AI Certifications in 2026: Which Are Worth It — and When They Aren't

What you need to know

• 2026 retired two of the three flagship AI exams. AWS's ML Specialty went on 31 March 2026; Microsoft's AI-102 followed on 30 June 2026. Their successors — AWS ML Engineer – Associate and Azure AI-103 — are the credentials that now matter.
• Certificates are tiebreakers, not door-openers. Hiring managers in both markets describe screening on portfolio evidence first, credentials second. A heavy cert list with no shipped work is a known red flag.
• India weighs credentials more formally than the UK — services firms and cloud partners count them for billing and partner tiers. UK screening leans on portfolio and interviews.
• The winning pattern is a pairing: one cloud cert matched to your target employers' stack, plus three shipped projects that prove you can use it.

The uncomfortable pattern: heavy cert lists and weak practical skill

Let us start with the finding nobody selling exam vouchers wants on a landing page. There is no rigorous public study proving that certifications predict weak engineers — anyone who quotes you a precise percentage here is inventing it. But there is a remarkably consistent qualitative pattern, and it shows up wherever hiring practitioners speak candidly. CertSelect's 2026 analysis of the US market puts it bluntly: certificates "do not get AI jobs — GitHub, demonstrated production work, and shipped models do", and hiring managers screen portfolio evidence first, certificates second. Practitioner communities on Reddit, surveyed in AI Tool Discovery's round-up, converge on the same hierarchy: real experience beats portfolio projects, which beat vendor certifications, which beat generic online completion certificates — with a strong Kaggle finish rated above any certificate at all.

Why does a long credential list sometimes read as a warning sign rather than an asset? The mechanism is simple opportunity cost. Every certification is six to eight weeks of structured study, and study optimised for a multiple-choice exam produces exam skill, not build skill. A candidate who spent eighteen months collecting seven badges made a choice — probably an honest, hard-working one — to optimise for the measurable thing. A screener looking at that CV sees eighteen months in which nothing was shipped, deployed, broken, or fixed. The certificates are not the problem; the silence around them is. This is exactly the failure mode our proof-of-work portfolio guide exists to prevent.

None of this means certifications are worthless. It means they are a second-order signal: powerful when they confirm evidence that already exists, close to inert when they are the only evidence on the page. Hold that framing — confirmation, not substitution — and every decision in the rest of this guide gets easier.

The three cert tracks employers actually recognise

When employers in India or the UK take a certification seriously, it is almost always one of three vendor tracks — AWS, Google Cloud, and Microsoft Azure — because those are the proctored, verifiable exams tied to the platforms companies actually pay for. But 2026 rearranged all three tracks, and a lot of career advice on the internet has not caught up.

AWS. The venerable Machine Learning – Specialty (MLS-C01) — for years the default answer to "which ML cert?" — was retired on 31 March 2026; existing holders keep the credential for three years from the date earned. Its practical successor is the AWS Certified Machine Learning Engineer – Associate (MLA-C01): US$150, 65 questions in 130 minutes, valid three years, focused on building, deploying and operating ML on SageMaker. An updated MLA-C02 version opens for registration on 1 September 2026, so if you are mid-preparation, check which version you will sit. For senior builders working with Bedrock and generative applications, AWS added a Generative AI Developer – Professional exam at US$300.

Google Cloud. The Professional Machine Learning Engineer (PMLE) survived the year unchanged: US$200 plus tax, two hours, 50–60 questions, with Google recommending three-plus years of industry experience including at least one on Google Cloud. Google certifications carry a two-year validity, the shortest of the three tracks. Practitioner consensus rates PMLE the hardest of the trio and the most respected for senior roles — and also the most niche if your market runs on AWS or Azure.

Microsoft Azure. The Azure AI Engineer Associate credential (exam AI-102) — long the enterprise favourite — was retired by Microsoft with a final exam date of 30 June 2026. Its successor, exam AI-103, awards the new Azure AI Apps and Agents Developer Associate credential: US$165 in the US, around two hours, built around Azure AI Foundry with the heaviest domain weighting on generative AI and agentic solutions. The rename tells you where Microsoft thinks the jobs are: agents and AI applications, not service configuration.

Pricing as of July 2026. AWS and Google bill in US dollars worldwide; the INR and GBP figures are approximate conversions at July 2026 exchange rates and will move with the currency. Microsoft prices exams regionally — Indian candidates have historically paid well below the US dollar price — so confirm your local fee at booking.

When a cert genuinely pays: gates, benches and borders

The honest case for certification is not "it proves skill". It is that certain doors are literally gated on it, and if you are standing in front of one of those doors, the exam fee is the cheapest key you will ever buy.

