AI Agents for PhD Students: The Research Jobs Worth Automating

A doctorate is a long project made of short jobs that keep coming back. The search you re-run every few weeks. The folder of PDFs you meant to read in March. The reference list that drifted out of one style and halfway into another. The progress note you write at 11pm the night before supervision.

None of that is the hard part. The hard part is the argument, and the argument stays yours for the whole whole programme. What surrounds it is admin with a research accent, and it arrives on a cycle.

That cycle is the entire point of this post. I run an AI agent platform, so the question people bring me is which parts of their work a machine can carry. For research work the answer comes down to frequency. How often does this job come back?

A chat window is good at a task you bring it. An agent is for the task that arrives whether you bring it or not.

Key takeaways

• Recurrence decides. A one-off question belongs in a chat window. A job that returns every week or every month for years is agent-shaped, and the scheduling is most of the value.
• The automatable jobs are the dull ones: new-paper monitoring on a saved query, reading-pile triage, cross-paper note synthesis, reference hygiene, journal reformatting, deadline tracking, weekly progress logs.
• Every automated job needs a human checkpoint. An agent that hands you twelve papers has not told you which three matter, and it cannot tell you a reference is wrong, only that it looks wrong.
• The boundary is not negotiable. An agent does not write your thesis, your chapters or your assignments, and it does not produce text you submit as your own. Disclose what your institution requires, and read that policy yourself.
• Cost is not the obstacle. Your first agent on Gravity is free with no card. Gravity is in private alpha, so you apply rather than sign up.

Why a chat window is not enough for a doctorate

The difference that matters is the trigger. You open a chat window. An agent opens itself. Across a project measured in years that stops being a convenience, because the expensive part of a recurring job is rarely the doing. It is the remembering.

Price out a weekly literature check honestly. There is the time to re-run the search, assuming you still have the string you used last time. There is the longer job of working out which hits you have already seen. Then the part nobody budgets for: the weeks you skip because you were writing, and the drift you find when you come back to it.

An agent is a standing instruction with a schedule attached. You describe the job once in a sentence, you say when it should run and what it should hand back, and then it runs without you in the room. What an AI agent actually is covers the mechanics, and AI agents versus workflow automation covers why this is not the same as wiring a flow together yourself.

Which is why much AI-for-academia advice lands badly on a research programme. It is written around the single dramatic task, the one where you paste in a paper and ask what it says. That task is real, and a chat window already handles it. The jobs that quietly cost you most are duller and they are periodic.

I would rather put a number here about how many PhD candidates use AI in their research. I went looking on 27 September 2026 and did not find one I trust enough to print, so I am not printing one. What I can show you is our own search data: Gravity's page on agents for students drew 1,697 impressions and 53 clicks at an average position of 5.0 in the 28 days to 25 September 2026, for a phrase that US keyword tools price at roughly 20 searches a month. The public volume figures badly under-read what this audience is looking for.

The research jobs worth automating

Seven jobs, and they share one property: each of them happens again. The third column is the one to read carefully, because every row leaves something behind that is yours.

This table describes the jobs, not any one product's connectors. Before you rely on a tool for the first two rows in particular, check what it can actually reach, because coverage of databases, repositories and paywalled journals differs between tools and I have not tested that for you.

Start with the reading pile, because the arithmetic there is most in your favour. Most of what a search returns gets discarded. In a sample of 195 medical systematic reviews registered on PROSPERO, the median project ran 65.8 weeks from registered start to publication, and the median review ended up including under 1% of what its de-duplicated search returned (Borah et al., BMJ Open, 2017). Those are medical systematic reviews from a 2017 analysis, not doctorates in general, so take the shape of the finding rather than the figure: a large majority of screening work exists in order to produce a no.

Screening to a no is mechanical, and mechanical is the test. A ranked pile with two lines of reasoning per item decides nothing for you. It changes what you open first, and on a pile that size that is most of the win.

Reference hygiene is the least interesting row on the table and the one I would automate second. Nobody has ever enjoyed discovering that a batch of entries lost their DOIs and others carry the wrong year. A check that runs every fortnight and reports what looks broken converts an end-of-project panic into a small recurring chore. The verification stays with you, because a checker sees the shape of an entry and not the paper behind it.

