VOLO · flyvolo.ai · SAMPLE · generated 2026-10-18
Cohort brief: what AI is changing in the work these degrees lead to
For each degree, the occupations it leads to, each broken into the tasks people actually do, with today's direction for every task and the verified records behind it.
What this brief does not do. It gives no score for any person, programme, role or occupation, no probability of losing a job, and no year by which anything happens. It does not say who to make redundant. Directions are judged task by task and are never added up. Where the evidence cannot show a direction, it says so. Every record links to its source; every judgement links to the public page, which stays free.
Degrees in this brief
How to read a task's basis: evidenced — at least one verified, sourced record supports the direction; inferred — our reading of the task, no record yet; uncertain — the direction cannot be judged.
Accounting
Trains you to see how money actually moves through an organisation — and to be the person who can say whether a number is defensible.
Where this degree leads, and what it reuses
| Path | What it reuses | The gap |
|---|---|---|
| Accounting, reaching judgement work faster flyvolo.ai/en/careers/accountant | All of it — the direct path. | Not knowledge but exposure. The traditional way in was to do entry and reconciliation until you understood the business; that is the automated part, so you have to seek advisory and exception work early rather than waiting to be given it. |
| Financial analysis / FP&A flyvolo.ai/en/careers/financial-analyst | The money-flow understanding, pointed forward instead of backward — the work is framing questions about what should happen rather than recording what did. | Modelling beyond spreadsheet familiarity, and the confidence to present a number you might be wrong about. |
| Controls, audit or compliance in another sector flyvolo.ai/en/careers/auditor | Control thinking, which transfers across industries far better than industry knowledge. Regulated sectors reward it and are slower to automate the accountable layer. | The sector itself, plus certification requirements that vary sharply by market. |
| Owning the automation rather than using it flyvolo.ai/en/careers/accountant | Control thinking plus enough domain knowledge to know what a correct output looks like. Someone has to choose the tools, define the checks, and be accountable when the output is wrong. | Enough technical comfort to evaluate a tool rather than trust a demo. Not a steep curve, but a real one. |
The occupations on those paths, task by task
Accountant / Bookkeeper
Impact index 68 · confidence medium · assessed 2026-09-09 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Entering and coding transactions | Automating | evidenced | Does not mean the headcount disappears. In small practices the same person does entry and advisory; automating the entry half changes the job's shape before it changes the job count. |
| Reconciliation | Automating | evidenced | The unmatched tail does not shrink in proportion to the matched volume. A practice that automates nearly all of its matching still needs someone who can chase what is left — and that person needs to have seen the rest to recognise what is wrong. |
| Statutory filing and tax compliance | Being augmented | evidenced | The liability argument protects the sign-off, not the preparation hours behind it. Expect fewer hours per filing, not fewer filings needing a named person. |
| Explaining the numbers to decision-makers | Still human-led | evidenced | One study of small firms, measuring associations rather than effects. Being hard to automate is not the same as being in demand: advisory work is concentrated in the senior half of the profession, and a junior whose entry work disappeared does not automatically arrive here. |
| Judging the ambiguous case | Still human-led | evidenced | One study of small firms on one platform, and its findings are associations. The volume of ambiguous cases is a fraction of total transactions; this task protects the necessity of the role, not the number of hours it takes. |
| Supervising the automation itself | New task | evidenced | New work is not the same as new headcount, and this task is usually absorbed by people already there rather than hired for. It also requires the judgement built by doing the work that is disappearing. |
Most recent verified records
- Employment of 22-25-year-olds in the two most AI-exposed occupation quintiles fell about 11% from November 2022 to June 2026, while the least-exposed quintiles grew about 10% (2026-08-12 · Labour impact · Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (August 2026)) record
- The IRS regulator of tax practitioners, who include CPAs, says AI-created returns and documents must be thoroughly reviewed before they reach clients or the IRS (2026-06-24 · Constraint · Internal Revenue Service, Office of Professional Responsibility — Issue Number 2026-19, Introductory Guidelines for Responsible AI Use in Federal Tax Practice (June 24, 2026)) record
- A field study of SME accounting finds accountants step in where the AI is unsure, and time moves from data entry to client communication (2025-03-31 · Labour impact · Jung Ho Choi and Chloe L. Xie — 'Human + AI in Accounting: Early Evidence from the Field', Journal of Accounting Research (first published 16 April 2026)) record
Something a student can check this week
Log your week in two buckets — production vs judgement — in hours. If judgement is under 20%, you have your answer and a number to take to your manager.
Financial analyst
Impact index 60 · confidence medium · assessed 2026-09-10 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Building and updating models | Being augmented | evidenced | Structure design is the interesting hour in a week of updating. Automating the update removes the hours that justified the headcount, while leaving the hour that justifies the title. |
| Gathering data and summarising filings | Automating | evidenced | The summaries are only as good as the reader's ability to notice what was left out. Juniors who never read a full filing lose the instinct for what a summary hides. |
| Deciding what to assume | Still human-led | evidenced | Assumption-setting is the senior half of the job by definition. Analysts are not hired into it; they arrive after years of the modelling work that is being compressed. |
| Explaining to the decision-maker | Being augmented | evidenced | The meeting in which capital is committed happens a few times a quarter. It protects the seat at the table, not the work that filled the rest of the calendar. |
| Orchestrating the analytical tooling | New task | evidenced | Explicitly peripheral by weight, and typically absorbed by one person on a team as an extra. It is not a job posting. |
Most recent verified records
- OpenAI launched a financial-services product for research, modelling and client materials, shaped by design partnerships with Morgan Stanley and Evercore (2026-09-10 · Capability · OpenAI) record
- Goldman Sachs made its in-house GS AI Assistant available to all employees after use by thousands of staff, for summarising complex documents, drafting initial content and data analysis (2025-06-23 · Deployment · Fox Business (reporting a Reuters-obtained internal memo)) record
- On the Finance Agent Benchmark — 537 expert-authored questions over recent SEC filings — the best model, OpenAI o3, reached 46.8% accuracy at $3.79 per query (2025-05-20 · Constraint · arXiv 2508.00828 (Bigeard, Nashold, Krishnan, Wu)) record
Something a student can check this week
Write down three assumptions in your current model that you would defend against your manager, with the reason. If you cannot find three, that is the gap to close.