Cloud-partner gatekeeping

AWS, Microsoft and Google all run partner programmes in which consultancies must employ minimum numbers of certified individuals to hold or advance a partner tier — and partner tier determines the deals a firm can bid on. Inside such a firm, your certification is not a personal vanity metric; it is inventory the business needs. Employees are frequently funded, given study time, and in some cases paid bonuses to sit these exams. If you work at, or want to work at, a cloud-partner consultancy in Bengaluru, Pune, London or Manchester, one matching cert is close to free money: the firm needs the badge count, and holding one moves you up the queue for the interesting client work.

Services firms and the bench

India's large IT services companies institutionalised this logic years ago. Certifications feed directly into how consultants are allocated to projects and presented to clients — a slide that says "team of six, five AWS-certified" wins procurement conversations. If your next two career moves are within that ecosystem, a current cloud AI credential has direct, mechanical value that no amount of GitHub greenery replaces, because the client's procurement team never opens GitHub.

Visa-adjacent and cross-border CVs

The third genuine payoff is portability. When a CV crosses a border — an engineer in Chennai applying to a London scale-up, or a UK graduate applying into a Dubai or Singapore team — the screener often cannot calibrate the local signals. They may not know whether your university is strong or your previous employer is respected. A vendor certification is one of the few lines on the page the reader can verify in thirty seconds regardless of geography, which is why cross-border applicants and anyone assembling an immigration-adjacent evidence pack tend to get more mileage from certificates than domestic applicants do. It is not that the cert is worth more abroad; it is that your other signals are worth less to a reader who cannot decode them.

If none of these three situations describes you — no partner-tier employer, no services bench, no border crossing — be suspicious of your own reasons for wanting the exam. Often what a builder actually wants is a structured syllabus and a deadline, and there are cheaper ways to buy those.

India vs UK: how each market weighs credentials

The dual-market difference here is real, and it changes the advice depending on which side of it you sit — or whether you intend to move between them.

In India, credentials carry formal, institutional weight. The services giants and the sprawling cloud-partner ecosystem treat certifications as countable assets: they gate partner tiers, appear in client proposals, and get parsed by recruiter keyword filters at volume. A fresher in Hyderabad with an AWS associate cert genuinely does clear screens that an identical fresher without one does not, because Indian hiring at scale relies on filterable proxies to survive applicant volumes that UK recruiters rarely see. The trap is over-rotation: because certificates visibly work at the screening layer, ambitious candidates stack five of them and neglect the portfolio — and then stall at the interview layer, where Indian product companies and GCCs increasingly interview exactly like their London and San Francisco counterparts: show us what you built. Our zero-experience playbook covers how to build the evidence layer alongside the credential layer rather than instead of it.

In the UK, the screening culture leans harder on demonstrated work. Startups and scale-ups — the fastest-growing slice of UK AI hiring — will typically open your GitHub before they notice your certifications, and mid-sized product companies weigh a strong repository and a crisp technical conversation over any badge. Certifications matter in specific UK pockets: the big consultancies (which face the same partner-tier arithmetic as their Indian counterparts), enterprise and financial-services environments with formal vendor relationships, and public-sector-adjacent work where procurement frameworks reward documented competencies. A UK candidate outside those pockets can usually defer certification entirely in favour of shipped work; a UK candidate inside them should treat one Azure or AWS credential as standard kit.

For builders moving between the markets — in either direction — the practical synthesis is: build to UK screening standards (portfolio-first, because that also satisfies India's interview layer) and carry one current cloud credential (because it satisfies India's filter layer and the UK's enterprise pockets, and it travels). One, not five.

Courses that beat certificates: fast.ai, Hugging Face, DeepLearning.AI

If your goal is skill rather than a gate pass, three course families consistently out-teach the certification syllabi — and two of them are free.

fast.ai's Practical Deep Learning for Coders remains the best top-down entry into deep learning: you train a working model in the first lesson and back-fill the theory as you go. It is free, PyTorch-native, and its projects are designed to be published — which means every chapter can become a portfolio artefact. In research-leaning and PyTorch-heavy shops in both markets, "I worked through fast.ai and here is what I built" lands better than most certificates.

Hugging Face's free courses at huggingface.co/learn — the LLM course and the agents course in particular — are the fastest route from zero to the tooling that 2026 job descriptions actually name: transformers, fine-tuning, evaluation, agent frameworks. Because the exercises live on the Hub, your coursework is public by default, and a Hub profile with real models and Spaces is itself a hiring signal in a way no PDF certificate can be.

DeepLearning.AI sits between the free courses and the vendor exams. Andrew Ng's short courses are free and current; the deeper specialisations are hosted on Coursera, which typically bills around US$49–59 a month as a subscription (pricing as of July 2026 — Coursera changes its packaging often). Practitioner communities rate the teaching quality highly while remaining honestly uncertain about the completion certificates as standalone hiring signals — which is exactly the right way to consume them: buy the learning, do not rely on the badge.

