AI Reviewer, Human Statistician or Doing It Yourself: The Real Cost of Each

You have a manuscript nearly closed, a date on the calendar and a budget that does not stretch to everything. In one tab, an automated reviewer promising a report in minutes. In another, an email from a statistician with a three-figure quote and a question back that you still have not answered. In the third, a twelve-part tutorial telling you this is a weekend's work. All three are sold as the same decision at three different prices, and they are not: they are three different jobs that happen to land on your desk in the same week. What follows separates what each one does, what each really costs — in money, in calendar time and in risk — and when each is the correct answer, including the situations where the correct answer is not my tool.

The three are not competing for the same job

A manuscript fails in three separate places, for three reasons that have nothing to do with each other. It fails through inconsistency: the p that does not match its degrees of freedom, the N in the Method that is not the one the table's df imply, the d that does not follow from the t printed beside it, the reference that does not exist. It fails through decision: the test is not the one your design calls for, the violated assumption was sidestepped in a way that will not hold, the causal claim is not something a cross-sectional design can carry. And it fails through competence: you did not know that had to be checked, so you did not check it.

Those three failures have three remedies and they are not interchangeable. Inconsistency is arithmetic: it is verified with code, and a mechanical check does not get tired at number eighty-seven. A decision has no closed answer; it depends on the literature in your field, on what you are going to claim, and on what you are prepared to defend when asked. Competence is not for sale in either of the other two: you either build it or you rent it every single time.

It also helps not to confuse categories, because the market mixes them. Paperpal and Writefull are language tools, and they say so themselves: Paperpal describes grammar checking, plagiarism checking, citation generation and checking, paraphrasing, and a battery of pre-submission format checks against journal requirements; Writefull describes academic language feedback, paraphrasing at several levels, and title and abstract generators. Neither advertises statistical verification or checking that your references exist in the databases (checked in September 2026).

And they do their own job well, which is not nothing: a manuscript with weak English dies on the editor's desk before anyone looks at a single number. They simply are not competing with this. A language tool and a statistical verification tool can be run on the same file on the same day without either getting in the other's way.

Route 1: the machine pass

In money. Single-digit euros per report, one-off payment, no prior meeting and no quote. In calendar time. Minutes: you upload the file and you have it. In risk. The lowest of the three for your wallet and the most deceptive for your judgment, for the reason below.

What it actually does is worth stating precisely, because “AI reviewer” sounds like a model reading your manuscript and forming opinions. The example report shows the mechanics: p values are recomputed from the test statistic and the degrees of freedom with code, not with a model; effect sizes are checked against the F or the t you wrote yourself; and every reference is looked up one by one in eight databases (Crossref, OpenAlex, Semantic Scholar, PubMed, DOAJ, arXiv and OpenLibrary). The prose of the report is written by a model, but around numbers it cannot change.

Part of that work is free and it costs little to say so. statcheck reads an article in PDF, DOCX or HTML, finds the tests written in APA format and recomputes every p from its statistic and its degrees of freedom; there is an R package and there is a web version, with no account and no payment. If that check is all you need, use it and save the rest. Its limits are equally clear: it only sees what is in canonical APA format in the body text, it does not go into regression tables, it does not check effect sizes, and it knows nothing about your bibliography.

The real risk on this route is misreading the result. A clean report does not say “this is right”, it says “I found nothing inconsistent in what you wrote”. If you ran the wrong test, the p is perfectly consistent with the wrong test and the report comes back without a single flag. And one sentence should not be dressed up: this is not peer review, it is the author's self-assessment of their own manuscript before submitting. It does not replace the editor, it does not replace the reviewers, and it accredits nothing to anybody.

Route 2: the human statistician

In money. Hundreds to thousands of euros, and there are public figures to orient yourself by. The large author-services companies sell this part separately: Editage lists it as “Statistical Review”, described as help with statistical design and inferences, with no price on the page; Enago lists it as “Statistical Analysis” and does publish fixed fees, with three options: $638 for methods and techniques, $765 for a statistical recommendation and $1,360 for the full analysis, delivered in 7, 8 and 15 days (checked in September 2026). An independent consultant almost always works by quote, because the price follows the scope. The order of magnitude is that one: hundreds for a bounded review of what you already did, thousands for a full commission of design, analysis and write-up.

