How to break into an AI career: a 90-day plan without a technical degree

Two extremes ruin the conversation about AI careers. The first says you need mathematics and a master's degree. The second says a course is enough and the salary arrives by itself. Neither describes the reality in which people move into these roles from support, editing, analytics, sales and teaching.

Reality is simpler and harsher at once. Companies do not need people who can train neural networks from scratch. They need people who can turn a vague request — "we want AI in this part of the process" — into a testable set of requirements, examples and quality criteria. Such specialists are scarce because the work demands discipline of thought and the ability to talk to several teams, not an ML diploma.

Who can get into AI work, and who should not try

It helps to split the entry into three levels. They differ not in prestige but in how much technical background is required on day one.

Entry level	What you need to start	Typical roles
No code	Clear writing, attention to detail, ability to follow a guideline and explain decisions	AI trainer, answer quality rater, data curator, AI editor, AI adoption specialist
Low code	Spreadsheets, regular expressions, JSON, the idea of an API, basic SQL, no-code automation	Model quality analyst, evaluation specialist, data operations, RAG content manager
With code	Python, APIs and vector stores, a working understanding of models and metrics	ML engineer, data engineer, AI application developer, MLOps

The first level is not a junior version of the second. Someone who defines quality criteria and spots contradictions in a guideline decides whether the product is usable at all.

Now the honest part. These roles do not suit people who want work without decisions and conversations: almost every entry role in AI is about alignment. They do not suit people who need a fast and guaranteed result: the market is young, titles drift, and some vacancies are a company experiment. And they do not suit people who tolerate routine badly: rating answers means hundreds of similar tasks while keeping the same bar on the three-hundredth example.

Five roles that really hire without a technical degree

Titles differ: the same tasks may be called "AI trainer", "data annotator", "AI quality analyst" or "AI content specialist". Read the task description, not the headline.

1. AI trainer

Reads request-answer pairs, compares options and finds defects: an invented fact, a broken instruction, the wrong tone, a missed condition. Then writes the reference answer or fixes the guideline so the error does not repeat. The second layer is inventing hard scenarios: where the user is confused, angry or asks for the impossible. Interviews test whether you can justify a choice by criteria rather than "it feels better", and whether you can spot a contradiction in a one-page guideline.

The classic beginner mistake is rewriting an answer to personal taste. Rating calls for the minimal intervention that removes a specific defect. The second failure is inconsistency: you counted something as an error on the first example and not on the fortieth. That is why such teams run calibrations, where raters discuss the same examples and agree on the boundaries.

2. Answer quality evaluator (evals)

Owns the system rather than the examples: criteria, scales, a representative sample and a report on what degraded after a prompt or model change. The role is closer to analytics: solid spreadsheets, simple metrics and clear written conclusions. What is valued is the ability to tell a metric that measures value from a metric that measures reporting comfort.

3. Data and knowledge base curator

When a model is connected to internal documents, answer quality depends not on the model but on what it was given. The curator removes duplicates and contradictions, marks versions and makes sure every answer has a source. The classic failure: two editions of a policy live in the base, the model takes the old one, support quotes outdated rules to a customer. Editors, methodologists and support specialists move into this role well.

4. AI adoption specialist

A role for companies that already bought AI and nobody uses it. The job is to find the steps where the model saves time, run a pilot with ten people, measure the effect and turn the experience into instructions. The key skill is conversational: explain the same thing to an accountant, a manager and a lawyer in three languages and never make them feel replaced. That is why companies usually promote from inside.

5. AI editor and content operator

Owns the output matching the brand, the facts and the task: remove vocabulary the company never uses, check numbers, rewrite a paragraph that sounds like a robot's press release. Half the job is process design: which generation steps repeat, where a human check is needed. A good entry point for journalists, linguists, translators and marketers.

What happened to the "prompt engineer"

The standalone vacancy is rarer than in the first hype wave: prompt writing dissolved into other roles. As a profession it survives where prompts are the product — in teams building agents and multi-step scenarios. The lesson for entry: build your plan around the skill of making a model's task testable, not around a job title.

The technical minimum you need even for a "no-code" role

"No code" does not mean "no technology". Entry interviews test whether you understand how the tool works and where it breaks.

How a model generates an answer

A working mental model is enough: a language model predicts the continuation of text from probabilities learned on a large corpus. It does not look facts up in a database and does not "know" them in the human sense. Hence two consequences you must explain in your own words: why the model confidently invents things, and why the wording of a request changes the result. Plus three terms that almost always come up: token, context window and temperature.

Why models fail

Errors fall into groups: invented fact, outdated information, wrong link, broken format, ignored part of the instruction, answer too generic, dangerous advice. The work starts when you stop saying "the model is dumb" and start classifying: what exactly happened and at which step. Classification turns a complaint into a task.

RAG, metrics and sampling

RAG means the model first retrieves fragments from your documents and then answers grounded in them. That scheme explains half the vacancies: companies need people who will put the documents in order. Separately, learn the difference between human and automated evaluation: the first is expensive and subjective, the second is fast and easily gamed. In practice they are combined, and the key word is "sample": a score on ten convenient examples means nothing.

