On the Collapsing Premium of Generic Analysis - Thoughts on AI, v0.3

Let us think from first principles about what happens to policy advice now that intelligence is capital.

I have been thinking about how think tanks that advise the Indonesian government, and those that advise the EU or other European governments, should position themselves. I do not know whether these consultants and their clients realize that the premium on generic analysis is basically collapsing. A lot of what used to be billed as thinking is now first-draft synthesis. We are probably only being hired to do verification, and even that is a different job than it used to be.

Note: the lag is probably longer in Jakarta than in Brussels. The direction is the same.

Why Outsource the Thinking?

Imagine it this way. If I were the president of Indonesia, or sitting on a case team in Brussels, why would I pay LPEM FEB UI or McKinsey, or wait for Bappenas to write me a memo? I could buy a bunch of GB300s, run a local GLM-5.3, and get a report that looks the same, if not tighter, after a few iterations. Outsourcing to the local machine is better because it stays private, and the cost of going from a super abstract thought to something that looks like a policy memo or a Statement of Objections annex is now close to zero.

This is the same shift as before. Intelligence used to be an enhancement to labor. It is rapidly becoming capital. If you do not own the capital, you rent other people’s synthesis at this year’s prices. That trade is dying for the generic memo. It is not dying for the person who still has to put a name on the file.

The Verification Trap

OK, so maybe we reposition as a verification service. Fine. But look at what that actually means.

Previously we did the whole stack: frame the question, write the model, estimate it, turn it into a memo the client could use. Now a government can produce an internal draft from its own machines. Not every ministry will. Not every DG will. Many still cannot specify the question, do not own the data work, and do not want an unsigned internal product. The tendency is still clear. When they can produce the draft, our job shrinks to checking it.

For an IO model, that is going line by line through whether the nested logit tree is even the right partition for Indonesian cement, European groceries, or ride-hailing; whether the BLP instruments are exogenous given how licensing and exclusive dealing work in that market; whether the outside good is doing mechanical work to deliver the markup the client wants; whether multi-homing and capacity were just assumed away. That is tedious. It does not scale the way a slide deck or an 80-page report used to scale across ministries and DGs. You cannot productize “spot the fatal identifying assumption” the same way you productized generic analysis. You cannot sell the same deck to five ministries. You sit with one file and argue about one assumption.

So yes, the work that remains is harder in a narrow sense. Harder hours, fewer of them, worse to scale. That is a worse business even if a few people still get paid well to do it.

Courts Will Run the Same Stack

This generalizes past think tanks. Think of a KPPU commissioner, an Indonesian court, or the EU General Court looking at a Commission file. On one side, the authority. On the other, the private parties. Lawyers versus lawyers. Economists versus economists. Two merger simulations, two demand systems, two stories about diversion and efficiencies.

The judge used to outsource the thinking because they had no capacity to evaluate whether the demand system, the diversion ratios, or the way efficiencies were modeled was garbage. That is no longer true in the old sense. Even if the judge does not understand IO, AI is now extremely good at explaining the argument from absolute scratch, one assumption at a time. Give it a harness and some scaffolds, the same checklist as above, plus the jump from reduced form to welfare, and it can flag the standard sloppiness: IIA left untested at the level that actually matters, an outside good that manufactures the markup, no robustness to the nest, a welfare claim that is not in the estimates.

Market-specific invalidity is different. A scaffold does not, by itself, know that this instrument dies because of how licensing works in Indonesian cement or how exclusive dealing works in EU groceries. That fact has to be in the harness, or in the person. Both sides will still run the same stack. The information asymmetry that used to force the bench to fully outsource the technical evaluation is just smaller than it was.

So the remaining premium is concentrated on the economist who can still sit there, under questioning, and explain why this specification is the credible one for this market, Indonesian or European, and what the decision should do if the key assumption is wrong. Data work and institutional mapping are why that person is not a rubber stamp. Putting a name on the affidavit is why the client still needs a person. None of that is the take-home memo.

The take-home assignment is dead because it is automatable. The oral exam is not. Policy advice is going the same way. The take-home memo is dead. The hearing is not.

The first draft is cheap. The live defense of the identification is not. That is the part of the old premium that does not go to zero.