Breaking Into AI: Why Proof Now Beats Paper, and How Students, Fresh Grads and Career Shifters Can Build It

I did not plan an AI career. I studied applied physics, ended up writing data pipelines for SARS-CoV-2 genome surveillance, took a master’s in data science, and only then walked into industry. Every one of those moves looked sideways at the time. If you are a student, a fresh graduate or someone thinking about switching careers, I want you to know that sideways is normal — and that you are probably closer than you think. But the rules of the door you are trying to walk through have changed, and it pays to understand how.

Here is my argument in one sentence: for AI roles, shipped and measured proof now outranks credentials — and because the entry-level squeeze is landing hardest on new workers, building that proof before you apply is the decisive move. A degree still helps. A certificate rarely hurts. But neither answers the question every hiring manager in this field is now asking of juniors: can you already do some of the job?

This guide is the one I wish someone had handed me, rebuilt as a proper case. It starts with the forces reshaping the AI job market and what the data actually shows, including where it is weaker than the headlines. It then sets out a simple framework — a proof ladder and a map of five career paths — and goes deep on each path: what the work is, how skills stack up, what the demand evidence says, and how people actually break in. After that come a few real cases, the strongest objections to my argument (some of which are partly right), what all this means for students, fresh grads and career shifters, a concrete 90-day plan, and an honest note on what the evidence cannot tell us.

Exhibit 1

1. UndergradApplied physicsA physics degree, not a computing one.
2. ResearchPhilippine Genome CenterData pipelines for SARS-CoV-2 genome surveillance.
3. Grad schoolMS Data Science, AIMAsian Institute of Management.
4. IndustryData scientist, AIAFirst industry role, in insurance.
5. ManagementData science manager, TDCX SingaporeLeading data work inside a global BPO.
6. NowLead AI engineer, UK consultancyEnterprise LLM delivery. Teaching at CIIT and Eskwelabs since 2023.

The author's own career, listed in order; none of these steps was planned as a step towards AI.

Context: why this matters now

Three forces are acting on the AI job market at once, and they pull in different directions. Understanding all three is the difference between a plan that works and one that feels productive but never converts into an offer.

The first force is price. Employers are paying more for AI skills, and the premium is rising rather than settling. PwC’s Global AI Jobs Barometer, which analyses close to a billion job advertisements each year, found that postings asking for AI skills carried a 56% wage premium in its 2025 edition, up from 25% the year before. The 2026 edition, covering more than a billion ads across 27 countries, put the premium at 62%. That average hides a wide range — from 16% in government roles to 118% at the top end — but the direction is unambiguous.

Exhibit 2

25%

56%

62%

2024 edition2025 edition2026 edition

The 25% figure is the previous year's premium as cited in PwC's 2025 release. The 2026 premium ranges from 16% (government) to 118% depending on sector. Advertised wages, not realised pay.

Source: PwC Global AI Jobs Barometer 2025 and 2026 (official releases)

View data table

Edition	AI-skill wage premium
2024	25%
2025	56%
2026	62%

The second force is volume. AI-skill demand is no longer a niche. PwC’s 2026 analysis found postings requiring AI skills grew 69%, against 9% for all job postings. Lightcast’s data for the Stanford AI Index 2026 shows the same shape from a different angle: 2.5% of all US job postings now mention AI skills, up 55% in a single year, 72% since 2022 and roughly 300% over a decade. LinkedIn’s Jobs on the Rise list for the US in 2026, as reported secondhand, put AI engineer at number one, AI consultants and strategists at two, data annotators at four, and AI and machine-learning researchers at five.

Exhibit 3

Postings requiring AI skills+69%

All job postings+9%

Source: PwC 2026 Global AI Jobs Barometer

View data table

Segment	Posting growth
Postings requiring AI skills	+69%
All job postings	+9%

For readers in Southeast Asia, the geography matters. Lightcast’s country comparison puts Singapore at the top of the world for the share of postings asking for AI skills, at 4.769% — nearly double the US share. Hong Kong, Luxembourg and Spain follow. If you are in the Philippines, the most AI-intensive labour market on the planet is a short flight away, and many of its employers hire remotely across the region.