Data and personal data

The topic that sinks the most candidates, deservedly so. Anything you paste into a third-party chat leaves for someone else's server. Customer personal data, medical records, restricted documents and trade secrets must not go to external services without a legal basis. In Russia this is governed by Federal Law 152-FZ "On Personal Data": processing requires a lawful basis and consent, and storing or processing Russian citizens' data has localisation requirements. Interviewers ask this to see whether you will put the company at risk in your first month.

Tools

Spreadsheets at pivot-table level, JSON as a format, the idea of an API (request, response, key, limits), SQL basics good enough to select and group rows, and one no-code automation tool. That is enough to pass a test task and not look helpless in week one.

A 90-day plan: from reading about AI to two interview invitations

The plan assumes 6-8 hours a week, which fits someone who works and studies in the evening. Over three months that is roughly 80-100 hours. The goal is modest and testable: not "master AI" but get two or three interviews for an entry role.

Period	What you do	Artefact you end with
Weeks 1-2	Learn how the model works and where it fails; pick an industry and a role	Two-page summary; a list of 15 real vacancies with their requirements
Weeks 3-4	Review 5-10 model answers a day: defect, classification, fix	30 test cases with ratings and comments
Weeks 5-6	Build criteria and scales, run your sample through them, count results	A mini evaluation system with a one-page conclusion
Weeks 7-8	Assemble a knowledge base for your industry and test grounded answers	A write-up: what went in, how you chunked it, where it improved
Weeks 9-10	Run a training adoption case: process, hypothesis, pilot, metric	A two-page case with before and after numbers
Weeks 11-12	Package the portfolio, rewrite the CV, apply and rehearse interviews	A public portfolio page, a CV, three rehearsals

The first two weeks are about language, not career: learn to tell a prompt from a guideline, RAG from fine-tuning, a hallucination from a data error. In parallel, pick an industry: "AI in general" does not sell at entry level, while someone who understands how e-commerce support actually works looks stronger. Your previous experience is an advantage here, not ballast.

Then come the artefacts. Thirty reviewed answers are already a portfolio because they show how you think. An evaluation system turns scattered opinions into a measurable value. The knowledge base gives you a second artefact for the curator role. A case with before and after numbers proves you can carry an idea to a measurable result, and the section on data you refused to send to a model adds trust.

The last two weeks are for selling, not learning: a public page with four artefacts, a CV tuned to one role, five applications a week and a rehearsed conversation. Applications without rehearsal turn into practice on real people, and you do not have many of them.

A portfolio instead of experience

Entry roles are rarely won with a diploma or a certificate — they are won with evidence of thinking. The four artefacts from the plan are the portfolio.

1. A set of test cases. Thirty rows: request, what the model was supposed to do, what it did, defect type, severity, fix. Bad: "the answer is bad". Good: "format broken: JSON with three keys was required, the model returned prose".
2. An error analysis and a guideline fix. How you found it, why it happens, what you changed, what improved and what stayed broken. The last point is mandatory: it shows you are not selling magic.
3. A mini evaluation system. Criteria, scales, sample, results table and a half-page conclusion. Conclusions must follow from criteria and actions from conclusions. "Quality should improve" is not a conclusion.
4. An adoption case with a metric. Process before, hypothesis, constraints, before and after results, and a section on data you deliberately did not send to the model. Even a training case works: reviewers check how you reason, not the scale.

One public page is enough: who you are, which domain you know, four blocks with links, a contact. Do not collect a hundred prompts, post chat screenshots without analysis, or show only successful examples. None of that answers the real question: can this person make quality decisions when nobody is around to advise.

CV, vacancies and the conversation about switching

The most common mistake is writing a CV "broadly, to fit everywhere". In AI that fails: roles differ, and hiring managers look for a match in tasks, not in enthusiasm.

Search for a group of titles, not one: AI trainer, AI quality analyst, quality rater, AI content editor, AI adoption specialist, knowledge manager, knowledge base curator. Add your industry to the query — "AI support", "AI healthcare", "AI banking" — and you will find vacancies where your background is an advantage. Separate must-haves (understanding criteria, accuracy, tolerance for routine) from nice-to-haves (specific tools you can learn in a week).

Write experience as task, action, result: not "handled customer requests" but "handled 40 tickets a day, built a library of 60 standard answers, halved response time on the most frequent question". A separate "AI projects" section holds the four artefacts with links and a result line each. The cover letter is four short paragraphs: the role and one concrete detail about the company, what you already have with a portfolio link, how your background transfers, and what you are ready to do as a test task.

The interview for an entry AI role

Theory is rare. Usually you get material and the interviewer watches how you reason. Three tasks come up most often.

Rate two model answers. You get a request and two options and must pick the better one. What is tested is not taste but the ability to build criteria: name what was required (facts, format, tone, completeness), compare on each point and explain what the rejected option loses. "The first reads better" loses to "the first is factually accurate and keeps the format, the second has a better tone but the sum is wrong, which is critical for a refund — so I pick the first".