Auditor
Impact index 56 · confidence low · assessed 2026-09-12 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Sampling and testing | Automating | evidenced | Says nothing about how audit hours are actually billed, and in many firms the hours freed here were absorbed into more testing rather than into fewer people. It also says nothing about smaller audits, where the tooling is often not bought at all. |
| Deciding where to look | Still human-led | evidenced | A judgement about the nature of the work, not a measurement. It also does not claim auditors are good at this: missing the risk that mattered is the defining failure of this profession and there is a long public record of it. |
| Asking the question that gets a real answer | Still human-led | evidenced | Nothing here measures how much of a modern audit is inquiry versus document testing, and in many engagements the inquiry is a formality completed by email. Where it is a formality, this task is not human-led — it is simply not being done. |
| The file and the opinion | Being augmented | evidenced | Nothing public measures documentation time saved in audit, and firm claims about it come from a party selling the transformation. What would settle it: a regulator's inspection findings on files prepared with these tools. |
| Signing it | Still human-led | evidenced | It protects the signature, not the headcount behind it. A firm can sign the same number of opinions with fewer people, and the signature says nothing about how many juniors were needed to get there — which is the number most readers of this page actually care about. |
| Owning what the machine drafted | New task | evidenced | That record is one consulting engagement at one firm, not a statutory audit, and it establishes that the failure happened rather than how often it does. New work appearing is also not new headcount: in a billable-hours business an unbilled checking duty is absorbed, which is exactly the condition under which it gets skipped. |
Most recent verified records
- Deloitte agreed to repay the last instalment of a US$290,000 Australian government report after fabricated references and a misattributed court quote were found; it acknowledged using Azure OpenAI (2025-10-07 · Constraint · Fortune) record
- The revised international fraud standard requires auditors to question management and others directly about fraud, and says automated tools do not replace professional skepticism (2025-07-08 · Constraint · International Auditing and Assurance Standards Board — International Standard on Auditing 240 (Revised), The Auditor's Responsibilities Relating to Fraud in an Audit of Financial Statements (final pronouncement, July 2025; effective for periods beginning on or after 15 December 2026)) record
- PCAOB economists reported registered audit firm staffing increasing steadily from 2015 to 2024, and rising per dollar of client revenue, with non-CPA accountants growing about twice as fast as CPAs (2025-07-01 · Labour impact · PCAOB Office of Economic and Risk Analysis — Data Points: Registered Firm Staffing Trends) record
Something a student can check this week
Take last year's file on one client and find the risk you decided not to pursue. Write down why, and whether you would decide the same today.
Business administration
Trains you to coordinate work you do not personally do — to read the numbers, split a job into steps, decide with incomplete information, and get people who do not report to you to act.
Where this degree leads, and what it reuses
| Path | What it reuses | The gap |
|---|---|---|
| Sales and account management flyvolo.ai/en/careers/sales-account-manager | Reading a customer's numbers well enough to argue in their terms, and coordinating the people inside your own company who have to deliver what you promised. | Graduates usually arrive without having been told no by a real buyer, and without a feel for how long a decision takes inside a company that is not their own. |
| Running a team on the floor flyvolo.ai/en/careers/first-line-manager | Splitting work into steps, and being the person who answers when a decision made by a system lands on someone. | Graduates rarely have experience of being accountable for someone else's output, which is the whole job, and almost never of delivering bad news to a person whose work they depend on. |
| Business operations flyvolo.ai/en/careers/operations-coordinator | The process-splitting, directly. Operations is where a written-down sequence becomes someone's daily job and where the exceptions get handled. | Graduates tend to be fluent in the ideal path and unprepared for the exception, which is where most operations time actually goes — the order that does not fit, the customer who is a special case. |
| Owning the business system flyvolo.ai/en/careers/business-systems-owner | Splitting work into steps, plus the part of the training that is about accountability: deciding what the system may decide on its own, and being the person who answers for that setting. | Graduates can describe a process but usually cannot yet read what a system is actually doing — the configuration, the audit log, the rule that fires at 2am — and that reading is what the role turns on. |
| Partnerships and channel flyvolo.ai/en/careers/partnerships-manager | Moving people sideways, applied outside the company: getting another organisation to put your product in front of its own customers when you have no authority over them at all. | Graduates underestimate how much of this is contract detail and follow-through months after the announcement, and overestimate how much is the meeting where the deal is agreed. |
| Buying, and who you buy from flyvolo.ai/en/careers/procurement-specialist | The commercial half of the degree used on the other side of the table: total cost rather than price, and a supplier relationship that has to survive the quarter in which you pushed hard on it. | Graduates can build a comparison and cannot yet judge the thing the comparison leaves out: whether this supplier will still answer the phone in a bad month. Nothing in a spreadsheet carries that, and it is what the job is paid for. |
The occupations on those paths, task by task
Sales / account manager
Impact index 58 · confidence medium · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Working out who is worth your time | Automating | evidenced | Faster qualification is not more deals. It changes where the hours go, not whether the customer buys — and a rep who now qualifies three times as many leads is being measured on a bigger number with the same closing skill behind it. |
| Putting the price and the proposal together | Automating | evidenced | What automates is assembling the quote, not deciding the number. Where the margin is set by judgement about this customer and this competitor, the document gets faster and the decision does not move — and the decision is where the money is. |
| Chasing, on time, every time | Being augmented | evidenced | Remembering is not persuading. A perfect follow-up list raises the floor for a disorganised rep and does almost nothing for a good one — so it compresses the gap between reps rather than replacing any of them, and that changes who gets hired, not how many. |
| The conversation where it is decided | Still human-led | evidenced | Being the hardest part to automate does not make it the biggest part of the day. In most SME sales roles this is a few hours a week sitting on top of many hours of assembly and chasing — so a role can shrink substantially while its irreplaceable core stays exactly where it is. |
| Telling the boss what will land | Being augmented | inferred | A better forecast changes what management does, not what the customer does. It can also make a rep's month look worse without their work being worse, which is a management problem dressed up as a data problem. |
| Running the machine that now sells with you | New task | inferred | New duties are not new headcount. This lands on the reps who are already there, usually without a title change and without coming off their number — which is how a role gets heavier while looking unchanged on the org chart. |
Most recent verified records
- Employment of 22-25-year-olds in the two most AI-exposed occupation quintiles fell about 11% from November 2022 to June 2026, while the least-exposed quintiles grew about 10% (2026-08-12 · Labour impact · Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (August 2026)) record
- Salesforce's CEO said its roughly 15,000 engineers had been flat for two years because of AI, while headcount grew mostly in sales (2026-05-27 · Labour impact · Salesforce, Inc. — Q1 FY27 Earnings Conference Call transcript (投资者关系自有发布, 2026-05-27)) record
- Singapore's financial regulator and the sector's skills bodies forecast that Gen AI will augment most financial-sector roles over five years rather than displace them (2025-04-01 · Forecast · Generative AI Jobs Transformation Map for Singapore's financial sector (MAS, IBF and Workforce Singapore)) record
Something a student can check this week
Take the last five deals you quoted and write down, for each, what you would have priced it at if it were yours to decide and why. Show that page to your manager and ask where you are wrong. You will find out in one conversation whether the gap is judgement or permission.
