Working in the Matrix
Working in the Matrix: Where Does AI Fit in Your Day?
In an earlier post I described an arbitrary measurement plane (the idea of the ‘average knowledge worker’ and the tools they use day to day) and suggested I would look more closely at where Generative AI tools (stand alone tools that to x, y and z) fit in that stack.
It is a thing to do, like feeding Vaal.
If one searches online for “Generative AI in the Enterprise” many of the results are from consulting firms and business publications describing the overall adoption rates of AI tools.
There was a Star Trek TOS episode called ‘The Apple’. Captain Kirk and his intrepid team (minus a few unfortunate red shirts) discover a society that is maintained and cared for by some sort of machine intelligence called Vaal. The villagers take part in ‘feeding Vaal’, which involves providing fuel to the ancient machine. They are fuzzy on the details as to why they have to do this, it’s just something they need to do. Shenanigans ensue, Kirk knows best, and Vaal is destroyed so that the villagers can live more authentic lives.
Looked at in a certain way this is an example of Goodhart’s Law, which states that states that when a measure becomes a target, it ceases to be a good measure. Reporting on the KPI becomes a game in and of itself, and true goals, such as actually doing things that make customers happy, suffer as a result.
A similar situation is conveyed, probably unintentionally, by some of the results we find when we go looking for literature on GenAI in the workplace. The articles come across as if to suggest that how much a company uses AI is the goal in and of itself.
You can see the Matrix, can’t you?
To talk about where in the day to day GenAI currently does fit and possibly could fit, and in the spirit of another previous post, we need a map of the matrix.
To build our matrix (really it’s a table, but we say matrix because it sounds cool) we can use two dimensions that we shall borrow from research on the topic of impacts of GenAI and work.
One dimension is found by describing the kind of work being done via reference to the tool use. For example, ‘working on emails’ or ‘Document’ work. Calling a spade a spade, we can suggest this means working in Word, Excel, Outlook, Powerpoint etc. The tools in the standard office tech stack.
Another dimension is described in terms of the ‘level’ of the work. Is it ‘Routine’ work, is it ‘Collaborative’ work or is it ‘Deep Work’?.
Here is such a matrix, with estimates of percent of the day spend on each task (i.e each cell in the matrix) by our average knowledge worker:
| Tool/Level | Deep Work | Collaborative | Routine | Tool Total |
|---|---|---|---|---|
| Word | 8% | 3% | 2% | 13% |
| Excel | 4% | 2% | 6% | 12% |
| Email (Outlook) | 2% | 8% | 10% | 20% |
| PowerPoint | 3% | 2% | 2% | 7% |
| Teams Calls | 1% | 14% | 3% | 18% |
| Teams Messages | 1% | 6% | 3% | 10% |
| “Just Thinking” | 12% | 3% | 5% | 20% |
| Level Total | 31% | 38% | 31% | 100% |
Note we put in a “Just Thinking” row as part of the tool stack. We are knowledge workers after all, and depending on the kind of knowledge work, for some tasks we need to assign time to that tool alone.
Next we can fill in the same cells again, but this time guesstimating how much our robot intern can help us marked from 1 (low expected use) to 4 (very high expected use).
| Tool/Level | Deep Work | Collaborative | Routine |
|---|---|---|---|
| Word | 🤖 🤖 🤖 | 🤖 🤖 | 🤖 🤖 🤖 |
| Excel | 🤖 🤖 | 🤖 | 🤖 🤖 🤖 🤖 |
| Email (Outlook) | 🤖 🤖 | 🤖 🤖 🤖 🤖 | 🤖 🤖 🤖 🤖 |
| PowerPoint | 🤖 🤖 | 🤖 🤖 | 🤖 🤖 |
| Teams Calls | 🤖 | 🤖 🤖 🤖 🤖 | 🤖 🤖 🤖 |
| Teams Messages | 🤖 | 🤖 🤖 🤖 | 🤖 🤖 🤖 |
| Just Thinking | 🤖 | 🤖 | 🤖 |
I gave the Just Thinking row a single point, rather than zero robot heads, because of Generative AI tools utility as a sounding board or mirror.
Now if we wanted to get really fancy we could do a couple of things:
First, we could compare the number of robot heads in a cell with the estimated time spend in the cell, and look for the biggest bang for our robot buck. Routine and Administrative email (10% of time, four robot heads), for example, would become a target for further investigation. So would collaborative Teams calls (14% of time, four robot heads).
Second, we could think about how much time we could reasonably claw back. With email this seems like a good opportunity to spend less time writing routine emails. With Teams Calls (aka time spend in meetings) I am not sure any time gets saved through automation. Human beings are still spending their time in the meeting, and no matter how much they can ‘multitask’ (on the call but doing something else, like emails) meetings are a well documented drag on productivity.
There would be another thing to keep an eye out for, which is the potential need to add a row for thinking about what the AI did. This ‘Verification Tax’ moves any profit we got from reducing cognitive load writing emails and reallocates some of it to checking what the robot intern did.
The Generative AI ecosystem response to a Verification Tax is to promote things called Agentic Workflows. I plan to visit AI Workflows and AI Agents in future posts, but the short version is we do not check as much of the work done by the robot intern at the individual task level, and instead we check on the results of something like a small project.
In the meantime, my suggestion would be to try building your own matrix, with your own estimates of 1, where you spend your time and 2, how much GenAI can help you in those individual cells.
https://en.wikipedia.org/wiki/Goodhart%27s_law
https://www.microsoft.com/en-us/worklab/work-trend-index/breaking-down-infinite-workday