Exhibit 4

Singapore4.8%

Hong Kong3.5%

Luxembourg3.4%

Spain3.3%

United States2.5%

Singapore's figure is 4.769%, shown rounded. The US figure is the national share from the same Lightcast analysis.

Source: Lightcast, data for the Stanford AI Index 2026

View data table

Market	Share of postings mentioning AI skills
Singapore	4.769%
Hong Kong	3.5%
Luxembourg	3.4%
Spain	3.3%
United States	2.5%

The third force is speed. The skills themselves are moving. PwC found that the skills employers ask for are changing 66% faster in AI-exposed jobs than in others. The clearest example is agentic AI — systems that plan and take actions, not just answer questions. Lightcast found agentic-AI skills went from 0.06% of US postings in 2024 to 0.23% in 2025, about 90,000 ads. That is still small. But it nearly quadrupled in a year, and it is the kind of skill no syllabus written three years ago could have included.

Exhibit 5

0.06%

0.23%

20242025

About 90,000 US postings in 2025. Small in absolute terms; the point is the rate of change.

Source: Lightcast, data for the Stanford AI Index 2026

View data table

Year	Share of US postings
2024	0.06%
2025	0.23%

Put the three forces together and you get a market that is paying more, hiring more and changing faster — which sounds like pure good news for anyone trying to get in. It is not quite. The next section explains why.

The evidence: credentials are loosening, but the entry door is narrowing

If you read only one section of this guide, read this one. Two findings, both robust, appear to contradict each other. They do not, and the way they fit together is the whole basis of my thesis.

Finding one: degree requirements are falling

PwC’s 2025 analysis found that between 2019 and 2024 the share of job postings requiring a degree fell in exactly the jobs AI touches most. In AI-augmented jobs — where AI helps people do their work — it dropped from 66% to 59%. In AI-automated jobs — where AI can take over tasks — it dropped from 53% to 44%. Employers are, slowly, shifting from credentials towards demonstrated skill.

Exhibit 6

Source: PwC 2025 Global AI Jobs Barometer

View data table

Job type	2019	2024
AI-augmented jobs	66%	59%
AI-automated jobs	53%	44%

Read the numbers carefully, though. A majority of AI-augmented postings — 59% — still ask for a degree. The trend is real, but “degrees don’t matter” is not what the data says. What it says is that a degree is becoming one kind of evidence among several, and that employers are increasingly willing to accept other kinds.

Finding two: entry-level roles now want senior skills

The harder news comes from PwC’s 2026 edition, which looked at 2.4 million US entry-level job postings. AI-exposed entry-level roles were seven times more likely to ask for senior-level skills than other entry-level roles. In plain terms: employers want juniors who arrive already useful.

That is the advertised side. The hiring side is starker. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen at the Stanford Digital Economy Lab used ADP payroll records covering millions of US workers to track what happened to employment by age and AI exposure. In their August 2026 revision, with data through June 2026, employment of 22- to 25-year-olds in AI-exposed occupations stood 19% below where it would have been had it kept pace with less-exposed peers. Experienced workers showed no comparable gap. The mechanism was reduced hiring of young workers, not more layoffs — and the decline was concentrated in occupations where AI mostly substitutes for tasks, while jobs where AI mostly complements people were flat or growing.

Exhibit 7

7×more likely that an AI-exposed entry-level role asks for senior-level skillsPWC 2026

2.4MUS entry-level job postings in that analysisPWC 2026

−19%employment of 22–25-year-olds in AI-exposed jobs vs. where it would be at peers’ paceSTANFORD DEL · 2026

The 19% figure is relative to less-exposed peers, not an absolute fall, and the authors describe it as descriptive rather than causal. Payroll data through June 2026.