First-line manager / team supervisor
Impact index 41 · confidence low · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Deciding who does what today | Automating | evidenced | A roster the machine produced still has to be defended to the person who got the bad shift, and that conversation stays with the supervisor. What automates is the hour of arranging, not the accountability for the arrangement — so this role can lose its most visible daily task and keep the one that makes it a job. |
| Seeing it go wrong before the report does | Being augmented | evidenced | More alerts is not more attention. Where detection got cheap, the usual result is a supervisor triaging a longer list of flagged items in the same hours — the work moved from finding to filtering, which reads as no change at all on a job description. |
| Writing the review and giving the feedback | Automating | evidenced | A drafted review is not a delivered one, and what costs a manager something is saying it out loud to someone who disagrees. It also creates a new failure mode rather than removing one: an assessment assembled from whatever was easy to log quietly promotes whoever is most legible to the system. |
| The conversation that keeps someone | Still human-led | inferred | Being the least automatable task does not protect the headcount around it. A company can halve its supervisors with this task fully intact — each remaining one simply does it with twice as many people. That is how an irreplaceable part gets thinner without ever being replaced. |
| Telling the floor above what happened | Automating | evidenced | The countersigning rule covers one hazardous sector. What automates is the summary, not the selection: deciding that a near-miss is worth the boss's attention this week is a judgement with career consequences attached, and a system that reports everything has effectively reported nothing. |
| Defending a decision the system made | New task | evidenced | New duties are not new authority. A supervisor asked to explain a system they are not allowed to change carries the blame for it without the power to fix it — a different job from the one they were promoted into, heavier, and invisible on the org chart. |
Most recent verified records
- Colorado's SB26-189 gives a person the right to meaningful human review after an automated consequential decision goes against them, with a plain-language explanation due within 30 days (2026-05-14 · Constraint · Colorado General Assembly, SB26-189 Automated Decision-Making Technology — session law chapter 131, signed 2026-05-14) record
- California's SB 7 would have barred relying solely on an automated system to discipline or dismiss a worker; it passed the legislature, was vetoed, and the veto was sustained on 2 March 2026 (2026-03-02 · Rule not enacted · California State Legislature — SB 7 (2025–2026), bill status and Governor's veto message (leginfo.legislature.ca.gov)) record
- 90% of US managers and 79% in Europe say their firms use software to instruct, monitor or evaluate workers, an OECD survey of more than 6,000 managers found (2025-12-19 · Deployment · OECD — How widespread is algorithmic management in workplaces? (policy brief, 19 December 2025), drawing on the OECD Algorithmic Management Survey of mid-level managers in France, Germany, Italy, Japan, Spain and the United States) record
Something a student can check this week
For one week, write down every time you changed what a system proposed — the roster, a flag, a rating — and what happened afterwards. Five entries tell you whether your overrides are judgement or habit, and that page is the case for being given more of them.
Operations coordinator
Impact index 62 · confidence low · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Turning a signed order into a real thing happening | Automating | evidenced | Orchestration automates the sequence, not the commitment. Someone still has to be willing to promise a customer a date and wear it when the supplier misses — and that promise is worth more than the scheduling. The job can lose most of its clicks and keep all of its liability. |
| Getting the confirmation back out of someone upstream | Automating | evidenced | It automates the chasing, not the relationship. The reason a supplier squeezes you in on a full day is that they know you, and that favour is not available to an automated reminder. Where supply is tight, the coordinator with the relationship still wins the slot — which protects some coordinators, not the headcount. |
| After the plan breaks | Still human-led | evidenced | Exceptions are where the value is, and they are also a small share of the hours. A role that is mostly routine coordination and partly exceptions can lose the routine and remain a job — for far fewer people, each of them on call more of the time. |
| The paperwork that has to be right | Automating | evidenced | Automated checking does not move who is liable. When a wrong number gets through, the penalty lands on the company and the explaining lands on the coordinator — so the checking gets cheaper while the exposure stays exactly where it was, and a check that is never wrong is also a check nobody reads any more. |
| Making the system say what is actually true | Being augmented | evidenced | A cleaner record makes the automation above it trustworthy, which is the point — and it also makes the coordinator's contribution invisible, because the work shows up as an absence of problems. That is a bad position to be in when headcount is reviewed. |
| Watching the automatic flow for quiet mistakes | New task | evidenced | Sampling is not coverage, and nobody has decided what rate is enough. This work is usually unfunded and unmeasured, so it is the first thing dropped in a busy week — which is exactly when the automation is running hardest. One federal agency required weekly reviews of its bots' logs, never gave the reviewers access, and then deleted the requirement. |
Most recent verified records
- C.H. Robinson's Q2 2026 10-Q shows average headcount down from 12,858 to 11,471 while revenue rose 19%, naming automation and AI to cut manual processes as part of its first restructuring initiative (2026-06-30 · Labour impact · C.H. Robinson Worldwide, Inc. — Form 10-Q for the quarter ended 2026-06-30 (SEC EDGAR)) record
- C.H. Robinson says AI agents now handle 95% of checks on missed freight pickups: one calls the carrier and another decides the next step, where teams once spent over half the day chasing (2026-01-26 · Deployment · C.H. Robinson — press release, "C.H. Robinson Launches AI Agents to Combat Industrywide Problem of Missed LTL Pickups" (26 January 2026)) record
- Eurostat: 6.08% of EU enterprises using AI applied it to logistics in 2025, the lowest of the seven purposes the dataset publishes (2025-12-31 · Deployment · Eurostat — Artificial intelligence by size class of enterprise (isoc_eb_ai), EU27, reference year 2025) record
Something a student can check this week
For one week, log every exception: what broke, roughly what it would have cost, how long you spent, what you decided. Five lines is enough. That page is the only version of your job that a spreadsheet-based headcount decision can see.
