{"course":{"slug":"ai-careers","title":"Working in AI","url":"https://addaly.com/learn/ai-careers","level":"beginner","track":"work"},"lesson":{"slug":"the-roles-that-exist","title":"The jobs, described as they feel on a Tuesday","url":"https://addaly.com/learn/ai-careers/the-roles-that-exist","position":1,"of":85,"minutes":7,"answer":"Most AI work is moving messy data and checking outputs, not training models.","keyIdea":"Most AI work is moving messy data and checking outputs, not training models."},"body":{"format":"markdown","text":"## Titles lie, work does not\n\n\"AI Engineer\" means at least six different jobs depending on who wrote the posting. Read the responsibilities section, ignore the title. Here are the roles that actually exist, described as they feel on an ordinary Tuesday rather than in a brochure.\n\n## Data engineer\n\nYou move data from where it is created to where someone can use it. Pipelines, schemas, backfills, an alert at 03:40 because last night's job died on a row containing a comma.\n\nA Tuesday: a dashboard shows Monday's orders as zero. You trace it to a payments provider that renamed a field from `order_id` to `orderId` on Saturday without telling anyone. You fix the parser, replay two days of data, and write a note so the next person loses ten minutes instead of three hours.\n\nThere are more of these jobs than any other kind on this list, in every country. Least discussed, most stable.\n\n## Data analyst\n\nSQL, spreadsheets, dashboards, and the question behind the question. Someone asks for \"signups by region\". What they are actually deciding is whether to keep paying for the Nairobi campaign. Most of the skill is finding that out before you write the query, because the query takes four minutes and the wrong question wastes a week.\n\nThis is the most common first job in the field for people who did not study computer science.\n\n## ML engineer\n\nYou put models into products and keep them alive: serving, latency, retraining, monitoring, rolling back when the new version is quietly worse. Training models is a slice of it. Training one from scratch is rare outside research labs.\n\nA Tuesday: recommendations got 400ms slower on Friday and nobody knows why. You find a feature lookup making one database call per item instead of one per request.\n\n## AI product manager\n\nYou decide what gets built and what \"good enough\" means, which in this field means writing down how the system will be measured before it exists. Also: arguing with legal about which data may be used, cutting scope, and telling people no in a way they can accept.\n\n## Prompt and evaluation work\n\nThe newest and least settled corner. The durable half is not writing prompts, it is evaluation: assembling the test cases that decide whether a system ships, grading outputs, hunting failure modes, turning \"this feels off\" into a number someone can act on. Job titles here change every eighteen months. The work does not.\n\n## Research scientist\n\nReading papers, running experiments that mostly fail, writing up the few that do not. A small number of jobs, concentrated in a few dozen labs and universities worldwide, nearly all wanting a PhD. Keep it on your map. Do not plan your life around it.\n\n## The ratio nobody prints\n\nPicture a 300-person fintech in São Paulo with a genuine data team: about four data engineers, six analysts, two ML engineers, one product person who covers AI alongside three other things, and zero researchers. Multiply that shape across thousands of companies and you have the actual job market.\n\n> Figure: One real data team, at a 300-person fintech. Bars (people): Data analysts 6; Data engineers 4; ML engineers 2; Product person covering AI, alongside three other things 1; Research scientists 0. Thirteen people and nobody training models from scratch. Multiply this shape across thousands of companies and it is the job market: analyst and data engineer are the two wide doors.\n\nIf you are choosing where to aim, aim where the jobs are and where you can produce evidence fastest. Analyst and data engineer are the two widest doors. ML engineering is far easier to reach from a data engineering job than from a course.\n\n## What they all share\n\nEvery role above is mostly one activity: something is wrong with the data and you have to find out why. Courses skip this because it does not demo well. Employers pay for it because it is where the time goes."},"check":{"question":"A 300-person logistics company in Jakarta hires its first AI person to build an assistant that answers staff questions about shipments. Three months in, where has most of that person's time gone?","options":[{"text":"Fine-tuning a language model on the company's shipment history.","correct":false,"tempting":"Assumes the hard part of AI work is training models, because that is what courses spend their weeks on."},{"text":"Getting shipment data out of four systems that disagree with each other, and building a set of test questions to check the answers against.","correct":true},{"text":"Comparing model providers and negotiating a contract with one of them.","correct":false,"tempting":"Treats the job as vendor selection: a real task, but one that takes days, not months."},{"text":"Rewriting the prompt until the answers read well.","correct":false,"tempting":"Mistakes the visible surface of the system for its main cost, because the prompt is the part you can watch changing."}]},"prev":null,"next":{"slug":"who-actually-employs-you","title":"The eight kinds of employer, and how many of each","url":"https://addaly.com/learn/ai-careers/who-actually-employs-you"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'The jobs, described as they feel on a Tuesday', in Working in AI. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-careers/the-roles-that-exist (Accessed: 31 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"The jobs, described as they feel on a Tuesday","courseTitle":"Working in AI","module":{"index":1,"slug":"the-terrain","title":"The terrain: who hires, for what, and where you fit"},"answer":null,"keyIdea":"Most AI work is moving messy data and checking outputs, not training models.","hero":{"kind":"figure","figure":{"t":"figure","label":"Figure 1.1","figure":{"type":"bars","unit":"people","title":"One real data team, at a 300-person fintech","series":[{"label":"Data analysts","value":6,"highlight":true},{"label":"Data engineers","value":4,"highlight":true},{"label":"ML engineers","value":2},{"label":"Product person covering AI, alongside three other things","value":1},{"label":"Research scientists","value":0}],"caption":"Thirteen people and nobody training models from scratch. Multiply this shape across thousands of companies and it is the job market: analyst and data engineer are the two wide doors."