Source: PwC 2026 Global AI Jobs Barometer; Brynjolfsson, Chandar & Chen, Stanford Digital Economy Lab (rev. Aug 2026)

How the two findings fit together

They fit because they describe the same shift from two sides. Employers are relaxing the proxy (the degree) and raising the bar (what you can already do). If AI can now handle many of the routine tasks that used to be how juniors learned on the job — first drafts, simple analyses, boilerplate code — then the remaining entry-level work is the part that needs judgement. A credential tells an employer you could probably learn that judgement eventually. A shipped, measured project shows you already have some of it.

There is one more nuance worth holding onto. PwC’s 2025 data found that jobs in more AI-exposed occupations still grew between 2019 and 2024 — by 38% — just more slowly than the 65% growth in less-exposed occupations. That is not a collapse. It is a market where the jobs are still there, but the competition for each one is sharper and the price of entry is higher.

Exhibit 8

Less AI-exposed occupations+65%

More AI-exposed occupations+38%

Source: PwC 2025 Global AI Jobs Barometer

View data table

Occupation group	Job growth, 2019–2024
Less AI-exposed	+65%
More AI-exposed	+38%

A word on the quality of this evidence, since you are making real decisions with it. PwC and Lightcast measure job advertisements, not hires; an ad that mentions “AI” may be keyword dressing rather than a real requirement. The Stanford study measures actual payroll, which makes it stronger, but it is US-only and descriptive. None of these sources is Philippine. I come back to these limits at the end; for now, the safest reading is that three independent methods — ad analysis, skills analysis and payroll data — all point the same way.

A framework: the proof ladder and the five paths

If proof is the currency, you need a way to tell how much you have. The simple model I use for this is what I call the proof ladder.

Exhibit 9

1. 01LearnedCourses, certificates, a degree. Necessary context; weak evidence on its own.
2. 02BuiltIt runs on your laptop. A notebook, a demo, a tutorial extended.
3. 03ShippedDeployed at a link someone else can use. At least one real user.
4. 04MeasuredOne honest number against a baseline: minutes saved, errors caught, cost per request.
5. 05SurvivedIt met real usage, broke, and you fixed it. You can explain what you changed and why.

Framework by the author. The rungs are cumulative: each assumes the one before it.

The ladder matters because each rung answers a different doubt in a hiring manager’s head. Learned answers “do they know the vocabulary?” Built answers “can they code?” Shipped answers “can they finish something outside a syllabus?” Measured answers “do they know whether it works?” And survived answers the question that matters most for any role touching production AI: “what happens when it breaks?” The entry-level squeeze in Exhibit 7 is, in ladder terms, employers moving the minimum from built to measured.

The second half of the framework is a map. Job titles in AI change every year or two, so don’t anchor on them. A more useful way to choose is to ask yourself two questions: how deep into the model do you want to go (from using models to building them), and how close to the customer and their problem do you want to be (from back office to in the room).

Exhibit 10

AI strategist / consultantDecides what is worth building and helps people adopt it

Forward-deployed engineerBuilds and ships AI inside a customer’s systems

Data & AI operationsKeeps data, labels and evaluations trustworthy

Research & model engineeringTrains, fine-tunes and improves models themselves

AI / ML engineerDepth of model work →Closeness to customer →

	AI / ML engineer	Forward-deployed engineer	AI strategist	Research / model eng.	Data & AI ops
In one line	Builds the AI system and keeps it running	Makes AI work for one customer, in production	Decides what is worth building and why	Makes the model itself better	Makes sure the data and tests can be trusted
A typical day	Code, APIs, data pipelines, tests	Customer calls, integration, debugging, shipping	Workshops, business cases, slide decks	Experiments, training runs, reading papers	Labelling, quality checks, evaluation sets
You will enjoy it if	You like building things that run reliably	You like people and messy problems as much as code	You like the why more than the how	You love maths and patient experiments	You are careful and like catching mistakes
First thing to build	A deployed app with tests	A tool a real user relies on	A business case for one AI use	A reproduced result from a paper	An evaluation set for a real task
Common way in	Software or data roles	Engineering plus client-facing work	Analyst, product or consulting roles	Graduate research	QA, analytics or domain expert roles

Framework by the author. Role placement is indicative; individual jobs vary. The table scrolls sideways on small screens.

