The I-Beam Society
The two limits that created the middle of the corporate pyramid are loosening. What’s left may slowly come to look like an I-beam.
The pharma sales force I know best is a textbook pyramid. From the bottom up: reps, district managers, regional managers and the BU head. A district manager’s week usually goes like this: collect the reps’ weekly reports, combine them into a district report, attend the regional meeting, then come back and explain headquarters’ latest requests to the reps. That takes up most of the week.
Look closely and it’s two jobs: passing information from below up the line, and passing instructions from above down it. This layer of work may well be the first place AI changes. The middle of the pyramid will slowly cave in, leaving a shape like the Chinese character 工, or an I-beam seen end-on. The top bar is the people who sign off and carry responsibility. The bottom bar is the people doing the work with a team of AI agents. The thin stem between them is the people whose job is to make the organization comfortable handing decisions to AI. Many companies feel what they’re short of is people who can use AI. My sense is that the stem is harder to find.
Where the middle came from
The pyramid was forced into existence by two limits. One: few people make good decisions, so the important ones get pushed upward. Two: one person can directly manage only seven or eight others, so a company of ten thousand has to stack layer on layer.
The people in the middle do the two jobs described above. Passing information up means boiling the hundred things happening on the front line down to the three the boss has time to read. Passing instructions down means turning one sentence of headquarters strategy into specific actions the front line can understand and carry out.
The internet made moving information almost free long ago, yet the middle didn’t disappear, because summarizing and relaying were never just transport. Deciding which item deserves to go up, or how a strategy should be put to this particular region, takes judgment. Moving information became free, judgment stayed expensive, and so the middle stayed.
Now judgment can be bought as well. Reading a hundred weekly reports and picking out the three that matter is exactly what large language models do well. They also do a good job of rewriting a strategy into specific advice for each rep, based on that rep’s customers. Important decisions no longer all have to be pushed upward, and one manager can handle more than seven or eight people. With both limits loosening at once, the middle layers no longer have to be filled by people.
The three strokes
There may be fewer people in the top bar, and each will carry more, because responsibility can’t be outsourced. AI can make a judgment, but it can’t sign anything, and it can’t sit across from a regulator and explain itself on anyone’s behalf. Every one of the thousands of decisions the system makes each day still has to end up under some person’s name.
In the bottom bar, some things still need a person there: building trust face to face, reacting to what happens in the moment, work done by hand. A rep still has to go and see doctors in person. But what to prepare beforehand, what to record afterwards and what to do next can be handed to a team of agents. This layer may not shrink much, but each person will be able to cover far more customers.
The stem has very few people in it. They pick which decisions can go to the system, write the business rules into it, then watch how it performs and judge when to give it more authority and when to take some back. Whether the top and bottom bars connect depends largely on them. I share some thoughts on how they might go about it in AI Has to Earn Its Decision Rights.
Where the line falls
A popular argument goes: AI will replace white-collar workers, so young people should learn plumbing. It’s half right. Plumbing is hard to replace, but that has little to do with it being blue-collar work. What makes it hard is that it can’t be written down as rules. Every house runs its pipes differently, there’s always a surprise on site, and the work takes hands. By that standard, a lot of white-collar work is also hard to write down as rules: negotiating a contract with complicated terms, persuading a key person not to leave, stepping up to take the blame when something goes wrong. Meanwhile picking orders in a warehouse, though blue-collar, is easy to write down.
So it may be more accurate to draw the line by whether the work can be written down as rules. If both the inputs and the outputs of a job can be spelled out clearly, it’s relatively easy to replace, whether it’s done in an office or on a factory floor.
The people in the middle
If a job mainly consists of tidying up information from below and passing it up, then passing requirements from above back down, that position is likely to feel the impact first.
The position may go, but the person still has somewhere to go. That district manager does more than summarize and relay. They go on visits with their reps, can see where a new hire is stuck, and keep the team steady when morale is low. None of that can be written down as rules, and it’s what makes them valuable. There are roughly three ways out, one for each stroke of the I-beam: move up and take on a real target along with the responsibility for it; move down, back to the front line, working with a team of agents; or move into the stem.
For companies, the hard part is the structure around people. Performance reviews, promotions and headcount are all designed for the pyramid, where being promoted means managing more people. What level should someone be if they work with twenty agents? How do you promote someone with no direct reports who decides how much authority the whole company gives to AI? I haven’t seen good answers to either question, and I haven’t worked them out myself.