Business systems owner
Impact index 52 · confidence low · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Shaping the process inside the system | Being augmented | evidenced | Faster configuration means more configuration, not less work. Every rule added is a rule that will later be wrong and have to be found — so the role's hours move from building to untangling, and untangling is invisible on a roadmap. |
| Keeping the records worth trusting | Automating | evidenced | Cleaner data makes everything downstream more trustworthy, which is the point — and it also removes the only visible artefact of this job. Nobody notices records that were never wrong, so the work reads as an absence, which is a poor position when the budget is set. |
| Deciding what the system may decide alone | New task | evidenced | Setting the boundary is not the same as being allowed to hold it. This role usually has the knowledge and not the authority: when sales wants the automation to send more, the person who can say no is further up, and being overruled repeatedly is how the boundary erodes without any decision ever being recorded. |
| Making the numbers mean something | Automating | evidenced | Self-service reporting does not remove the person who knows the number is wrong. It multiplies the number of people confidently quoting a figure whose caveat they never saw — so the work shifts from producing reports to correcting them in meetings, which is slower and harder to defend as a headcount. |
| Getting the team to use it properly | Still human-led | evidenced | This is the first task cut when the role is under pressure, because it produces nothing shippable. Cutting it does not show up as a problem for two quarters, and then shows up as data nobody trusts — by which time the cause is no longer attributable. |
Most recent verified records
- China's cyberspace regulator reported 868 filed generative AI services and 530 registered applications as of 30 April 2026, and requires every live application to display which filed service it uses (2026-04-30 · Constraint · 国家互联网信息办公室——《关于发布生成式人工智能服务已备案信息的公告(2026 年 3 月至 4 月)》) record
- Census BTOS supplement: 16% of AI-using US firms replaced existing software or equipment with AI-integrated solutions (2026-04-01 · Deployment · U.S. Census Bureau, Center for Economic Studies (CES-WP-26-25)) record
- The UK Ministry of Justice links and deduplicates people's records across prison, probation and court case systems with a probabilistic model in production, having judged rules unsuitable (2025-10-06 · Deployment · UK Ministry of Justice — Algorithmic Transparency Recording Standard record, "MoJ : Splink Master Record" (GOV.UK, published 6 October 2025; phase: production)) record
Something a student can check this week
Write down the three numbers your system is supposed to move and where each one currently stands. If you cannot fill in the second column without asking someone, you are administering an instance rather than owning an outcome — and that gap is the whole conversation.
Partnerships / channel manager
Impact index 44 · confidence low · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Working out which partner is real | Automating | evidenced | A longer shortlist is not more partnerships. The bottleneck in this job has never been finding candidates, it is that most signed partners never sell anything — so faster sourcing mostly produces a bigger pile of dormant agreements, and someone still has to explain the pile. |
| Agreeing what each side actually owes | Still human-led | inferred | Being the hardest part to automate does not make it most of the week. In most partnership roles this is a handful of conversations a year sitting on top of many hours of research, chasing and reporting — so the role can shrink a great deal while the irreplaceable part stays exactly where it is. |
| Getting the two systems actually connected | Being augmented | evidenced | Cheaper integration is why the number of live connections grows and why most of them carry no volume. A connected partner and a selling partner get reported as the same milestone, and only one of them is revenue. |
| Keeping them selling after the announcement | Still human-led | inferred | This task surviving does not protect the headcount around it. One manager can nominally cover thirty partners; the work does not disappear, it thins out per partner until the quiet ones are never called — and nobody logs the calls that were not made. |
| Counting what the channel actually produced | Automating | evidenced | Automating the count does not settle the argument, it sharpens it. A number that is harder to flatter makes more partnerships look dormant, and the person who has to say so out loud is the same person whose role is justified by the channel being worth having. |
| Governing what partners say in your name | New task | evidenced | The obligation moved and the authority did not. Your contract usually lets you object after the fact, not approve in advance — so this is a monitoring duty over someone else's output with no ability to stop it at source, and nobody has decided how much of it to sample. |
Most recent verified records
- Pagaya tells shareholders it has cut the time to onboard lending partners by 2-3x, and that every new partner now joins through pre-built product integrations (2026-06-30 · Deployment · Pagaya Technologies — Q2 2026 Shareholder Letter, filed with the SEC as Exhibit 99.2 to Form 8-K (July 30, 2026)) record
- AWS now measures partnership success across its partner programs with automated Partner Revenue Measurement, replacing manual reporting, reconciliation and tracking (2026-01-30 · Deployment · AWS Partner Network (APN) Blog — "Unlock revenue insights in the new Attributed Revenue Dashboard" (21 April 2026), with the launch notice "New Partner Revenue Measurement gives visibility into AWS service consumption" (30 January 2026)) record
- AWS launched an AI agent in Partner Central that checks partners' AI specialisation applications and gives immediate feedback, saying it cuts processing time by up to 70% (2025-11-30 · Deployment · Amazon Web Services — What's New, "AWS expands AI Competency with new Agentic AI categories" (posted Nov 30, 2025); with the APN blog "Powering Next-Level Partner Success: Innovations for Growth and Scale in 2026" (01 DEC 2025)) record
Something a student can check this week
Rank your partners by revenue they actually produced last quarter, not by how promising they are. Count how many produced nothing. That single count is the conversation automated attribution is bringing to this role, and it is better to open it than to answer it.
