Monday, October 26, 2020

Using NLP to enable flatter organizations

You’re a developer working on the Foobar feature at company X. Unbeknownst to you, Genghis, the company CEO, just had dinner with other industry leaders and became excited about this new thing called Beezlyboo. Genghis mentions his excitement at Thursday’s executive council meeting, with the company’s top layer of middle management listening in.

Mandukhai, the middle manager two layers above you, was present. She conveys the takeaways in her regular Friday morning coffee catchup with the managers that report to her, including the Foobar team lead, your boss, Subotai.

Fighting for relevance after a bumpy year, Subotai grows convinced that working on this Beezlyboo thing might be his ticket back into Mandukhai's good graces.

You enter the office on Monday morning, unsheathe your sword, and balance it carefully on top of your monitor before sitting down and getting to work. Little do you know that two weeks from now, half your time will be spent on this new thing called Beezlyboo.

How Middle Management Emerges

This is an unfortunately common description of how decisions are made in many of today’s Khanates companies. Why is that? Let’s take a look at how communication and decision making evolve as companies grow larger.

Here’s how many companies begin:
2 people companies

Two founders, working closely to get their new startup off the ground.

A more mature company might look more like this:
Small company

That’s Google circa 1999. Individuals hold increasingly defined roles. Management and process are beginning to creep in, but all anyone really wants to do is get on with things, management be damned. Grabbing a colleague to discuss something is as easy as shouting across the office.

But as companies grow larger, the majority end up looking more like this:Large company

This is of course an over-simplification as every organization maintains informal networks, cross-functional teams, and parallel networks. But it is a useful representation to understand how hierarchies simplify the sharing of information by creating highways for it. Additionally, we will assume that every team operates in a mesh (even if we were not skilled enough to draw those relations).

See those ‘o’ nodes? That’s middle management. Layers and layers of it. Its main purpose is to simplify decision making and communication by concentrating ownership of different parts of the business in the hands of a select few.

For the more mathematically inclined, middle management splits the previously shown complete graphs with Kn=n(n−1)2K_n = \frac{n(n-1)}{2}Kn=2n(n1) edges into multiple complete subgraphs with Mn=∑inKm(i)+1M_n = \sum_i^nK_{m(i)+1}Mn=inKm(i)+1 edges, where iii runs over managers and m(i)m(i)m(i) is the number of people they manage. We assume that reporting edges correspond to lines of communication and that members of a team are fully connected in complete graphs.

The benefits are clear. When CEO Genghis or Kublai, another middle manager, want to figure out what’s going on with the Foobar team (highlighted above in Millenial pink), all they need do is call Subotai. By concentrating ownership of Foobar in Subotai’s hands, his team is shielded from unnecessary communication, allowing them to focus on their work.

To illustrate: K5000=12,497,500K_{5000} = 12,497,500K5000=12,497,500, but introducing 556 managers2 each with m(i)=10m(i) = 10m(i)=10 reports gives M556=30,580M_{556} = 30,580M556=30,580. That's only 0.2%0.2\%0.2% of the original and a radically simpler org structure!

But there's also an obvious downside: poor decision making. While the ‘x’ nodes focus on the work at hand, the ‘o’ nodes are busy talking and aligning amongst themselves. This creates two problems: the ‘o’s lose touch with the actual work being done, while the ‘x’s cannot easily see, communicate, or influence beyond team boundaries. Which, remember, is why you’ll be surprised when Beezlyboo lands on your desk in two weeks' time.

What Companies Are Doing About This Today

So far we've been talking about idealized companies. In reality, there are informal networks that permeate company hierarchies. There are also good and bad ‘x’s and ‘o’s. We know a lot about what a good ‘o’ is (or should be). There’s an entire industry out there dedicated to this, complete with degrees (MBAs) and reading material (those management books you tend to find in airport bookshops).

One thing ‘o’s obsess over is how best to organize things, which is great. That's their job, after all. The problem is that their solutions are often clunky. Do any of these sound familiar?

  • Culture stuff: the ‘o’s realize the company’s formal and informal hierarchies are too strong, so they begin promoting things like guilds, societies, team building activities, random coffee catchups between senior leads and juniors, etc.
  • Reorgs: featuring endless tweaks to the traditional hierarchical org structure (functional, divisional, matrix/cross-functional… you name it). Ironically, it’s the ‘o’s that instigate these — often to the complete bewilderment of the ‘x’s.
  • Surveys: because making the ‘x’s click through an annoying ten-page annual survey is a great way of discovering how they really feel. Right? Right.

If these solutions sound drab and disengaging, it's because they are. Ronald Coase, the 1991 Nobel laureate in economics, published an influential essay in 1937 on why firms even appear, and mentions that the benefits of larger firms are quickly outweighed by the overheads of maintaining their complex hierarchies and the misallocations of resources which stem from this. In a more modern work, Gary Hammel and Michele Zanini demonstrate in Humanocracy -- another one of those airport books -- that:

A

2018 Gallup study

found two-thirds of US employees and 51 percent of US managers to be disengaged at work.

In 2018, there were 146 million employees in the US workforce. [...] Of these, 20.5 million were managers and supervisors. In addition, there were 6.4 million employed in administrative support functions. In total, then, the bureaucratic class comprises 26.9 million individuals or 18.4 percent of the US workforce. [...] claiming more than $3.2 trillion in compensation, or nearly a third of America’s total wage bill.