In calendar time. The number that matters here is not the hours of work, it is the round trip. You send the data, five questions come back, you take three days to answer them because one of them forces you to go and check how you coded a variable, and so on. Counting in weeks is realistic; counting in days is how calls get missed.

In risk. The risk is not that the analysis comes out wrong, it is picking the wrong person to hand it to. The most reliable signal is the inverse of the one people look for: be wary of anyone who gives you a plan and a price without having asked you anything about the design, and trust the one who starts by sending back uncomfortable questions. Thesis mills versus a statistical consultant lays out the difference with examples, and how to choose a statistician for a doctoral thesis goes through the criteria one by one.

What only a person does is decide, and then stand behind it. A machine can flag that you did not report the sphericity test; whether the violation matters, and what to do instead, is a judgment that depends on your design, your N and what you are going to conclude. And there is a second half that gets forgotten when comparing prices: when Reviewer 2 knocks down your analysis, somebody has to write the two paragraphs that defend it, or sit with you the week before the viva. That part is not downloadable.

Route 3: doing it yourself

In money. Zero, or whatever the license costs if you do not already have R or Python in front of you. In calendar time. This is where the whole bill sits, and the honest number is worth stating. Learning to report a t test, an ANOVA or a multiple regression properly, with its assumptions and its effect sizes, is days of real work, not an afternoon. Learning a multilevel model or a structural equation model properly is months, and the first one you do alone you will do twice: once badly and once well.

In risk. The highest of the three, for a reason that more studying of what you already know to study does not fix. The error that costs you the rejection is rarely the one you were unsure about — you go and look that one up — it is the one that never occurred to you to check. A seasoned reviewer spots in two minutes that you did not verify something, and the problem is that you cannot search for what you do not know exists.

Even so, there are three conditions under which this is the correct answer rather than the cheap one: the design is standard and documented to the point of tedium, you have plenty of time before the deadline, and you are going to repeat this analysis. A first-year doctoral student who will run the same design five more times, or a group that has settled into a stable line of work, is buying something other than the saving: they are buying not depending on anyone next time.

There is also the mixed route, which usually works out cheapest and which almost nobody proposes: you do the whole thing yourself and pay only for a review of what you did. Someone going through your syntax, your assumptions and your table for a while costs a fraction of the full commission, and gives you back exactly what you could not see from inside it. The orders of magnitude for each are in how much statistical analysis for a thesis costs.

Where my tool does not replace a statistician

This is worth writing down without hedging, because it is where people who buy a report thinking they bought something else go wrong. The reviewer checks the consistency of what you wrote; it does not choose your model. If the model is the wrong one, everything else can come back spotless and the manuscript is just as dead.

Nested designs. Pupils within classrooms, patients within therapists, measurements within people. Whether the nesting has to be modeled, with what random-effects structure, and what to do when the model will not converge, are decisions, not checks. Structural equation models. Which alternative model has to be fitted for the fit to demonstrate anything, and what you conclude when two different specifications fit equally well. Measurement invariance. When partial invariance is enough to compare means across groups, and when you have stopped comparing the same thing.

Missing data. The mechanism is not dictated by the percentage; choosing between multiple imputation, full-information maximum likelihood and complete-case analysis depends on why the data are missing, which you know and a methodologist argues. Survival and irregular repeated measures. What counts as the event, how censoring works, and what to do with those who drop out. Small clinical samples. Where the power is what it is and the real decision is what can be claimed with that N.

The rule that separates the two is simple: anything a reviewer is going to ask “why did you do it that way?” about is human territory, because the answer is a justification and not a computation. If you are in any of those cases, the machine pass is still worth running at the end for what it is worth — figures that add up and references that exist — but the decision that comes before has to be made with someone. That is what I do in statistical consulting, and it is why both services live on the same site instead of cannibalising each other.

The cost nobody quotes you: when you decide

The same three routes cost wildly different amounts depending on when you take them, and that variation is larger than the difference between the routes themselves. A design decision costs a one-hour conversation before you collect data. After collection, that same decision is no longer for sale at any price: all that is left is making the most of what exists, and making the most of a crooked design is more expensive and slower than doing it straight, on top of producing a worse result.

The typical case arrives like this: sample already collected, insufficient power for the effect being chased, and a request to “compute post-hoc power to justify the N”. Power computed afterwards from the observed effect is a transformation of the p value and justifies nothing; a reviewer who knows that will flag it. What can be done at that point — reframing the question, reporting the interval, being honest about the precision — is a statistician's work, and it is repair work.