Procurement / supply chain specialist
Impact index 59 · confidence low · assessed 2026-09-13 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Finding and comparing suppliers | Automating | evidenced | Finding a supplier is not qualifying one. Nothing here says the candidates produced are real, solvent, or able to deliver at the volume asked, and in several industries the list that matters is not public at all. |
| The order and the paperwork | Automating | evidenced | The matching rests on how the work is done, not on a measurement of how much of it is automated, and the answer differs between a company with an ERP and one running on spreadsheets. The signature rule covers the US federal government only. Automating a match also does not resolve the mismatch, which is where the time goes. |
| Negotiating with someone who will remember | Still human-led | evidenced | The deployment covers tail-end suppliers on standard terms and the trust finding comes from an experiment with simulated suppliers; neither shows how long-term strategic relationships respond, and neither is an argument that buyers negotiate well. It also does not cover commodity categories bought at auction, where the relationship genuinely does not exist. |
| Knowing a supplier is in trouble | Being augmented | evidenced | Nothing here says whether companies act on what monitoring shows them. A known-risky supplier that nobody replaced is the ordinary outcome, and that is a decision problem rather than a detection one. |
| Deciding who goes short | Still human-led | inferred | A judgement about where the decision sits, not a measurement of how often shortages happen — which differs enormously by industry and by year. It also does not claim the decision is made well, or made by procurement rather than above it. |
| Answering for who is in the chain | New task | evidenced | Requirements differ sharply by market and by company size, and what would settle how much the duty weighs is an enforcement action naming a buyer, which has not surfaced. New work appearing is also not new headcount — in most companies this lands on the buyer who already has the supplier relationship. |
Most recent verified records
- Japan's IT promotion agency surveyed 1,799 companies and reported generative AI in use in 31.3% of procurement departments against 71.9% of IT departments (2026-07-30 · Worker adoption · IPA (Information-technology Promotion Agency, Japan) — DX動向2026 (DX Trends 2026), full report PDF) record
- Anthropic's Economic Index for May 2026 records supplier research and purchase-order handling as delegation-shaped, while negotiation and price analysis sit below the median of all occupational tasks (2026-05-01 · Worker adoption · Anthropic Economic Index — Cadences (June 2026 release), dataset for the May 2026 window) record
- Volkswagen's procurement monitors suppliers' finances continuously and runs software-based media monitoring over more than 40,000 direct suppliers, 96% of its purchasing volume (2025-12-31 · Deployment · Volkswagen AG — Annual Report 2025 (published on the occasion of the Annual Media Conference on March 10, 2026), risk report and sustainability report) record
Something a student can check this week
List your categories by how many suppliers could actually deliver. The short lists are where your job is.
Computer science
Trains you to model a problem precisely enough that a machine can execute it — and to reason about what happens when it does that at scale, under failure, and with someone attacking it.
Where this degree leads, and what it reuses
| Path | What it reuses | The gap |
|---|---|---|
| Software engineering, entering deliberately flyvolo.ai/en/careers/junior-software-developer | All of it — this is the direct path. | Not knowledge but evidence of judgement. The traditional first job taught judgement by paying you to write routine code for two years; that arrangement is weakening, so you need to demonstrate review-quality thinking before anyone would normally ask for it. |
| Domains where being wrong is expensive flyvolo.ai/en/careers/backend-developer | Failure reasoning and fundamentals, in the places that keep a human firmly in the loop: payments, security, data integrity, infrastructure. | Depth in one specific area, which only time produces. Most graduates have breadth and no depth, which is exactly backwards for this path. |
| Between engineering and the problem | Problem modelling plus the ability to read a system, applied where the value is understanding both the machine and what a person needs from it: solutions engineering, developer experience, technical product. | Communication treated as a first-class skill rather than a soft one — writing documentation someone can actually follow, and running a conversation with a person who does not share your vocabulary. |
| Engineering thinking in a non-software field | The whole toolkit, applied where almost nobody can both understand the domain and build. That combination is scarce and does not compete head-on with generation. | The domain itself, restarting from near zero — and the patience for a first year that feels like going backwards. |
| Testing and quality engineering flyvolo.ai/en/careers/software-tester | Failure reasoning, turned into a job: working out how a system breaks before a user does, and building the harness that keeps proving it. | Graduates usually arrive able to write tests for code they wrote themselves, which is the easy case. The job is testing code someone else wrote, under a deadline that belongs to them. |
| Putting AI into production inside a company flyvolo.ai/en/careers/ai-implementation-lead | Reading a system plus problem modelling, applied to a company that has bought the tools and now needs someone to decide what they may do unattended. | Graduates underestimate that most of this is organisational — getting two departments to agree, and standing there when the thing gets something wrong — rather than technical. |
| The layer people actually touch flyvolo.ai/en/careers/frontend-developer | Problem modelling, in the one place where the undefined cases are visible to a stranger: the states nobody drew — empty, half-loaded, offline, the error that arrives while the user is still typing. | Accessibility, which a CS degree almost never teaches and which in many markets is a legal requirement rather than a nicety. Note what that means for this path: the part of the job the degree skipped is the part with a law behind it. |
| The pipes, and who answers for what comes out flyvolo.ai/en/careers/data-engineer | Failure reasoning, aimed at the failure mode this degree teaches you to expect and most people never think about: the one where nothing throws an error and the number is simply wrong. | Graduates have written queries; they have not maintained something that breaks at 3am because a system upstream changed a field without telling anyone. Nor has anyone yet asked them to say, in one sentence and on the record, what a number means. |