They argue that, if we could repurpose a fraction of this managerial bloat into more productive pursuits, we would free up a lot of unproductive excess wages and therefore raise productivity, overall compensation (for the rest of us), and employee engagement.

Genghis: Okay, kiddo, that's great... but we're trying to run a Khanate company that stretches from eastern Europe to China. Maybe our team building activities are a little boring, and we do like a good reorg, especially after we invade acquire a new company, but what are you proposing? That we fire all our generals and officers and turn into those Valve hippies?
Ogedei (You): Not at all. Hear me out.

What the Future Might Look Like

With everyone and their cat — not to mention the OECD — talking about the Future of Work these days, isn’t it time we began trying to tackle company organization with more pizazz? You know, just throw some tech and AI at it.

Genghis: Oh god, here we go again with your AI thing.

He's right: here we go.3 One of the most significant transformations due to play out over the next decade is the death of email in within-company communications and its replacement by tools like Slack and Teams. Not only do Slack and Teams introduce a new medium for companies to communicate in, they also allow for modern third-party integrations via their APIs. That’s something email tools like Gmail or Outlook never quite got right.

On top of this, there's an unfolding generational shift. Millennials will continue to graduate to positions of power as the 2020s progress — and they’ve grown up using online chat protocols and social networks.

Ogedei: So here’s the thing, Dad. The way we organize ourselves in companies will change because young people like me communicate in new and different ways.

The first wave of startups attempting to tackle this — like Know Your Team (KYT), Ally, Lattice, Donut, Cadence, and a handful of others — are already well on their way.

Let's take KYT as an example. It's a fantastic tool that takes the magic of being a good 'o' and pours it into a suite of effective communication tools to help you manage your reports. Backed by years of research and a kick-ass CEO in Claire Lew, KYT is a great example of a company pushing the envelope of how we think about management.

If middle managers in today’s corporations are obsessed with airport books and ways of organizing people, then the true organizational geeks are military commanders and theorists. Organizing militaries has been a thing since time immemorial, from Sun Tzu to Clausewitz.4

The Mongols weren’t too shabby either. Instead of Slack, they used a system of postal relays called Yam. Instead of executive councils, they had Kurultais. The Mongols adapted quickly and didn't try to change the religious habits of conquered populations. By virtue of necessity, they also had a pretty devolved leadership structure — individual Khans ruled parts of the empire which were thousands of miles apart. We could learn a lot about how management can be done better from the Mongols.5

What we're building

Say you're a middle manager. You report to someone, a few others report to you. A good chunk of your time is spent synthesizing information (mostly updates) flowing across your node in the graph, from the top down and from the bottom up.

This is where we come in. At Middle Out, we're working on a solution that analyses user conversations, summarizes important segments into updates, and proposes these updates to users for final editing and approval before making them visible to the rest of the company. We then take these updates and show each user the ones our model believes are most useful to them. This is tantamount to getting rid of the fog-of-war that shrouds parts of any company from individual employees. In short, you'd see Beezlyboo coming a mile away.

Why did this not exist before? Why do people still write weekly status reports themselves?

Mostly because the NLP models required to detect interesting sequences of messages and then summarize them simply weren't there only a year or two ago.

To detect important messages we roll our model across each message and consider its context (a window of past and future messages centered around the message we're considering) and score them. A second rules-based model then goes out collecting groups of highly scored messages in preparation for summarization.

You've probably heard of GPT-3 making waves across the tech community, but other models like Google's PEGASUS have achieved impressive results as well... and these are all 2020 models. Imagine what we'll have in 2025. We make use of one of these models to then summarize the highly scores groups of messages into updates which are then proposed to the user.

Now for the cool part. By having access to updates from across the company, we can quickly call to arms disparate employees to begin working on something together. That's open allocation at scale and we're working on a tool to facilitate that too.

We go from static graphs to dynamic ones which adapt to what you're working on and what's relevant to you right now:New company

Genghis:

Nice. Let's axe all our generals.

6 Ogedei:

Wait! Not so fast.

We still need some human managers. Why? Because things like leading a team in a new direction, passing on knowledge, pastoral care, and motivating others are all uniquely human abilities. They're also all abilities that great managers tend to excel at.

In short, we'd like to see a world with flatter hierarchies where employees self-organize and decide on direction together, aided by a smaller set of experienced managers.

Mandukhai: I agree! I'd much rather spend time passing on my knowledge to young recruits than having to deal with leading the Khanate's company's largest army group where all I do is sit in meetings all day!

The Middle Out App

Here are some early highlights of what we're building.

Genghis message

We piece together what you're working on by securely analyzing your communications across chat and email to propose updates to you.

Genghis message

We gently nudge you to keep a curated list of updates by asking you to approve, edit or reject proposed updates once a day.

Updates shared

Accepted updates are shared with the rest of your company and presented to each user through a news feed of updates that are most relevant to them. You can of course also search across updates or look at the updates of a specific user.

By knowing what others are up to you can easily vote on direction together with the right employees from across the company.



from Hacker News https://ift.tt/3e2B4Ub

No comments:

Post a Comment

Note: Only a member of this blog may post a comment.