Anthropic just asked nearly 52,000 Americans how they feel about AI. The people most afraid it will take their job are the ones who use it least, and for accountants, that pattern is pointed straight at us.
On June 12, Anthropic published the first Anthropic Public Record, a nationally representative survey of 51,993 Americans run through YouGov. The headline finding lands on our doorstep: job loss is the number-one fear about AI, named by 64% of respondents, and it ranked first in every single state. Not misinformation. Not rogue robots. Your livelihood.
Two patterns sit on top of each other, and together they describe the accounting profession almost exactly. First, worry about job loss rises with education. Americans with postgraduate degrees are nearly 10 points more worried than those with a high-school education or less. Anthropic's read: the most anxious workers are the ones whose work already overlaps with what AI is being asked to do. Research, analysis, document drafting, reconciliation. That is the job.
Second, and pulling the other way, people who use AI at work every day are markedly less worried about losing their job than people who never touch it: 54% versus 70%, a 16-point gap. The same split shows up on the second-biggest fear, cognitive dependency, where daily users again worry less, 46% versus 62%.
So the most exposed workers are the most afraid, except the ones who have actually put their hands on the tools. It is tempting to draw a straight line: use AI, fear less, problem solved. Do not. The reason matters, because this is exactly where most AI commentary gets sloppy.
This is a correlation, not a cause, and Anthropic says so plainly: hands-on experience may build skill and fluency, or it may simply reveal AI's limits, and likely it is both. Elsewhere they note the heaviest users' calm 'probably reflects differences in the outlook of early adopters.' Translation: the arrow might run backward. Maybe unworried people adopt early, and the adopting did not make them calm. The data cannot tell us which way it runs.
Here is the fair version, and the useful one. Among everything this survey measured, hands-on use is the only thing that moves alongside lower fear. It is not a guarantee. It is the one lever pointing the right direction. For a CAS or advisory practitioner, high education and work squarely in the overlap, that reframes the choice. Not be brave or be scared, but build fluency by doing, because abstract dread changes nothing and a week of real use tells you exactly where the tool helps and where it falls apart. One caveat, said out loud: Anthropic sells AI, so read a survey from an AI company concluding that AI familiarity calms people as a strong dataset with a house interest, not gospel.
Savant Labs now lets finance teams describe a task in plain language inside Claude Cowork or Microsoft Copilot and have it run as a deterministic, audit-ready workflow wired to the ERP (CPA Practice Advisor, Jun 15 2026). For advisory work, that is the bridge between 'AI drafted it' and 'I can defend it in a workpaper.'
Every meeting, contact, client org, and action item captured in Abacor is now reachable by asking Claude in plain English (CPA Practice Advisor, Jun 4 2026). If you run client advisory off scattered notes, this turns 'where did we land with that client' into one question.
Suralink's new agentic AI layer plugs Claude and Microsoft Copilot into its request-list and PBC workflows (BusinessWire, Jun 3 2026). Chasing client documents is the next chore the agents come for. Worth a look before your next engagement kickoff.
In the same Anthropic survey, AI firms scored dead last on trust, below the federal government at 20% and far below independent experts at 43%. Your clients' default toward the vendors is skepticism. Sell the advisory outcome, never the logo on the tool.
Read that gap as a map, not a motivational poster. It does not prove that using AI cures the fear, because early adopters may have started out calmer. But it is the single sharpest signal in a 52,000-person survey that the people closest to the tools carry the least dread about them. For a profession sitting in the overlap, that beats another think-piece predicting our extinction.
What it is: Anthropic's AI assistant, and the most direct way for an advisory practitioner to get the hands-on reps this survey rewards. Beyond chat, Projects let you load a client's chart of accounts, prior workpapers, and your firm's templates so every answer is grounded in that engagement, not the open internet.
What it does well: Reasoning over messy, unstructured client material. It turns a founder's rambling email into a clean list of advisory questions, sketches variance commentary, and stress-tests a cash-flow assumption. Paired with a QuickBooks connector or a platform like Abacor or Suralink, it works against live data instead of a stale export.
What it doesn't do well: Arithmetic at scale and anything you cannot verify. It will state a confident number that is wrong, so it belongs on the draft-and-review side of your process, never the final-sign side. Tie out every figure the way you would a new associate's.
Pricing: Claude Pro runs about $17 per month billed annually, or $20 monthly. Team is about $20 per seat per month annually, $25 monthly, minimum five seats, and keeps your conversations out of model training by default, which is the version a firm handling client data should standardize on. Confirm current tiers at claude.com/pricing before you budget.
This week's Workflow Lab is the cheapest way to move yourself from the 70% toward the 54%: build one Claude Project for a task you already repeat every month. I built mine for monthly advisory commentary across a handful of CAS clients.
Setup runs about twenty minutes. Create a Project, then load its knowledge with the raw materials you reach for every month: the client's trailing P&L, your variance-narrative template, the firm's tone-and-format guide, and two or three past commentaries you were proud of. That context is what separates a generic answer from one that sounds like your practice.
Then hand it the month's numbers and ask for a first-draft commentary in your format: flag the three biggest variances, propose a plausible driver for each, and list the follow-up questions to put to the client. You are not publishing what it writes. You are skipping the blank page and starting from a draft you edit, which is where the hours actually go.
The point is not the output. It is that after a month of this you will know firsthand, not from a survey, exactly where the tool earns its keep and where it cannot be trusted near a client deliverable. That knowledge is the thing the 70% do not have, and it is the only thing in the data that tracks with worrying less.
Pick one recurring advisory task you would never hand a brand-new associate on day one, a tricky cash-flow projection or a messy client-cleanup diagnosis, and run it alongside Claude this week, not instead of you. Keep score: where did it save real time, and where did it confidently get it wrong? You will finish the week holding the one thing the 70% do not have, a firsthand map of where the tool helps and where it cannot. That map, not a headline, is what makes the fear manageable.
P.S. Advisory and CAS folks: reply with the one client task you wish AI could take off your plate but do not yet trust it to. I'm collecting them for a future Workflow Lab on what's actually safe to delegate.