| Systems whose behaviour is learned flyvolo.ai/en/careers/machine-learning-engineer | Fundamentals and failure reasoning, on systems where no one can explain line by line why they did what they did — so the reasoning has to be statistical instead of mechanical. | Experiment design and enough statistics to be suspicious of your own result. And one thing coursework rarely states outright: a model that improves on the offline metric and gets worse for the people it is used on is the normal case, not an accident. |
The occupations on those paths, task by task
Junior software developer
Impact index 62 · confidence medium · assessed 2026-09-09 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Writing routine code | Automating | evidenced | Generating it is not the same as owning it. Someone still has to decide it is correct, and that someone has historically learned to judge by writing it first. |
| Debugging unfamiliar systems | Being augmented | evidenced | This is where the ceiling on generated code shows up most clearly, and it is the skill that separates a junior from someone who can be left alone. |
| Reviewing generated code | New task | evidenced | Reviewing well requires having written the thing being reviewed. If the writing work that trained that judgement is the work being automated, this task has a supply problem within a few years. |
| Turning a vague request into a spec | Still human-led | inferred | This is normally a senior responsibility; listing it as a junior task describes where the role is heading, not where most junior jobs are today. The intermediate rungs are what disappeared. |
| Deciding how pieces fit together | Still human-led | evidenced | Saying seniors are safe is not a statement about entry. The path to senior ran through two years of routine code, and nothing has replaced that path. |
Most recent verified records
- Employment of 22-25-year-olds in the two most AI-exposed occupation quintiles fell about 11% from November 2022 to June 2026, while the least-exposed quintiles grew about 10% (2026-08-12 · Labour impact · Stanford Digital Economy Lab — Brynjolfsson, Chandar & Chen, "Canaries in the Coal Mine?" (August 2026)) record
- Thirty-one of forty-five US federal agencies answered yes to generating code using AI in the 2025 federal inventory of commercial AI, some naming GitHub Copilot, Claude Code, Codex and Poolside (2026-04-13 · Deployment · OMB — 2025 Federal Agency AI Use Case Inventory, consolidated commercial off-the-shelf AI use cases (CSV, 45 agencies × 20 questions)) record
- The UK Home Office's engineering standard requires every AI-assisted output to be reviewed and approved by a person before production, and keeps accountability for designs with the team (2026-03-20 · Constraint · UK Home Office — Engineering Guidance and Standards, SEGAS-00020 "Use AI" (last updated 20 March 2026)) record
Something a student can check this week
Take a generated PR — yours or a colleague's — and write a review that finds a real problem, with the reasoning. Do this weekly and keep them.
Backend developer
Impact index 54 · confidence low · assessed 2026-09-12 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Endpoints and the plumbing between them | Automating | evidenced | Generating an endpoint is not the same as deciding it should exist, what it must guarantee, or what happens when it is called twice. The typing was rarely the expensive part of this work. |
| Keeping it correct when things happen at once | Still human-led | evidenced | Uber reports that a tool now drafts fixes for low-level data races in its Go code, with code owners accepting most and rejecting some that passed tests but were wrong; that is races in one codebase, not transactions or ordering across services. Incidents are almost never attributed publicly to generated concurrency code, and that silence is not evidence that it does not happen. |
| The data model, and changing it while it is in use | Still human-led | evidenced | Google reports that most of the code in one large migration of IDs was written by AI while every change was reviewed and rolled out by people; that is a code change at one company, not a live schema migration. The claim is about who decides and answers for it, not about who types it. |
| Who is allowed to see what | Still human-led | evidenced | The UK's cyber security agency advises writing authorisation logic closer to by hand, and a benchmark found about half of the models' working backends exploitable; neither is a defect study separating generated from hand-written code in the same codebase, which is what would settle it. Regulated environments already require human sign-off here, which may be doing more of the work than the difficulty is. |
| What it costs and how fast it is | Being augmented | evidenced | Nothing here measures how much of this is now tool-driven in practice, and the answer differs enormously between a team with an observability budget and one without. |
| Being the one the pager wakes | Still human-led | evidenced | This says nothing about whether on-call load is rising or falling, which is the question most engineers actually care about and the one nobody publishes. |
Most recent verified records
- The UK's National Cyber Security Centre advises developers to move towards writing code by hand when building authentication or authorisation logic with AI assistance (2026-06-18 · Constraint · National Cyber Security Centre (UK) — blog, "The 'vibe coding spectrum' approach to AI-assisted software development" (Toby W, Principal Security Architect; published 18 June 2026)) record
- Thirty-one of forty-five US federal agencies reported a commercial AI coding tool in the 2025 federal inventory, naming GitHub Copilot, Claude Code, Codex and Poolside (2026-04-13 · Deployment · OMB — 2025 Federal Agency AI Use Case Inventory, consolidated commercial off-the-shelf AI use cases (CSV, 45 agencies × 20 questions)) record
- On 392 backend tasks, the best model's code was correct and secure only 35% of the time, and about half of the models' working programs could be exploited (2025-05-30 · Constraint · Vero, Mündler et al. (ETH Zurich, LogicStar.ai and others) — "BaxBench: Can LLMs Generate Correct and Secure Backends?", arXiv:2502.11844v3 (30 May 2025; ICML 2025)) record
Something a student can check this week
Find one invariant in your system that must never be violated, and check whether anything actually enforces it or whether everyone just avoids breaking it.