You don’t have to pick forever. Most people move around this map; my own path started bottom-right in genomics research and ended up top-right. Pick the box that matches what you enjoy now, and build your first rungs of proof for that box. The rest of this section goes through each path in turn.

Deep dives: the five paths

Exhibit 11 summarises what each path demands at each rung of the ladder, and the strongest public demand signal for each. The subsections below explain the reasoning.

Exhibit 11

	AI / ML engineer	Forward-deployed engineer	AI strategist	Research / model eng.	Data & AI ops
Demand signal	#1 on LinkedIn Jobs on the Rise 2026 (US)	Postings up more than 1,000% year on year, Jan–Aug 2026	#2 on LinkedIn Jobs on the Rise 2026 (US)	#5 on LinkedIn Jobs on the Rise 2026 (US)	Data annotators #4 on LinkedIn Jobs on the Rise 2026 (US)
Proof that gets interviews	A deployed app with tests and a README	A tool one real user relies on weekly	A one-page business case with a baseline number	A reproduced paper result, code public	An evaluation set for a real task, with labelling guide
Proof that gets offers	Measured quality and cost per request, plus error analysis	Before-and-after numbers from that user, plus what you cut	A recommendation someone acted on, and what happened	A small extension or ablation beyond the paper	Agreement rates between labellers, and errors you caught
Proof that gets promoted	A system that survived real traffic and an incident	A handover: someone else runs it without you	An adopted programme with tracked results	Published or production-adopted research	An evaluation suite that gates releases
Most common failure mode	Demo-ware: impressive, never deployed	All people skills, no shipped code	Slides with no numbers	Theory with no working code	Treated as clerical; never shows judgement

Rungs are the author's framework. LinkedIn rankings are US-only and reported secondhand; data annotators stand in for the wider data & AI operations family. The table scrolls sideways on small screens.

Source: Demand row: LinkedIn Jobs on the Rise 2026 (US), via VnExpress; Lightcast via Fortune (FDE)

Path 1: AI / ML engineer

What the work is. The AI engineer builds the system around the model: the data pipelines that feed it, the retrieval layer that gives it context, the API that serves it, the tests that catch regressions and the monitoring that tells you when quality slips. In 2026 most of this work involves calling large models rather than training them, but the engineering is not lighter for that — it is different. A retrieval system that answers from the wrong document version is an engineering bug, not a model problem.

The skills ladder. Python beyond notebooks, SQL, git and one cloud platform are the floor. Above that sit the skills juniors usually lack: designing an evaluation set, measuring cost and latency per request, handling rate limits and failures gracefully, and knowing when a smaller, cheaper model is good enough. Senior AI engineers are distinguished less by knowing more frameworks than by knowing which failure is coming next.

Evidence of demand. AI engineer topped LinkedIn’s US Jobs on the Rise list for 2026, as reported secondhand. It is the broadest of the five paths and the one most job ads mean when they say “AI”.

How people break in. Usually sideways from software engineering or data roles. For a junior without that background, the realistic route is a portfolio that reaches measured on the ladder: one deployed project with a test set, an honest accuracy number and a cost per request. The common failure mode is demo-ware — a slick chatbot that never left localhost and has no answer to “how do you know it works?”

Path 2: Forward-deployed engineer

What the work is. The forward-deployed engineer (FDE) makes AI work for one customer, inside that customer’s systems: scoping a messy problem, integrating with their data and identity systems, shipping, and staying until it runs without them. The role originated at Palantir in the early 2010s and, according to The Pragmatic Engineer, Palantir had more FDEs than conventional software engineers until around 2016.

The skills ladder. Everything the AI engineer needs, plus the ability to run a discovery conversation, cut scope without losing the client, and explain trade-offs to executives. The senior version owns an outcome rather than a ticket.

Evidence of demand. Lightcast data reported by Fortune shows FDE postings up more than 1,000% in January–August 2026 compared with the same period a year earlier, with a median advertised salary above $188,000 against about $145,000 for traditional US software engineers. Those are US figures, and the role is still small in absolute terms, but no title in this guide is growing faster.

