Every institution that matters runs on desks. The underwriting desk that prices a risk. The credit desk that approves a facility. The trading desk that sizes a position. The claims desk, the procurement desk, the strategy desk. Different industries, same anatomy: a queue of decisions about the future, a body of evidence of uneven quality, and a person — experienced, accountable, busy — converting one into the other under time pressure.
For a century, the conversion mechanism has been discretion. Not recklessness — discretion at a good desk is trained, supervised, and bounded by policy. But at the moment of decision, the final synthesis happens inside a practitioner's head, and the confidence attached to it is expressed in the vocabulary of the desk: I'm comfortable with this. This feels rich. I've seen this pattern before. That vocabulary has carried institutional decision-making a long way. It is also, by any measurable standard, unauditable. Nobody can compute the Brier score of a feeling.
We think the next decade belongs to what we call the post-discretionary desk, and the term is chosen carefully — because the common prediction about AI and decision desks is wrong in an instructive way.
The wrong prediction
The common prediction is replacement: the model takes the decision, the human leaves the desk. Some version of this is recited in every industry we work near, with either enthusiasm or dread depending on the audience. We think it misreads what desks actually do and what the current generation of AI actually offers.
Decision desks are accountability structures, not just decision factories. Someone owns the underwriting result, the credit loss, the position. Institutions cannot delegate ownership to a system that cannot be examined — and the awkward truth about most AI tooling is that it has reproduced the desk's oldest weakness rather than fixing it. A model that answers everything, fluently, with no audit trail from evidence to confidence, is discretion with better grammar. It moves the unaccountable feeling from a person's head into a system's weights, and calls the move progress. Replacing a senior underwriter's intuition with a model's intuition is not transformation. It is redecorating.
The right prediction
What actually changes the desk is something narrower and more structural: the arrival of pre-outcome intelligence — systems whose product is not an answer but a governed probability, delivered before the outcome resolves, with its justification attached.
The distinction does real work. A governed probability differs from a confident answer in ways a desk can hold on to. It arrives with lineage: the evidence pool it drew on, screened for contamination before use. It arrives with a constructed confidence — built from measured dimensions like evidential support, contradiction, novelty, and the system's own historical reliability — rather than an asserted one. It arrives with the system's honest position on the ladder: a calibrated estimate where the evidence supports one, structured scenarios where it partially does, and a refusal where it does not. And after the outcome resolves, it becomes a permanent entry in a track record that can be scored. The feeling has been replaced by an instrument reading.
"Post-discretionary," then, does not mean the practitioner leaves. It means the practitioner's relationship to confidence changes. On the discretionary desk, the human generates the confidence and defends it with seniority. On the post-discretionary desk, the human interrogates a stated, audited confidence and decides what to do about it — including overruling it, on the record, for reasons that themselves become part of the institution's learning. Judgment is not eliminated; it is relocated to where it earns its keep — on the question of what to do — and removed from where it never could be audited: the question of how likely.
The historical rhyme
This transition has happened before, and the precedents are instructive precisely because they now look boring.
Insurance once ran on the discretion of individual assessors until actuarial tables converted mortality from opinion into instrument — and the underwriter's job became more valuable, not less, because it now started from a defensible number. Credit once ran on a banker's reading of character until ratings and scores gave lending a shared, auditable language of default risk — imperfect, periodically humbled, and nonetheless indispensable. In both cases the pattern was the same: a domain of professional feeling acquired a layer of governed measurement, the measurement became the starting point of judgment rather than its replacement, and within a generation, operating without the instrument came to look not traditional but negligent.
Pre-outcome intelligence is that layer for operational prediction at large — for the thousands of daily institutional questions that have never had their actuarial table because they were too varied, too fast, and too dependent on unstructured evidence. What has changed is not the desks' appetite for instruments. It is that governed, calibrated, evidence-screened prediction has become buildable.
What desks should demand
If the destination is desks that run on governed probabilities, the route runs through procurement standards — and here the practitioners hold more power than they use. Any system proposing to put probabilities in front of an accountable desk should face four questions, and the desk should walk away from any vendor who stumbles on them.
Can it show its evidence? Every probability should carry its evidence pool, and the pool should have been screened — stale, off-domain, and junk material identified and excluded before the number was formed, not explained away after.
Is its confidence constructed or asserted? A stated 74% should decompose into measured components a reviewer can inspect. If the confidence cannot be taken apart, it is a tone of voice.
Does it ever refuse? A system that has never declined to answer is a system whose every answer carries an asterisk. Abstention under polluted or insufficient evidence is not a weakness in a prediction system; it is the single fastest test of whether it has one honest bone in its architecture.
Is its record scoreable? Calibration — the match between stated confidence and realised outcomes, measured with standard instruments like Brier scores — must be checkable on the system's resolved history. Not claimed. Checkable.
These four questions are, not coincidentally, a description of what we build, and we are content to be held to them in the same order and with the same severity. The proof, for us and for anyone in this category, lives in resolved outcomes and published calibration — nowhere else.
The desks that adopt this standard first will gain something compounding: an institutional memory of governed predictions, scored against reality, accumulating while their competitors are still paying senior salaries for confident feelings. The discretionary desk had a long and honourable run. Its successor will keep the judgment, keep the accountability — and finally have instruments worthy of both.
IP Factory HQ builds POE 1, the Predictive Outcome Engine — governed intelligence systems that identify patterns, quantify probabilities, and inform decisions before outcomes occur.