Software tester / QA engineer
Impact index 72 · confidence medium · assessed 2026-09-10 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Manual regression testing | Automating | evidenced | This is the largest block of hours in most QA roles. At Uber the test engineers who ran the scripts were moved to building the system's context, and the company says headcount did not grow in proportion rather than that it fell; whether the remaining tasks can absorb everyone whose week this filled is not established here. |
| Writing test cases and automation | Automating | evidenced | Generated tests confirm what the code does rather than what it should. Teams that measure coverage rather than defects will conclude this task is solved, which is a measurement problem that hurts testers. |
| Exploratory and adversarial testing | Still human-led | evidenced | The hypothesis about where to look is built by having done the regression work that is disappearing. This task is protected by experience that the pipeline no longer produces. |
| Deciding whether it ships | Still human-led | evidenced | On many teams this call belongs to an engineering manager, not a tester. Where it does, protecting it protects someone else's job. |
| Testing systems with a model inside | New task | evidenced | Few practitioners is a statement about now. The discipline is young enough that its eventual size is unknown, and it may end up inside the model teams rather than in QA. |
Most recent verified records
- Uber says its AI mobile-testing system runs 1,013 end-to-end tests continuously in CI and has turned its test engineers from "clickers" into people who build its context (2026-07-30 · Deployment · Uber — "DragonCrawl: A Generative, Intent-Based Framework for Scalable Mobile End-to-End Testing", arXiv:2607.28750 (30 July 2026; accepted at ASE-Industry 2026)) record
- A survey of 200 SRE and DevOps leaders reported 43% of AI-generated code changes still need manual debugging in production after passing QA and staging (2026-04-14 · Constraint · VentureBeat) record
- Meta says its LLM-based test generation is deployed in its continuous integration system, and that its coverage-test generator has been deployed at Meta for two years (2026-01-30 · Deployment · Becker, Harman et al. (Meta) — "Just-in-Time Catching Test Generation at Meta", arXiv:2601.22832 (30 January 2026; FSE Companion '26)) record
Something a student can check this week
Find one bug this week that no existing test or generated test would have caught, and write up how you thought of it. That write-up is your job description.
AI implementation lead
Impact index 16 · confidence low · assessed 2026-09-11 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Deciding what to try first | Still human-led | evidenced | Hard to automate is not the same as durable. This whole role exists because a transition is under way; when the tools become ordinary the way spreadsheets did, choosing what to automate becomes part of running a function rather than a function of its own. The safety here is the safety of scaffolding. |
| Checking it on this company's own work | Being augmented | evidenced | A good evaluation tells you the tool works on the cases you thought to collect. It says nothing about the cases nobody thought of, and those are where the expensive failures live — so a passing score is a reason to start watching, not a reason to stop. |
| Wiring it into what is already there | Being augmented | inferred | Cheap integration is why pilots multiply and why most of them are still running as pilots. Being able to connect a tool in an afternoon does not establish that anyone changed how they work, and the two get reported as the same milestone. |
| Deciding what it may do unsupervised | Still human-led | evidenced | Guardrails being human work does not mean they get written. In most companies nobody has been given this task, so the boundary is settled by default — by whoever configured the tool — and that default is invisible until something goes out that should not have. |
| Getting people to actually use it | Still human-led | evidenced | This is the task most likely to be cut, because it is slow and produces no artefact. A company that buys the tools and skips this gets the licence cost and the pilot, and reports both as progress. |
| Being there when it gets something wrong | Still human-led | evidenced | The directive names a senior official, not the implementation lead, and covers one government. Carrying the blame is not the same as holding the authority: this role often answers for decisions it cannot overrule, and that combination is the most common reason people leave it. |
Most recent verified records
- Census BTOS supplement: 18% of US firms used AI in a business function, and 57% of those in three functions or fewer (2026-04-01 · Deployment · U.S. Census Bureau, Center for Economic Studies (CES-WP-26-25)) record
- Canada's federal rule on automated decisions makes a named senior official responsible for overrides, failures, corrective action and recourse (2025-06-24 · Constraint · Treasury Board of Canada Secretariat — Directive on Automated Decision-Making (modified 24 June 2025)) record
- China's labour ministry put 17 new occupations and 42 new work-types out for comment, adding a 'generative AI system tester' work-type under the existing 'generative AI system operator' occupation (2025-05-08 · Labour impact · 中国就业网(人力资源和社会保障部)——《职业上新!42 个新工种亮相》,2025-05-09) record
Something a student can check this week
Name the one number your work should move, then find out who owned that number before you existed. Go and ask them whether they think you are helping. Their answer, not your dashboard, is where your funding comes from next year.
Frontend developer
Impact index 66 · confidence low · assessed 2026-09-12 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Turning a design into a working interface | Automating | evidenced | A screen that renders is not a screen that ships. It says nothing about whether the result is maintainable, whether it matches an existing design system, or how it behaves in the states nobody drew — empty, loading, error, half a list. |
| The states nobody drew | Still human-led | inferred | This is a judgement about where the work sits, not a measurement of it. Teams do not count how much of their frontend code is edge-case handling, so there is no figure to check this against, and the balance differs enormously between a marketing page and a trading screen. |
| Making it survive real devices and networks | Being augmented | evidenced | A government standard requires this testing for its own services; nothing here says how many hours it takes elsewhere or whether teams do it at all — a great deal of shipped frontend work never gets this attention, and the tools existing does not mean they were run. |
| Making it work for everyone, because it is required | Still human-led | evidenced | A regulator's action has listed the barriers left on sites that relied on an automated overlay, without naming the sites; what would settle it further is an enforcement action that names a specific interface and what was wrong with it. Requirements differ by market and by whether the product is consumer-facing. |
| Owning the components everyone else uses | Still human-led | evidenced | Whether this is a job or a side duty depends entirely on team size, and how many teams staff it deliberately is not something anyone counts. |
| Building interfaces for things that answer back | New task | evidenced | New work appearing is not the same as new headcount: this is being absorbed into existing roles, and nothing here says anyone was hired to do it. |
Most recent verified records
- The US Justice Department extended state and local governments' web accessibility deadlines to 2027 and 2028, saying generative AI does not yet reliably automate fixing inaccessible content (2026-04-20 · Constraint · US Department of Justice, Civil Rights Division — Interim final rule, Extension of Compliance Dates for Nondiscrimination on the Basis of Disability; Accessibility of Web Information and Services of State and Local Government Entities (Federal Register, 20 April 2026)) record
- Figma reported that in March 2026 it began enforcing the AI credit limits it had introduced across all its seats in 2025, and started selling credit add-ons and pay-as-you-go usage (2026-03-01 · Constraint · Figma, Inc. — Form 10-Q for the quarter ended 2026-06-30, filed 2026-08-05 (SEC EDGAR)) record
- AI chatbots on UK government services must let users know answers are not from a human and might be inaccurate, and how to reach a person, says the Service Manual (2025-09-23 · Constraint · Government Digital Service — Service Manual, "Using artificial intelligence (AI) in services" (published 23 September 2025)) record
Something a student can check this week
Take one screen in your product and list every state it can be in. Then check how many of them actually exist in the code.
