Signal vs. Noise
Is That True? The Ranking This Site Uses Says Paralegals Are Safe. Three Other Sources Disagree.
Two pools of borrowers, identical on every metric that gets scored. One remodeled. One consolidated debt. The details decide.
By Scot Free
Every blueprint on this site carries a box called Will AI Take This Job?, and every one of them cites the same source: a Microsoft Research study that scored 785 occupations for how much of the work current AI can actually do.
While pulling ranks for a different piece, I ran paralegals and legal assistants through it. The dataset puts them at #668 of 785 — among the least AI-exposed occupations in America, well below the median occupation at #392.
That did not match anything I knew about the legal industry. So I went and checked, which is what this column is for.
What three other sources say
1. The federal employment projection. BLS projects paralegal and legal assistant employment at 0% change from 2025 to 2035 — 404,900 jobs falling to 403,800, a net loss of 1,100 — against 3% growth for all occupations.
2. The BLS's own stated reason, which is the part that settles it. The agency writes that employment growth for these workers may be limited by advances in technology, including artificial intelligence, and that AI is expected to make paralegals more efficient at tasks such as conducting research and preparing documents, which may reduce demand.
That is not an industry opinion or a vendor's marketing. That is the federal government naming AI, by name, as the reason an occupation of 400,000 people stops growing.
3. An independent exposure index. The International Labour Organization's 2025 generative-AI exposure work places paralegals and legal assistants in the high exposure gradient, with two of the occupation's top five O*NET tasks classified as displaceable.
And a fourth data point worth holding: over the same period BLS projects lawyers to grow 4% while paralegals stay flat. Within one industry, the exposure lands on the support tier and not on the bar-admitted tier. That is a very specific pattern, and it is consistent with the automation story rather than against it.
So: one source says least exposed. Three say exposed, one of them being the federal projection itself.
Both can be true, and this is the useful part
The reflex here is to pick a winner. That's the wrong move, because the sources are not answering the same question.
- The Microsoft study measures observed use of a general-purpose AI assistant. It looks at what people actually asked an AI assistant to do, maps those conversations to occupational work activities, and scores how much of each occupation's work showed up and got completed. It is a measurement of what was seen happening, in that tool.
- The ILO index measures what AI could do to the tasks — a task-level analysis of displaceability, independent of whether anyone is currently doing it.
- The BLS projection measures what employers are expected to hire — an industry-employment model with AI named as an input.
Now apply that to law specifically. Legal AI adoption is not happening in a general-purpose chat window. It is happening inside Westlaw, inside CoCounsel, inside Harvey, inside firm-internal document systems — tools built for the industry, sold to the industry, sitting behind a firm login.
A study measuring general-purpose assistant conversations cannot see any of that. Not because the methodology is sloppy, but because that is not where the instrument is pointed.
Which means the honest reading is: paralegal work is highly exposed to AI, and paralegals are not showing up in general-purpose AI assistant conversations. Both statements are supported. They look contradictory only if you assume the score measures automatability, which it does not claim to.
Two pools of borrowers
Here's the version of this problem I learned in finance, and it's the clearest illustration of why the details matter.
Take two pools of second mortgages. Run every metric: credit scores, loan-to-value, debt-to-income, borrower age, property type, geography. The two pools are indistinguishable on all of it.
One pool borrowed to remodel. The other borrowed to consolidate debt.
Over time, those two pools do not perform the same way. The remodel borrower is adding to an asset. The consolidation borrower is managing a liability, and the fact that they needed to consolidate is information about them that the credit score has already priced in once and will price in again.
Nothing in the scored variables captures that difference. You only find it by asking what the loan was for — a question no model asks unless someone goes and looks.
The paralegal case has the same shape. The score is correct about what it measured. The thing that predicts the outcome is sitting one layer down, in a detail the instrument was never built to capture.
What this means for the boxes on this site
I use that dataset on every blueprint, so I owe readers a straight accounting of where it holds and where it doesn't.
Where it holds. The seats on this site are overwhelmingly physical trades — HVAC-R, pipefitting, electrical, machining, welding, water operations. Those occupations run 75% to 90% physical work activity, there is no specialized AI tool quietly doing it behind a login, and a general-purpose measurement and a task-level measurement would reach the same conclusion. The blind spot described here does not bite on a seat where somebody has to physically be in the crawlspace.
Where it needs a caveat. Any seat with a substantial analytical, documentation, or software-mediated component. On this site that means the substation technician seat, where roughly a third of the work is testing, documentation, and relay setting calculations, and where the exposure rank is already the least flattering on either board.
So the box on every blueprint now carries two limits rather than one. It measures what AI chatbots can do, not robots. And it measures general-purpose assistant use, not tools built into an industry. Where those two gaps overlap, the score will read low on work that is in fact being automated.
The move
Three questions for any ranking, index, or score that lands in front of you — and they generalize well past AI:
- What did the instrument actually observe? Not what the headline claims it means. The underlying measurement, stated plainly. If you can't find that in the source, the source isn't ready to be quoted.
- Where can it not see? Every instrument has a field of view. Naming the edge of it is usually faster than evaluating the whole methodology, and it tells you which cases to distrust.
- Does anything independent agree? One source is a reading. Three sources measuring different things and pointing the same direction is a finding. Three pointing different directions means go look at what's actually happening.
The Scot Free Take
This is the site checking its own instrument, and the honest version is that the dataset isn't flawed, it's narrower than the way it gets quoted, including by me.
The mortgage pools taught me this before AI existed. You can have two sets of borrowers who match on every variable anyone bothered to score, and the thing that actually separates them — what the money was for — sits in a field nobody collected. Models are only as good as the questions they were built to answer, and almost every misuse of a model is somebody asking it a question it never claimed to answer.
So the discipline isn't skepticism about rankings. It's specificity: what did this thing measure, where can't it see, and does anything else agree. Run that on three sources and you usually learn more in twenty minutes than the headline would have told you in a year.
Get the MoneyZoo LAB Report and instant access to The Side Door Playbook. One seat, one receipt, one move, every Monday. Subscribe here.
Sources
- Microsoft Research, "Working with AI" (Tomlinson et al., 2025), github.com/microsoft/working-with-ai, CC BY 4.0 — Paralegals and Legal Assistants (SOC 23-2011) rank #668 of 785; physical work activity share approximately 31%. Rank computed by this author from the published applicability score file.
- BLS, Occupational Outlook Handbook, Paralegals and Legal Assistants (SOC 23-2011), last modified Aug. 27, 2026 — 2025 median pay $62,890; employment 404,900 (2025) to 403,800 (2035), a change of 0% and −1,100; about 40,800 annual openings; the stated outlook language on technology and artificial intelligence limiting employment growth and reducing demand for research and document preparation tasks.
- BLS, Employment Projections — lawyer employment projected to grow 4%, against little or no change for paralegals and legal assistants.
- International Labour Organization, 2025 refined generative-AI exposure index, as applied to O*NET-SOC 23-2011.00 — high exposure gradient; two of the occupation's top five O*NET tasks classified as displaceable.
- Legal-sector AI adoption context: American Bar Association 2024 technology survey (30.2% office use) and Clio legal trends research (79% broader legal-professional use), as reported in legal industry compilations.
A note on these sources: the ILO exposure gradient and the adoption percentages are reported here as cited by third parties rather than retrieved from the primary publications, and should be verified against the originals before being relied on. The BLS projection figures and the agency's own outlook language are from the Handbook directly, and they are the load-bearing evidence in this piece.
