The accuracy trap
Here is the part most takes on this miss.
In M&A advisory the money on the line is enormous, and the accuracy of the model matters more than almost anything else about it. If a machine builds it, someone has to go through and check every single detail anyway. Once you are checking every cell, you have ended up back in the same place, and it can be faster to build it yourself and know.
I run a software company, so I watch the same pattern in code. AI writes something quickly and then you spend a long time correcting it. That will keep getting better, and I assume it gets better fast.
The reason it bites harder in banking is time pressure. When a deal needs to happen, it needs to happen very quickly and with extreme accuracy. There is no room in that window for a long correction cycle.
So the industry lands somewhere that sounds stupid and is economically rational: it is faster to have someone stay up all night and do it. Banks pay well enough that they can expect that out of people.
Consulting is more exposed, and the reason is worth knowing
A banker who spent five years in healthcare coverage framed the contrast better than I do. In banking, money changes hands. You have to close the deal. In consulting, the firm shows up, does the diligence, hands over its thoughts, and quite often nobody reads the report anyway. The obligation is different, and the tolerance for an answer that merely looks plausible is different with it.
His frame for the tool itself is Excel. Investment banking existed before Excel, and Excel did not wipe out investment bankers. It let them do more. AI looks similar from a seat inside the work. It can give you a preliminary baseline, help you understand the drivers, let you tweak things quickly. You still proof all the numbers at the end, and a human still takes responsibility for them.
His forecast is the one I would bet on. Fewer analysts doing grunt work, because a lot of grunt work gets eliminated. The industry still here.
The blocker is often not a modeling problem
Let me defend consulting for a second, because there is real nuance in it and the nuance is the point.
Every company contains pockets of information that nobody fully understands and that sit in no system anywhere. I was on a post-merger integration once where we kept staring at two functions that should obviously have been combined and had not been. We looked for the structural reason. There was not one. The two groups just did not like each other. There is no way a model was ever going to tell us that the group director here has a problem with that guy over there.
What fixed it was running a workshop and getting them to get along.
That is not an exotic case. Businesses are full of minute human details that get overlooked, and people keep expecting a centralized system that somehow knows everything to absorb them.
What a company is worth gets decided by people in a room
There is an essay by the economist Friedrich Hayek about price being something close to the pinnacle of information. A price communicates an enormous amount without needing much explanation attached to it.
Now point that at what bankers do. What is the price of a company? It is genuinely hard to say. There is no objective basis sitting underneath the number. We back into it with models and explanations, which is what a DCF and a comps set are doing. They give you a defensible route to a figure.
Then get into an actual room. When a board is discussing a merger, a lot of what moves the outcome has nothing to do with finance. Who is going to run the combined company. What this director thinks of that executive. Whether the CEO is upset this week or in a good mood. Those things swing the result considerably, and none of them are quantitative.
That is why no centralized system gathers all of this and replaces it.
Why a client picks a bank at all
Look objectively at the advice different banks give when they go in and pitch. It is generally not that different from firm to firm. So how does a client choose? They pick somebody they trust.
A lot of a deal happens outside the deliverable. Deals get made on the golf course. The ideas that change the shape of a transaction often get brainstormed in conversation and never make the pitch deck. That strategic layer comes from insight you cannot train a model on, and it is a good reason to read this job as a service business first.
Augment is the word I would use
AI augments the job. That is the whole of my forecast, and I would not stretch it any further than that.
What it leaves behind is a requirement about people rather than about tools. Someone who has technical ability and emotional intelligence at the same time is the one who can navigate the things happening in that room, understand what is actually being demanded, and pick up the minute information floating around it.
How to tell if this has expired
Three markers to watch, since this is a dated take and should be treated as one.
- The check-it-anyway ratio flips. Today, verifying a machine-built model takes long enough that building it yourself stays competitive. If verification gets meaningfully cheaper than construction, the central argument here weakens.
- Analyst classes change shape. Fewer analysts doing grunt work with the industry intact is the forecast above. Sustained cuts to junior hiring would confirm the first half of it.
- Something moves in the room. The parts that hold up best are the ones happening between people: the trust decision, the personal politics inside a board discussion, the two teams who do not like each other. I do not expect that to move soon.
For now, M&A advisory is probably one of the last things to go. That is a claim with a mechanism behind it, and mechanisms break. Check it against what you see when you get there, and if you are still building toward that seat, the learning pathway is where to start.