How people break in. Experience bars vary widely. The Pragmatic Engineer reported that Palantir hires FDEs with about a year of post-college experience, while Ramp prefers five or more years for its senior FDEs. Salesforce reported in March 2026 that 40–50% of moves into its Agentforce FDE team were internal transfers. For most readers, then, this is a second or third job rather than a first one — but it is a realistic target for career shifters who already sit close to customers. I have written a separate deep dive on the role for anyone considering it seriously.

Path 3: AI strategist / consultant

What the work is. Deciding what is worth building, and helping an organisation adopt it. That means mapping processes, sizing benefits, sequencing a portfolio of use cases, writing the governance rules and — most underrated — managing the change in how people work.

The skills ladder. At the bottom: structured problem-solving, business-case maths, clear writing. In the middle: enough technical literacy to know what a model can and cannot do reliably, and to tell a vendor demo from a production system. At the top: a track record of recommendations that were acted on and paid back.

Evidence of demand. AI consultants and strategists ranked second on LinkedIn’s US Jobs on the Rise list for 2026, as reported secondhand.

How people break in. From analyst, product, operations or consulting roles. The failure mode is the strategist who has never built anything and therefore cannot tell what is hard. My advice to aspiring strategists is the same as to engineers: build one small thing yourself, end to end, so that your slides are grounded in what actually breaks.

Path 4: Research and model engineering

What the work is. Making the model itself better — training, fine-tuning, evaluation research, efficiency work. It is the most mathematically demanding path and the one where frontier labs concentrate.

The skills ladder. Linear algebra, probability, optimisation and deep-learning frameworks at the floor; the ability to design a clean experiment and read results sceptically in the middle; original contributions at the top.

Evidence of demand. AI and machine-learning researchers ranked fifth on LinkedIn’s 2026 US list, as reported secondhand.

How people break in. Usually through graduate research. This is the one path where my “proof beats paper” thesis is weakest, and I say more about that in the counterpoints. Even here, though, the proof that stands out is code: a reproduced paper result that someone else can run, and ideally a small extension of it.

Path 5: Data and AI operations

What the work is. Keeping the inputs and the tests trustworthy: labelling and annotation, data-quality checks, building and maintaining evaluation sets, reviewing model output, and spotting drift. It is the least glamorous path on the map and, in my experience, one of the most important. A model can only be as good as the evaluation that tells you whether it is good.

The skills ladder. Care and consistency at the floor; writing labelling guidelines, measuring agreement between labellers and spotting systematic errors in the middle; owning an evaluation suite that gates whether a model ships at the top.

Evidence of demand. Data annotators ranked fourth on LinkedIn’s 2026 US list, as reported secondhand.

How people break in. From QA, analytics, annotation or domain-expert roles. The failure mode is being treated as clerical labour: if you never move beyond labelling into designing the tests, the role can be automated around you. The way up is to own the judgement — the guidelines, the edge cases, the decision about what counts as correct.

Case examples

Frameworks are easier to trust when you can see them operating. Here are three cases, two public and one from my own record.

Case 1: Salesforce fills a scarce role from the inside. In March 2026 Salesforce reported that it had tripled its Agentforce forward-deployed engineering team in six months, that 40–50% of moves into the role were internal transfers, and that since September 2025 it had run a six-week onboarding programme for new FDEs. It said it values curiosity over credentials. Venture firm a16z, writing about the same role, advised founders to hire for hustle and curiosity rather than advanced degrees, noting the job does not require a PhD. The lesson: when a role grows faster than the market can supply it, employers fall back on people whose work they have already seen. For a career shifter, the most underrated route into AI is the internal one — becoming visibly useful with AI inside the organisation you already work for.

Case 2: proof as a number, not a title. My first industry role, as a data scientist at AIA Digital+, involved building a data pipeline for next-best-action models that reduced the full-time effort needed for that model development by up to 95%. Later, at TDCX AI in Singapore, I led work on an LLM-based simulator that cut agent time-to-proficiency by at least 20%. Neither of those numbers came from my degrees. Both are what made the next conversation easier. The lesson: the career currency is a measured change in something a manager already tracks — effort, time, error rate, cost.