},"alt":"Figure: One real data team, at a 300-person fintech. Bars (people): Data analysts 6; Data engineers 4; ML engineers 2; Product person covering AI, alongside three other things 1; Research scientists 0. Thirteen people and nobody training models from scratch. Multiply this shape across thousands of companies and it is the job market: analyst and data engineer are the two wide doors."}},"blocks":[{"t":"h2","id":"titles-lie-work-does-not","runs":[{"text":"Titles lie, work does not"}]},{"t":"p","runs":[{"text":"“AI Engineer” means at least six different jobs depending on who wrote the posting. Read the responsibilities section, ignore the title. Here are the roles that actually exist, described as they feel on an ordinary Tuesday rather than in a brochure."}]},{"t":"h2","id":"data-engineer","runs":[{"text":"Data engineer"}]},{"t":"p","runs":[{"text":"You move data from where it is created to where someone can use it. Pipelines, schemas, backfills, an alert at 03:40 because last night’s job died on a row containing a comma."}]},{"t":"p","runs":[{"text":"A Tuesday: a dashboard shows Monday’s orders as zero. You trace it to a payments provider that renamed a field from "},{"text":"order_id","code":true},{"text":" to "},{"text":"orderId","code":true},{"text":" on Saturday without telling anyone. You fix the parser, replay two days of data, and write a note so the next person loses ten minutes instead of three hours."}]},{"t":"p","runs":[{"text":"There are more of these jobs than any other kind on this list, in every country. Least discussed, most stable."}]},{"t":"h2","id":"data-analyst","runs":[{"text":"Data analyst"}]},{"t":"p","runs":[{"text":"SQL, spreadsheets, dashboards, and the question behind the question. Someone asks for “signups by region”. What they are actually deciding is whether to keep paying for the Nairobi campaign. Most of the skill is finding that out before you write the query, because the query takes four minutes and the wrong question wastes a week."}]},{"t":"p","runs":[{"text":"This is the most common first job in the field for people who did not study computer science."}]},{"t":"h2","id":"ml-engineer","runs":[{"text":"ML engineer"}]},{"t":"p","runs":[{"text":"You put models into products and keep them alive: serving, latency, retraining, monitoring, rolling back when the new version is quietly worse. Training models is a slice of it. Training one from scratch is rare outside research labs."}]},{"t":"p","runs":[{"text":"A Tuesday: recommendations got 400ms slower on Friday and nobody knows why. You find a feature lookup making one database call per item instead of one per request."}]},{"t":"h2","id":"ai-product-manager","runs":[{"text":"AI product manager"}]},{"t":"p","runs":[{"text":"You decide what gets built and what “good enough” means, which in this field means writing down how the system will be measured before it exists. Also: arguing with legal about which data may be used, cutting scope, and telling people no in a way they can accept."}]},{"t":"h2","id":"prompt-and-evaluation-work","runs":[{"text":"Prompt and evaluation work"}]},{"t":"p","runs":[{"text":"The newest and least settled corner. The durable half is not writing prompts, it is evaluation: assembling the test cases that decide whether a system ships, grading outputs, hunting failure modes, turning “this feels off” into a number someone can act on. Job titles here change every eighteen months. The work does not."}]},{"t":"h2","id":"research-scientist","runs":[{"text":"Research scientist"}]},{"t":"p","runs":[{"text":"Reading papers, running experiments that mostly fail, writing up the few that do not. A small number of jobs, concentrated in a few dozen labs and universities worldwide, nearly all wanting a PhD. Keep it on your map. Do not plan your life around it."}]},{"t":"h2","id":"the-ratio-nobody-prints","runs":[{"text":"The ratio nobody prints"}]},{"t":"p","runs":[{"text":"Picture a 300-person fintech in São Paulo with a genuine data team: about four data engineers, six analysts, two ML engineers, one product person who covers AI alongside three other things, and zero researchers. Multiply that shape across thousands of companies and you have the actual job market."}]},{"t":"p","runs":[{"text":"If you are choosing where to aim, aim where the jobs are and where you can produce evidence fastest. Analyst and data engineer are the two widest doors. ML engineering is far easier to reach from a data engineering job than from a course."}]},{"t":"h2","id":"what-they-all-share","runs":[{"text":"What they all share"}]},{"t":"p","runs":[{"text":"Every role above is mostly one activity: something is wrong with the data and you have to find out why. Courses skip this because it does not demo well. Employers pay for it because it is where the time goes."}]}],"check":{"question":[{"text":"A 300-person logistics company in Jakarta hires its first AI person to build an assistant that answers staff questions about shipments. Three months in, where has most of that person’s time gone?"}],"options":[{"text":[{"text":"Fine-tuning a language model on the company’s shipment history."}],"correct":false,"tempting":[{"text":"Assumes the hard part of AI work is training models, because that is what courses spend their weeks on."}]},{"text":[{"text":"Getting shipment data out of four systems that disagree with each other, and building a set of test questions to check the answers against."}],"correct":true,"tempting":null},{"text":[{"text":"Comparing model providers and negotiating a contract with one of them."}],"correct":false,"tempting":[{"text":"Treats the job as vendor selection: a real task, but one that takes days, not months."}]},{"text":[{"text":"Rewriting the prompt until the answers read well."}],"correct":false,"tempting":[{"text":"Mistakes the visible surface of the system for its main cost, because the prompt is the part you can watch changing."}]}],"order":[2,1,3,0],"why":null},"prev":null,"next":{"slug":"who-actually-employs-you","title":"The eight kinds of employer, and how many of each","minutes":8}}}