Why do these beat certificates for skill-building? Because their unit of progress is a working artefact, not a bank of multiple-choice questions. Exam preparation optimises recall of service limits and API names; course projects force you through the debugging, data-wrangling and deployment friction that interviews probe. The output of eight weeks of fast.ai or Hugging Face study is three repositories; the output of eight weeks of exam cramming is one line on a CV. If those repositories are chosen well — see our three-project portfolio guide — the difference in interview conversion is not close.

The pairing strategy: one cloud cert + three shipped projects

Everything above converges on a single, boring, effective formula: one recognised cloud certification, matched to the stack your target employers run, sitting next to three shipped projects that prove you can use it.

The pairing works because the two signals cover each other's weaknesses. The certificate is verifiable but shallow — it proves you studied, to a proctored standard, on a named platform. The projects are deep but unverifiable at a glance — anyone can claim a repository; a screener has to trust it. Together they corroborate: the cert says "the knowledge is real and current", the projects say "the knowledge produces working software", and the combination survives every layer of the funnel — recruiter filter, hiring-manager skim, and technical interview — where either signal alone fails at least one layer.

Choosing the one cert is a market-reading exercise, not a rankings exercise. Pull ten job postings you would genuinely accept, in your city or target city, and count the clouds. Bengaluru enterprise and GCC roles skew AWS with a strong Azure minority; UK financial services and public sector skew Azure; data-heavy product teams in both markets are where GCP appears. Sit the exam for the cloud that appears most — the "best" certification you cannot map to a real employer is a worse investment than an average one you can. Then build the three projects on that same cloud, so every line of the pairing reinforces the others. Keep the credential current (remember Google's two-year window and Microsoft's annual free renewal), and let the projects age publicly — a repository with a year of commits is better evidence than one created last month.

Budget honestly for the whole pairing: US$150–200 for the exam (₹13,000–17,500 / £115–155 at July 2026 rates), perhaps a Coursera month or two for structured preparation, and £0/₹0 for the projects beyond free-tier cloud usage and your evenings. As total career investments go, it is remarkably cheap; the scarce input is the discipline to ship the unglamorous half of the pairing.

How to present certs on CV and LinkedIn without looking junior

Presentation is where good credentials go to die. The same certificate reads as senior or junior depending entirely on placement and framing.

On the CV: certifications live in a short section near the bottom — name, issuer, year, nothing else. Two entries maximum; list only current, recognised credentials and drop the expired and the trivial (a foundational-level badge under a professional one subtracts, not adds). They must never appear in your headline, your summary, or — the classic junior tell — as a substitute for a projects section. The hierarchy a screener should see is experience, then shipped work, then education, then credentials as a quiet footnote that survives the applicant-tracking keyword pass. Our ATS-proof CV guide covers the mechanics of that keyword layer in detail.

On LinkedIn: put credentials in the dedicated Licences & Certifications section, where recruiter search actually indexes them — and keep them out of your headline. "AI Engineer | 7x Certified" is the single most reliable junior signal on the platform, because the people with the strongest evidence never need it: their headline says what they build. Let the certification appear once, in its structured field, with the verification link attached; spend the headline and About section on shipped work, following the positioning logic in our LinkedIn playbook for recruiter search. The one acceptable moment of cert visibility is the week you earn it — a single post about what you built while studying (not a badge photo with a motivational caption) converts the announcement into evidence.

The rule underneath both: a certification should be discoverable, never load-bearing. If removing it would collapse your professional story, the story needs more work — and the same applies to your GitHub profile, where pinned repositories, not badges, do the talking.

Decision flowchart: your background → cert, course, or portfolio first

Run yourself through this in order; stop at the first branch that matches.

1. Does a gate you face demand a credential? Partner-tier employer, services bench, client contract, cross-border CV. → Cert first, matched to the gate's cloud — then projects immediately after.
2. Can you already build and ship on your own? Working engineer, comfortable in Python, has deployed something. → Portfolio first: three projects, then one cert to document the stack you now demonstrably use.
3. Do you know the theory but not the tooling? Data scientist retooling toward LLMs, academic entering industry. → Course first: Hugging Face or fast.ai to convert knowledge into artefacts, then reassess — often the portfolio emerges from the coursework and the cert becomes optional.
4. Are you starting from zero? Student or career-changer without code confidence. → Course first, portfolio second, cert last — and only once postings you want actually mention it. A credential earned before you can build reads as exactly what it is.
5. Considering a specific employer? Match their stack before booking anything — and if it is a startup, run our startup vetting checklist before you spend a rupee or a pound preparing for their cloud.

Costs approximate as of July 2026; exam fees exclude local taxes.

The last word belongs to the framing we opened with. In 2026's market — post-retirements, post-hype, portfolio-first on both sides of the map — a certification is a fine full stop and a terrible opening sentence. Decide what your evidence says first; then, if a gate in front of you reads badges, buy exactly one.