Journal reformatting deserves its place because of when it hits you. A desk rejection arrives, you pick the next venue, and the manuscript needs different headings, a different reference format and a different section order. The writing does not change. The packaging does. That is the cleanest job on the list to hand over, and it is the one people are most surprised to find there.

Weekly progress logging looks least like automation and pays off longest. An agent that asks you on Friday what moved, files the answer and keeps the running log gives you two things at once: a supervision agenda you did not assemble at the last minute, and a record of a year you will otherwise remember as a blur.

What must stay yours

An agent does not write your thesis. It does not write your chapters, your assignments or your abstract, and it does not produce text that you submit as your own work. I am saying that flatly because this category is has plenty of pages that imply the opposite while being careful never to write it down.

Four things are not delegable, and none of them are close calls.

• The argument. What you are claiming, and why it follows from what you found. That is the doctorate. Everything else in this post is logistics arranged around it.
• Interpretation. A summary of a paper is not a reading of it, and the judgement about what a result means for your question is precisely the thing you are being examined on.
• The claims you put your name to. If an agent hands you a sentence you could not defend in a viva, it does not go in. A source you have not opened is not a source you have.
• The prose. Writing is how you find out whether you understand something.

Then disclosure, which is where this gets decided in practice. Institutions differ from each other, and I do not know how current any given policy is. I do not know what yours says, this post cannot know, and any page that assures you it is fine is guessing on your behalf. Read your own institution's policy on AI use and disclosure before you automate anything that touches a submitted document, and disclose whatever that policy requires you to disclose.

The practical version is shorter than the policy version. Keep a note of which jobs you handed over. If you can tell your supervisor in one sentence exactly what the agent did and what you did, you are almost certainly on the right side of the line. If that sentence is hard to say out loud, you have your answer.

One thing I will not help with. People search for an AI that will produce their university assignments, and some of them search for an AI that will disguise the fact afterwards. Not here. That is not what agents are for, and it is not a job I am interested in serving.

Literature review needs more room than a section

Screening a literature review properly is its own workflow, and compressing it into four paragraphs here would do it badly. Inclusion and exclusion criteria, de-duplication, the extraction table, the audit trail that lets somebody else repeat what you did: each of those carries a decision that changes the result.

So it gets its own post. Using an AI agent for a literature review works through where the automation helps and where it quietly corrupts the method. Read that before you point anything at a screening step.

For the wider tool landscape, AI research assistants is our category page. I am naming no products in this post on purpose. I have not run the research tools in this space against each other, and repeating their feature lists back to you would pretend that I had.

Setting one up

Pick one job. Not a system, one job, and preferably the one that made you sigh most recently.

Choose the job you have already skipped twice

The best candidate is something you know you should do weekly and do not. Skipped work is the clearest signal that a job needs a trigger. If you cannot choose between two of them, I would take new-paper monitoring, because it fails silently: nothing breaks, and you notice the gap months later when a paper you should have read in spring turns up in somebody else's related work.

Write the schedule into the sentence

Describe the job the way you would describe it to a new research assistant, and put three things in the description: what to look at, how often, and what a useful result looks like. Here is a complete instruction: every Monday, check for new papers matching this search, and tell me what is new and what changed since last week. On Gravity you describe that in plain words instead of building a workflow, and the recurrence is part of the description rather than something you configure afterwards.

Keep a checkpoint that you personally read

Every automated research job needs a place where a human looks, and it should sit where a mistake is still cheap. Read the ranked pile before you cite anything out of it. Read the reformatted manuscript before you submit it. Confirm a deadline at its source before you trust a reminder about it. An agent running unwatched for a term can be wrong about something, and the question is only whether you find out early or expensively.

Start on the free agent before you spend anything

Your first agent on Gravity is free and takes no card, which is enough to find out whether the job you picked is genuinely agent-shaped. After that, Pro is $5 a month during the alpha, and you can buy more usage if you run out. Max is $20 a month for more agents and more usage. Current plans are listed at gravity.fast/pricing. A run that fails on a platform error, with no usable output, is refunded under the refund policy.

No agent is going to finish a doctorate. It can be finished a good deal less chaotically by somebody who stopped hand-carrying the parts of it that repeat.

FAQ

Sources