Find the defect in a guideline. You get a page of rules. Typical defects: two clauses contradict each other, a criterion nobody can verify, a missing exception, a rule so broad it forbids normal cases. A guideline two people apply differently destroys the whole evaluation, which is why this skill beats writing pretty prompts.

A mini adoption case. "Support gets five hundred tickets a day and we want faster replies. What do you suggest?" A good answer starts not with "connect a model" but with "which requests are most frequent and what is the cost of an error in each type". Then hypothesis, metric, constraints, pilot plan.

Structure any answer as thesis, argument, example, conclusion. First the short direct answer, then the reason, then a case from your portfolio, then a one-sentence conclusion. And do ask clarifying questions: "just to check, do you mean factual accuracy or format compliance" signals that you will ask such questions on the job instead of guessing.

Ask the hiring manager three things: what success looks like in the first three months, who decides in disputed cases, and what hurts most right now. Keep the red flags in mind: annotation without criteria or calibrations, "do everything related to AI" without a team or metrics, a promise of growth into ML engineering without a single engineer on staff, and paying for training before employment.

Rehearse the conversation in advance. Improvising at your first live interview is a bad strategy. Say your answers out loud, then run a full dialogue with an AI coach: it asks questions, you answer, and afterwards you read a debrief of where the answer sagged and where an example was missing.

How long it takes and what it costs

A realistic timeline without a technical background is three to six months at 6-8 hours a week: three months for skills and portfolio, another one or two for applications and test tasks. "Career change in three weeks" belongs to a different genre.

Money is barely needed at the start: model documentation, open company guides, free tiers and public evaluation material cover most of the programme. It makes sense to pay for a model subscription if the free limit is too small and for a one-off consultation with someone already in the role. You do not need a "profession of the future" course with a job guarantee or a "prompt engineer certificate": entry roles look at artefacts and at how you speak.

And patience. The first twenty applications may bring no reply — the market is young and many vacancies are filled internally. Track conversations, not rejections: if ten interviews leave you unsure what is missing, it is time to collect feedback and rework the portfolio.

Risks nobody mentions in the courses

• Some roles are temporary. Mass annotation and simple answer rating are being automated and getting cheaper. Roles that hold up require judgement, domain expertise and responsibility for decisions.
• Automated evaluation changes the job. Models increasingly act as a first filter while humans review disputed cases. Less routine, but you must be able to audit the audit.
• There is an entry ceiling. Repetitive annotation by someone else's checklist hits a ceiling within two years. A good entry role has criteria, calibrations and a visible growth path.
• The "do more" pressure. AI speeds work up and that quickly becomes an expectation to do twice as much for the same money. Ask in the interview how the team measures workload.
• Info-products around the topic. Guaranteed employment, ad salaries without a source, testimonials only as screenshots, "two seats left" — all signs you are being sold air. Ask to see three portfolios of graduates who actually work in the role.

A first-week checklist

1. Pick an industry you already know and one entry role.
2. Write down fifteen real vacancies and underline the repeating requirements.
3. Explain out loud, in your own words, what a token, a context window and temperature are.
4. List ten types of model errors you have already met.
5. Review five answers using the pattern request, expectation, actual, defect, fix.
6. Create a test-case spreadsheet and add the first five rows.
7. Write down which data at your work must never go to an external service.
8. Look at how an open knowledge base in your industry is organised.
9. Draft one sentence about yourself: "I know X and want to own the quality of AI answers in X".
10. Say the switching story out loud and note the questions that stumped you.

Questions people ask most

Can I get into an AI career without a technical degree?

Yes, in entry roles: AI trainer, answer quality rater, knowledge base curator, AI adoption specialist, AI editor. No machine learning degree is required, but technical literacy is: how a model generates an answer, where it fails, what context, RAG and quality metrics mean.

Do I need to learn Python?

For no-code roles, no: spreadsheets, JSON and the idea of an API are enough. Python becomes relevant if you move towards automated evaluation, data engineering or building AI applications.

How long does the switch take?

Three to six months at 6-8 hours a week: three months for skills and portfolio, another one or two for applications, test tasks and search.

What matters more: a certificate or a portfolio?

A portfolio. Entry roles look at your answer reviews, your criteria system and your adoption case, because they show how you decide. A certificate can help pass an initial filter, but the conversation will be about practice.

Which role should I pick first?

The one where your previous experience becomes an advantage: support and sales move into AI training and adoption, editors and translators into AI editing and knowledge curation, analysts into quality evaluation.

How do I tell a good vacancy from a dead end?

There are written quality criteria, there are calibrations and a person who decides disputed cases, and growth within a year is visible. Annotation by a checklist without described processes and without a team leads to a ceiling.

What next

Start with one action today: write down five model answers that annoyed you and, next to each, what exactly was wrong. In a week you will have a defect list, in a month a thirty-row table, in three months a portfolio you can take to an interview.

If you want the full picture first, read the map of new AI jobs: it covers how the product development cycle changes and which roles appeared at each step. For the conversational part, use interview preparation and the piece on soft skills, and to rehearse a live dialogue, open the Coach mode in HintMeTalk AI.