The term Artificial Intelligence has been around since 1956. Scientists quickly figured out that it was more efficient to teach computers to learn than to program them with every single skill needed to complete a given task. This became known as 'Machine Learning'. 'Deep Learning', a subset of Machine Learning that uses Neural Networks’ - a set of algorithms modelled on the human brain, enabled AI to tackle even bigger problems.

Nowadays we hear all the buzz around Machine Learning use cases such as Voice Recognition, Medical Diagnosis , and even Self-driving Cars .

But Machine Learning and Deep Learning in particular have enormous potential for the entire economy, so why hasn't every business already started using AI? And what about AI applications for time tracking and personal productivity ?

For one, AI can be super hard to implement: envisioning use cases, hiring a team to implement, and then figuring out how to integrate that AI into a business's workflow can be a complex undertaking.

Second, AI is computationally intensive and expensive. Buying and running the compute power necessary to implement deep learning can be cost prohibitive.

Third, and most importantly, Machine Learning needs huge amounts of input data to yield accurate results. This is why I’m still not sold on AI in Time Tracking, although we can achieve some AI-related claims with what you already have.

So, let me expand on the last point: I’m no AI expert but I know that, before Machine Learning, you had to give your program a set of instructions, each covering a specific case or scenario, similar to this:

• When A happens, Do B

• When X happens, Do Y

Let’s say you’re coding a traffic game:

• When the red light is on, the cars must stop.

• When the green light is on, the cars start moving.

• Etc.

This works well for ‘simple’ programming challenges, but if you had to write a simple face recognition program that sorts images into two categories (e.g. butterflies and cats), you can only imagine the number of instructions you’d need to give to your program.

It’s actually impossible: try for a second listing ALL the characteristics of a butterfly, then those of a cat….

Today, with Machine Learning, you give your program millions of input images of butterflies and cats, telling the program: this is a butterfly, this is a cat, this is a butterfly, this is a cat…

Machine Learning then creates the set of instructions on its own and keeps adjusting as you feed it more data.

So, for Machine Learning to work, you need a very large set of data, the more data the better the algorithm.

Put simply, it’s better to have a team of good AI engineers and tons and tons of data than a Nobel Prize winning team in algorithms (if there was such a thing) and a limited amount of data.

Now let’s go back to Time Tracking: an ideal algorithm would automatically figure out the client to which to allocate the tracked data.

This one is by far the biggest problem to solve in Time Tracking, just think about it: a piece of software is installed on all digital devices you use during the day, collecting information on your work, location, files you open, people you contact… and, by the end of the day, all those time entries are accounted for and allocated to different clients, contacts, tasks, projects….

This is huge: no more stopwatches, timers, manually entering time entries, going through your inbox history, call logs or agenda… to build your timesheet.

Also imagine you are part of a team: no need to check on every team member to add their timesheets, or if you are a timekeeper no more chasing your boss at the end of the month because you need to submit their hours….

AI time tracking would solve this with a push of a button, right? I highly doubt that. I have personally tested some apps that (claim to?) use AI time tracking and I ended up spending more time correcting the automatic allocations the apps created than logging my time. Maybe I did not give it enough time, and maybe after weeks or months of training the algorithm would have given better predictions but:

a. My work, contacts, tasks and projects will probably be different three or six months from now, so I’d need a new algorithm;

b. As mentioned earlier, for Machine Learning to give accurate results it needs a good algorithm and lots of data, and we’re not talking about megabits or gigabits of data, but much, much more, in the order of Terabits.

You’d need to feed the machine loads of input and expected results to fine-tune the algorithms. This works well for self-driving cars, Google search (made on all of Google’s users), Amazon and recommendation engines, done on millions of viewers and shoppers, but will not work on the relatively small amount of data collected from my work log for the day, even for a year.

So, is all hope is gone? Will Time Tracking always be a necessary and boring task, like brushing your teeth?

Probably not. But maybe the answer is not in AI. The thing is not everything in our work has be about finding a solution in AI, even if it’s compelling to think this way.

I mean, wouldn’t we all want to be able to take a magic pill that would make our worst work nightmare go away? Well, I’ve been in this industry long enough to know that with a simple and disciplined approach, you could achieve 80% of the AI claims – no need for a magic pill!

How is that possible?

Think about it: most of the time the thing you’re working on has an indication of the task or client you want to allocate it to. If it’s an email, then the recipient’s address can give you the client name, or a hint. If it’s a spreadsheet then it’s the name of the document or the case number in the name of the file. If it’s a call, you can easily look up the contact name in the address book. When nothing of the above applies, you can safely assume that the work is related to your calendar….

None of this requires AI, a ‘rules’ engine telling the software to take an action based on a keyword or a combination of keywords will suffice.

Of course, this won’t cover 100% of your work and you’ll still need to review the automatic allocations made by the app, but it does the heavy lifting and does not require years of training an AI algorithm.

There are also lots of scenarios in which adjusting the output makes for a tedious task. Imagine you start working in a document in MS Word; originally MS Word, names it ‘Document1’. You then spend a good 25 minutes working on that ‘Document 1’ before renaming it to, say, ‘Research for Project X’. Even if you create a rule matching Project X to a client, you need the engine to be smart enough to assign the first 25 minutes - when the document was ‘Document1’ - to ‘Research for Project X’.

At Chrometa- yes, that’s us - we have built a powerful rules engine that automatically allocates collected time based on keywords (single keywords, combinations of keywords, email addresses, domains, phone numbers, website addresses….) and can also go back in time and adjust any gaps not covered by the rules. We don’t claim it allocates 100% of the time it collected as we honestly don’t know how any software could do that but it’s very good at recognizing patterns and keywords, and it offers a simple interface to work with.