The machine's moment is the opposite: at the end, with the manuscript written. Before that there is nothing to check, because there are no written numbers that can contradict each other yet. Running it while you are still deciding on the analysis will give you nothing useful; running it the week before you submit will. That order — person at the start, you in the middle, machine at the end — is the cheapest overall, and it is, oddly enough, the reverse of the one almost everybody follows.

The short rule for choosing

Before paying for anything, say your question out loud and look at how it starts. If it starts with “is what I did written up correctly?” — do the figures add up, do the references exist, is the APA format right — that is a mechanical check, and paying a person to do it by hand is throwing money away. If it starts with “did I do the right thing?” — the test, the assumption, the model, what I am allowed to conclude — no machine will answer that in a form you can defend, because the answer is an argument. If it starts with “do I want to be able to do this myself next time?”, then the time you are about to spend is not a cost, it is the thing you are buying.

And they are not mutually exclusive. The combination that works for most people is the same across almost every project: you do what you already know how to do, you pay for a short consultation at the moment a design decision you do not command shows up, and you run the manuscript through the machine the week before submitting, when the text exists and you can still fix things without redoing anything.

If after all this you still cannot see which one is yours, the most practical signal is what you have least of: short on money and long on time, route three; long on money and short on time, route two; short on peace of mind before hitting submit, route one, which is what it is there for.

Frequently asked questions

Can an AI reviewer replace a statistician?

For the checking part, yes, and with an advantage: recomputing hundreds of p values from their degrees of freedom and looking up every reference in eight databases is work a person does worse, slower and more expensively. For the decision part, no. Choosing the test your design calls for, deciding what to do about a violated assumption, or justifying why you modeled the nesting one way rather than another are judgments you have to be able to defend to a reviewer, and an automated check does not generate arguments, it generates flags.

How much does a statistician cost for a thesis or a paper?

It depends on scope and on timing. The few public figures out there give you the range: Enago publishes fees for its statistics service, $638 to $1,360 depending on whether you want a recommendation or a full analysis, and Editage lists its own with no price (checked in September 2026); most independent consultants work by quote because scope drives everything. In practice it runs from a short review consultation, the cheapest thing you can commission, to a full package of design, analysis and write-up. What is worth knowing before asking for a quote is where you are in the process: the same work costs considerably more if the data are already collected under a design that has to be patched.

Does an automated report count as peer review?

No. It is the author's self-assessment of their own manuscript before submitting. No independent reviewer signs it, no journal backs it, and it does not belong in your cover letter as a credential. It is there so you reach submission with the figures adding up and the bibliography checked, which is exactly the kind of failure that wastes time at the first editorial screen.

Is it ethical to use an AI tool to review my own manuscript?

Reviewing what you wrote is not the same problem as generating the content. Checking figures and references is the equivalent of running a spell checker: it does not change what you did, it points at where you wrote it up wrongly. What does need care is the journal's policy on AI use, which varies a fair amount between publishers today, and not uploading identifiable participant data to any tool. The data-processing agreement, the named sub-processors and the fact that the manuscript is not stored are published on the site precisely because that is the reasonable question to ask.

What if the report flags something I do not understand?

That is the most useful of the three situations, even though it does not feel like it. A flag you do not understand is a competence gap located by name, which is infinitely better than a vague one. At that point you have all three options again, but with the question already narrowed: look it up yourself, which is now feasible because you know what to look for; ask it in a short consultation, which is cheap when the question is specific; or decide that the point does not affect your conclusion and leave it, which is also a legitimate decision if you can justify it.

Is there anything free that does part of this work?

Yes, and it is worth trying before paying for anything. statcheck recomputes p values from the test statistic and degrees of freedom in a PDF, DOCX or HTML file, has an R package and a web version, and costs nothing. It covers one part of the problem — tests written in APA format in the body text — and leaves out regression tables, effect sizes and the bibliography. If your manuscript is straightforward and your results run in the text, that may be all you need.

Keep the sentence that orders the decision: the machine checks what you wrote, the person decides what you did, and you buy the ability to stop asking. Choosing badly is usually not choosing the wrong route, it is choosing it late — a design decision is worth a conversation before you collect data and cannot be bought afterwards. If what you have in front of you is precisely that decision — which model your design calls for, what to do about an assumption that does not hold, how to defend what you have already done — that is what my statistical consulting is for.