I had an idea while designing an analytical prospecting system that I had to kill. Killing it produced a better one.
When you build one of these pipelines you run into a class of work that machines are not good enough at and that is too expensive to have a highly paid person in a high-cost country do. There is a lot of that work. Verifying that a name in a filing and a name in a database are the same human being. Reading an ambiguous document and deciding which of two readings is right. Checking that a machine-extracted fact is actually in the source. Judgment calls at the edge of a rule.
The obvious move is to send that work somewhere cheaper. The interesting move is to notice that the reason it was never worth sending anywhere is that coordinating it cost more than the work did, and that this is the constraint AI actually removes.
First, the version that does not work
My initial sketch of the prospecting system had inexpensive offshore staff reviewing prospects’ social media content, on the reasoning that there are only a few hundred thousand people in the target population and the labor cost would be trivial against the value of a single new client.
That reasoning is sound on cost and wrong on everything else, and it is worth being specific about why, because the failure mode generalizes.
Cheaper labor does not change what is permitted. If extracting content from a platform violates that platform’s terms, it violates them whether the person doing it sits in New Jersey or anywhere else. Obligations follow the data and the firm, not the worker. Routing an activity through a lower-cost jurisdiction to make it feel more acceptable is a way of laundering a rule you did not want to follow, and it does not survive contact with a regulator or a plaintiff.
Automating it does not change it either. A related idea I considered was skipping the humans and using a headless browser to screenshot pages, then a vision model to read the screenshots. It is technically clean. Legally it is the same act. An automated session that captures content is collection, and having a model read the image rather than parsing the DOM changes who wrote it down, not what happened. If the underlying access was not permitted, the screenshot is not a workaround, it is evidence.
And the data protection obligations attach the moment you move the data. Sending personal information about individuals to contractors in other jurisdictions triggers vendor oversight and incident response duties, and under the amended Regulation S-P those duties reach the service providers touching the data. Cross-border transfer terms, data processing agreements, and access controls are not paperwork you add later. They are the design.
So the honest conclusion is that the offshore idea was right about the economics and wrong about the scope. The work that should go there is not bulk collection of data you were not entitled to collect. It is judgment on data you were entitled to have.
What actually goes there
Once you draw the line that way, a real category emerges, and it is bigger than it first looks:
- Entity resolution. Is the John Sullivan in this SEC filing the same John Sullivan in this county property record? A model gets this right most of the time and confidently wrong the rest of the time, and confidently wrong is the expensive case.
- Citation verification at the margin. The automated check catches the clear failures. The remaining cases are ones where the claim is arguably supported, and arguably is a human word.
- Ambiguity adjudication. The three model runs disagreed. Someone has to decide.
- Edge-of-rule judgment. Does this outreach pass the “how did you know that?” test. Reasonable people differ, which is the definition of a case a model should not be closing on its own.
- Quality sampling. Pulling a random sample of records the system was confident about, to find out whether the confidence was justified.
Every one of those operates on data that already passed the compliance gate. None of them involves going to get anything.
The coordination is the thing AI changes
Distributed human review has always been possible. What killed it was overhead. Someone has to decompose the work, write instructions for each piece, route it to someone qualified, check what comes back, handle disagreements, and keep quality steady across a group of people who cannot ask a follow-up question easily. That coordination cost is why this kind of work stayed in-house and mostly stayed undone.
That is the part that is now cheap.
System
Identifies a case it should not close alone: the models disagreed, confidence is below the threshold, or the compliance gate flagged a rule-edge case.
Minimize
Build the task with the least data that makes it answerable. Strip identity where the task does not need it. One record, one question.
Instruct
The model writes the task in plain language, with the decision rule, two worked examples, and a path for when the reviewer cannot tell.
Route
Send it to a reviewer qualified for this task, in a jurisdiction cleared for this data, under a data processing agreement and an NDA.
Verify
A second reviewer, or a model, checks a sample. Systematic disagreement goes back to the instruction, not the person.
Log
Who decided, on what evidence, under which version of which rule, on what date.
The pattern worth extracting: the AI is not doing the judgment and it is not managing people. It is doing the packaging. It decomposes the problem, writes a task that can be answered without the surrounding context, decides who should see it, and reconciles what comes back. The human contributes the one thing that was actually scarce, which is a judgment call, and contributes it without needing to hold the whole system in their head.
That inverts the usual framing. The common version of this story is humans supervising AI. What is actually more useful, and more buildable, is AI organizing work so that human judgment can be applied in small well-formed units, at the exact points where it is needed, by people who could never have been onboarded to the full problem.
The gates, stated plainly
If you build this, these are not optional and they are cheaper to build in than to retrofit.
| Gate | The rule |
|---|---|
| Data minimization | The reviewer sees what the question requires and nothing else. “Does this document support this claim” needs the document and the claim, not a name. |
| Jurisdiction routing | Which data classes may be seen from which countries, checked before the task is created. |
| Contractual coverage | A data processing agreement, an NDA, retention and deletion, a right to audit, and breach notice, before any data moves. |
| Access controls | Individual accounts, no shared credentials, no local storage, no export, session logging. |
| Vendor oversight | Periodic review, not a signature at the start. Amended Reg S-P makes this explicit. |
| Audit trail | Who decided, on what, under which rule version, and when. An undocumented human decision is worth less than an automated one. |
| Pay and disclosure | Would you describe the arrangement accurately to the people whose data is being reviewed? If that description is uncomfortable, the arrangement is the problem. |
When it is worth it
Not always. This only pays when the work has high judgment density and low context requirements, meaning each unit needs real thought but not much background. Entity resolution fits. Anything requiring deep familiarity with a specific client relationship does not, and trying to force it produces worse output than the model you were trying to improve on.
There is also a floor. Below a certain volume the setup cost dominates and you should just do the work yourself.
What I would take from all of this, though, is the general shape rather than the specific application. The scarce input in most analytical systems is not compute and it is not model capability. It is human judgment applied at exactly the right point. Most systems either ask for far too much of it, which means nobody sustains the process, or none at all, which means the confident errors go straight into production. Building the machinery that finds the handful of cases that genuinely need a person, and puts a well-formed question in front of the right person, is where a lot of the practical value in this technology actually sits.
That is a design problem, not a model problem.