Case 3: hackathons as compressed proof. Before and alongside my industry work, I competed in public hackathons: third at USHER 2022, second of 123 teams in the youth category of the UN Big Data Hackathon 2022, first at the DOH-DAP Hack4Health 2024 with a fraud-detection platform for unified health data, and first at ICSC Hack-A-Street 2025. The lesson: a good hackathon forces you up the ladder in 48 hours — you must ship something, demo it to strangers, and defend it under questioning. Those are the same muscles an interview tests.

Counterpoints

An argument worth acting on should survive its strongest objections. Here are the four I hear most often, stated as strongly as I can, with my honest assessment of where each one is right.

Implications by reader type

The same thesis implies different moves depending on where you are standing. Here is how I would apply it to each of the three readers this guide is for.

If you are a student

Your advantage is a low cost of failure. You can try two or three directions this year and lose nothing but a few weekends. Your risk is the opposite of a career shifter’s: you have time but no evidence that you can finish something outside a syllabus. So use the time to climb the ladder early. Turn one class project into something with a real user — a student organisation, a professor, a family business — and measure it. Keep the path open, but lean towards the AI engineering or research paths if you enjoy the maths, and treat a data or AI operations internship as a feature, not a fallback: it teaches you how real data behaves, which most graduates learn painfully in their first job.

On maths: the research path needs a lot of it — linear algebra, probability, optimisation. The applied paths need solid statistical intuition: averages versus distributions, sampling, precision versus recall, and healthy suspicion of any single number. Take the maths courses seriously now, because you will never again have this much structured time to learn them.

If you are a fresh graduate

You are the reader the entry-level squeeze affects most directly. Your advantage is recent, rigorous training and a thesis or capstone you can turn into a shipped project in a few weeks. The single highest-return move available to you is to take that thesis off your laptop: deploy it, give it an evaluation set, work out its cost per request, and write a one-page README on what broke. That moves you from learned to measured on the work you have already done.

Then go narrow. Pick a domain — finance, health, logistics, government — and build a second project on real public data from it. Apply selectively, linking the project that matches each role, rather than sending the same CV everywhere. And look seriously at Singapore and regional remote roles: Exhibit 4 shows where AI-skill demand is densest.

If you are switching careers

You have less free time than anyone else on this page, and more of the scarcest input in applied AI: domain knowledge. You know where the work actually hurts — which process is slow, which report nobody trusts, which tickets keep coming back. The Salesforce case suggests the best route is often internal. Automate one painful step in your current job, with permission and non-sensitive data; measure the minutes saved; write it up; share it. Aim first for a hybrid role — the “AI person” on an operations, finance or QA team — rather than jumping straight to a pure engineering title.

Is it too late at 30 or 40? No. What changes with age is your free time and your tolerance for a pay cut, so plan for that honestly, and expect it to take longer than a bootcamp advert suggests.

If you work in BPO or shared services

This subsection is for the many Filipinos reading from a production floor, a QA desk or a workforce team. An IMF working paper using Philippine Labour Force Survey data found that roughly a third of Filipino workers are highly exposed to AI — but about 60% of those are in roles where AI is more likely to complement than replace them. BPO carries the highest displacement risk, which is exactly why it is worth moving now.

Exhibit 12

~1 in 3Filipino workers in highly AI-exposed occupations — about 60% of them in roles AI is more likely to augment

The IMF reports about one third highly exposed; of those, about 60% are also highly complementary (AI more likely to augment than displace). BPO carries the highest displacement risk. Exposure is estimated from occupational tasks, not observed job losses.

Source: IMF Working Paper WP/25/43 (Cucio & Hennig, 2025)

The IT-BPM industry crossed $40 billion in revenue with 1.9 million workers in 2025, and its 2026 targets are $42 billion and 1.97 million workers. It is full of processes that are documented, measured and repetitive — exactly where AI systems pay back first. That makes the people who know those processes from the inside a natural fit for the teams that will rebuild them. A BPO floor already runs on the measurement discipline AI projects need: handle time, quality scores, calibration. Few industries train people to think in baselines so thoroughly.

