of steps, the easier the automation will be to build.
If the task requires some very nuanced step that involves extreme expertise, non-trivial open ended research or extensive background knowledge it might not be the right fit for automation. Similarly, it might not be the right task for an intern or newer employee.
Manageable "context":
Some tasks require fetching external context to be automated. The AI might need background information like detailed rubric for a grading task or explanation of the slang in the data it's processing. This is very easy to do and actually one of the beauties of AI automation because easy to pass along background context to the AI.
Problems arise when the context becomes too large. If you need to pass in entire books, 50 page PDFs or large corpuses of data then you will run into issues. In a similar way to if you handed an intern a textbook and told them to accomplish some niche task using that information they will struggle.
AI models currently operate with fixed context windows. A context window is the maximum amount of information you can pass into a single prompt, the maximum context you can provide. Context window sizes have grown significantly in the past year but more does not always mean better. Filling the entire context window on a prompt often causes the model performance to plummet. The models will begin to disregard key instructions, hallucinate and cut corners if your prompts are too long.
Ensuring the context required for the task is manageable is a must.
Note: There are many strategies used to manage context windows. Retrieval Augmented Generation (RAG) is a popular strategy for reducing large chunks of context into only the most relevant pieces for the AI. This is more of an art form than a science and while it's absolutely possible to include in your automation, it makes things much more complex.
Public APIs
A key limitation preventing certain automations is access to necessary data or platforms. It is much more difficult to integrate data sources into your automations if the owner isn't providing an API or worse yet, actively preventing your automation from accessing information.
At gumloop we have many integrations built out and can create new ones very quickly if the data source/platform has an API (ex: we created a Notion node in an hour because their API was easily accessible). We only run into obstacles if for some reason, the API is private.
Note: There are ways around this. Brute force web scraping, automated web navigation and 3rd party APIs exist as means of sidestepping API restrictions. These can actually be easy to implement but it's often a case-by-case situation. Make sure to carefully look into your ability to access data before committing to creating an automation.
Automating with Gumloop
Gumloop is a platform automating workflows with AI and a simple drag-and-drop interface. We're trying to give everyone at a company the tools to be as impactful as an engineer is with AI. Automate your entire workflow (as long as it's worth automating ;) ) with Gumloop's no-code automation builder.
Try it out completely for free here.
Reach out if you have questions
If there's something you think should be possible or an integration the platform is missing please reach out to us at founders@gumloop.com. We're continuously building based on user feedback. More likely than not, if you request something doable, you'll see the feature live on the platform within a week or so.