Data engineer
Impact index 50 · confidence low · assessed 2026-09-12 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Building the pipeline | Automating | evidenced | The pipeline running is not the same as the numbers being right. Building it was never where the time went in a mature data team; keeping it true was. |
| When the data is wrong and nothing errored | Still human-led | evidenced | A judgement about the nature of the failure rather than a measurement of how often it happens. Silent data errors are not quantified anywhere, partly because by definition they are found late or not at all. |
| Owning what a number means | Still human-led | evidenced | That benchmark is about querying rather than about who owns definitions, and its 21.3% is one agent on one model at the end of 2024. It establishes that enterprise schemas defeat current models; it does not establish that anyone is being hired to maintain semantics. |
| What it costs to keep asking | Being augmented | evidenced | Says nothing about whether costs are rising or falling overall, and cheap querying tends to increase the number of questions asked, which can move the bill in either direction. |
| What must be kept, deleted, or never collected | Still human-led | inferred | Requirements differ sharply by market and by data type, and what would settle this task is an enforcement action naming a data team's retention or deletion practice rather than the company as a whole — a regulator saying who was supposed to delete what, and by when. |
| Feeding the systems that answer in sentences | New task | evidenced | New work appearing is not new headcount. In most companies this is being absorbed by whoever already owns the warehouse, and nothing here says otherwise. |
Most recent verified records
- The US commodities regulator reported in the federal AI inventory that it runs an isolation-forest model daily to flag potentially erroneous data loads (2026-04-13 · Deployment · OMB — 2025 Federal Agency AI Use Case Inventory (entry CFTC-001)) record
- Thirty-five of forty-five US federal agencies reported running a knowledge retrieval system over their own agency information (2026-04-13 · Deployment · OMB — 2025 Federal Agency AI Use Case Inventory, consolidated commercial off-the-shelf AI use cases (CSV, 45 agencies × 20 questions)) record
- On 632 enterprise data workflows drawn from real warehouses, a code agent solved 21.3% — against 91.2% on the earlier academic benchmark Spider 1.0 (2024-11-12 · Constraint · Spider 2.0: Evaluating Language Models on Real-World Enterprise Text-to-SQL Workflows (arXiv:2411.07763)) record
Something a student can check this week
Take one metric on a company dashboard and ask two teams what it means. If the answers differ, you have found the work.
Machine learning engineer
Impact index 47 · confidence low · assessed 2026-09-12 — an index of how much of this job's task mix is in motion, not a probability of job loss. Full page
| Task | Direction today | Basis | What this does not establish |
|---|---|---|---|
| Building a model for the problem | Automating | evidenced | This site's four directions cannot tell those two mechanisms apart, and the difference matters to anyone planning around this page: a task that became purchasable can become unpurchasable again when the vendor's price, licence or capability moves, in a way that a task a machine learned to do does not. It also says nothing about the domains where a bought model is not an option — narrow, proprietary, latency-bound or regulated ones — which are not rare. |
| Deciding what counts as good enough | Still human-led | evidenced | A judgement about the nature of the work, not a measurement of how it is staffed. Plenty of teams do not do this at all and ship on a vendor's published benchmark, which is the failure this describes rather than evidence against it — but how many is not counted anywhere. |
| Getting data the thing can learn from | Being augmented | evidenced | Nothing here measures how much of a given team's time this takes, and the answer differs by an order of magnitude between a team with an existing data asset and one starting from nothing. Using a model to label data that trains a model also has known failure modes this judgement does not weigh. |
| When it quietly stops working | Being augmented | evidenced | The count here covers only the federal systems that chose to answer, and it describes processes, not how well they work. A model running unwatched for a year is common — where nobody watches, this task is not done at all. |
| Hand-crafting the inputs | Automating | evidenced | The evidence is individual companies describing their own systems, not a measurement across the field. Domain-specific feature design is also still load-bearing in some regulated settings where an unexplainable input is not allowed, which this judgement does not separate out. |
| Answering for what it does to people | New task | evidenced | A rule on the books that does not yet apply, and it names the organisation, not the individual engineer; requirements differ sharply by market and by what the system is used for, and nothing here shows who inside a company has been told the work is theirs. |
Most recent verified records
- The 2025 US federal AI inventory lists 227 deployed systems their agencies rated high-impact; testing is recorded as complete for 44, in progress for 81 and left blank for 102 (2026-04-13 · Deployment · OMB — 2025 Federal Agency AI Use Case Inventory (computed over the individually-reported CSV)) record
- Instacart replaced separately trained query-understanding models with one fine-tuned LLM serving millions of weekly searches, moving its engineers' core challenge away from feature engineering (2025-11-21 · Deployment · Instacart — "Building The Intent Engine: How Instacart is Revamping Query Understanding with LLMs" (Nov 21, 2025), by five Instacart engineers) record
- In the 2024 US federal AI inventory, of 76 AI systems in operation that described their monitoring, 47 relied on manual model updates and only 5 had a full pipeline with drift detection (2025-01-23 · Deployment · Office of Management and Budget — 2024 Federal Agency AI Use Case Inventory, consolidated raw data v2 (GitHub, committed 23 January 2025)) record
Something a student can check this week
Take the benchmark your team quotes and find out whether its data could have been in the model's training set. If you cannot rule it out, the number means less than the team thinks.
Method and sources
Every occupation is broken into the tasks people actually do. Each task carries one of four directions and states what that judgement does not establish. Evidence enters only after a named person has read the source; a vendor's own claims never move a direction. Method: flyvolo.ai/en/method · open data under CC BY 4.0: flyvolo.ai/en/data.