• Agents and senior agents: you know which tickets repeat and why. Start with the career-shifter track below and aim for data and AI operations or an internal automation team.
• QA analysts: quality forms and calibration sessions are evaluation work. Learning to build evaluation sets for AI tools is a short step from what you already do.
• Team leads and workforce managers: you already think in handle time, quality scores and staffing. That is how a forward-deployed engineer or AI strategist scopes a project.
• Trainers: you know how people learn a new process. AI adoption is mostly that problem.

What to do next: your 90 days

Whatever your track, the learning order is deliberately boring, and the most common mistake I see is collecting courses instead of finishing things:

1. Fundamentals. Python, SQL, git and basic statistics — enough to build without copying a tutorial.
2. Working with models. Call an LLM API, train one simple model, and add search over your own documents.
3. One shipped project. Deployed, used by a real person, with one number you measure. This is the step most beginners skip.
4. Evaluation and reliability. A test set, error analysis, cost per request, and basic monitoring.
5. Specialise. Only now pick a path from the map and a domain you care about.

Pick the track below that sounds most like you. Each one has the path that usually fits, the skills to learn first, one project to build, and a 90-day plan that walks you up the proof ladder.

At the end of the 90 days, check yourself against this list. A junior portfolio that gets interviews usually has:

1. One project that is actually deployed. A link someone can click — not a notebook on your laptop.
2. A real user. Even one person who uses it weekly beats a hundred GitHub stars.
3. A number. Minutes saved, errors caught, accuracy on a test set, cost per request. Pick one and measure it honestly against a baseline.
4. A write-up of what broke. Bad data, prompts that failed on edge cases, costs that surprised you. This is what you will talk about in interviews.
5. A domain. Insurance, health, logistics, education, customer service. Domain knowledge is the fastest way to look senior early.

If you can tick all five, you are on the measured rung, which is where the entry-level bar now sits. If you can tick four, you know exactly what to do in the next 30 days.

Limitations and methodology

I have tried to keep every number in this guide traceable, and I want to be equally clear about what those numbers cannot tell you.

• Job ads are not jobs. PwC and Lightcast analyse postings. A posting that mentions “AI” may be aspirational or keyword-stuffed, and wage premiums are advertised pay, not realised pay. Postings also overstate demand in hot fields, where employers advertise more speculatively.
• The strongest hiring evidence is American. The Stanford payroll study is the best evidence on actual hiring, but it covers US workers, and its authors describe it as descriptive rather than causal. The PwC entry-level analysis is also US-only. I have found no equivalent hiring data for the Philippines.
• Exposure is not displacement. The IMF study estimates exposure from the tasks in each occupation. It tells you where AI could change work, not where jobs have been lost.
• Some figures are secondhand. The LinkedIn rankings come from a news report rather than LinkedIn’s own publication, and the Salesforce figures are the company’s own account of its programme. I have labelled them accordingly.
• The sources have interests. PwC sells AI advisory services; LinkedIn and Lightcast sell labour-market data; Salesforce sells AI agents. None of that makes their numbers wrong, but it is a reason to prefer findings that several independent sources agree on — which is why I lean hardest on the pattern across sources rather than any single figure.
• My own observations are one practitioner’s. The field notes describe what I have seen teaching and hiring. They are useful as illustration; they are not a sample.
• The framework is mine. The proof ladder and the five-path map are tools for thinking, not measured categories. Real jobs blur the boundaries.

The door is narrower, and proof is the key

The AI job market in 2026 is paying more, hiring more and changing faster than almost any other part of the economy — while quietly raising the bar at the front door. Employers are relaxing the proxy and raising the standard. That combination punishes people who wait for a credential to do the talking, and rewards people who arrive already holding a shipped, measured, battle-tested piece of work.

So pick your track above. Put the next 90 days in your calendar — literally, as blocked time. Ship something small in the first 30, even if it is rough. Then measure it, write down what broke, and show it to someone. Sideways is normal. Proof is what turns sideways into forward.