Guglielmo Paesano AI engineer and founder of AIGROW

I build AI systems that do people's work, and I measure where they still need people.

contact@guglielmo.ioDiscord, 4K members

Can AI do your work? I test it and publish what happens.

Send a task from your job. I rebuild it with invented documents, give it to 12 current AI models from seven makers, three times each, and publish every answer: the right ones and the wrong ones.

Send a task to testHow a test is run

work tasks tested so far
6
current models, from seven makers
12
answers marked, every one published
216
answers with an error a person would have to catch
5

The tests so far

Each test is one ordinary office task, built with invented documents that hide the traps a hurried reader falls into. A model passes a run only when every check passes, and it counts as reliable on a task only when it was right three times out of three. All of them were run on 16 November 2026.

  1. Can AI check an expense claim against the company policy?

    A surveyor hands in ten expenses after a week on the road. Seven of them cannot be paid as claimed, each for a different rule in a one-page policy.

    A finance assistant’s task. Finance.

    All twelve models, every time36 of 36 answers passed every check

    No answer failed a check.

    Read the test

  2. Can AI compare supplier quotes and pick the cheapest?

    A bakery needs 1,150 printed cake boxes and has three quotes on the table. The quote with the lowest unit price is the most expensive one to accept.

    A bakery owner’s task. Purchasing.

    All twelve models, every time36 of 36 answers passed every check

    No answer failed a check.

    Read the test

  3. Can AI find the cancellation deadline in a contract?

    A dental practice leases a sterilisation unit on a three-year agreement that renews by itself. Miss the notice date by a day and the practice pays for another year.

    A practice manager’s task. Contracts.

    All twelve models, every time36 of 36 answers passed every check

    No answer failed a check.

    Read the test

  4. Can AI reconcile a bank statement?

    A print shop closes October with a cash book and a bank statement that disagree by 951.38 pounds. The difference is made of six ordinary things, one of them the bank taking a direct debit twice.

    A bookkeeper’s task. Bookkeeping.

    11 of 12 models, every time35 of 36 answers passed every check

    Most often wrong, in 1 of 36 answers: Spots the cheque paid in on the 30th and the cheque written on the 31st, neither at the bank yet.

    Wrong at least once: GPT-6 Luna.

    Read the test

  5. Can AI shortlist job applicants against the requirements?

    Seven people apply for a warehouse team leader job with five stated requirements. Two meet all five. A wrong shortlist either wastes an interview or turns away someone who qualifies.

    A warehouse manager’s task. Hiring.

    All twelve models, every time36 of 36 answers passed every check

    No answer failed a check.

    Read the test

  6. Can AI write a staff rota that follows the rules?

    A cafe has eight people, fourteen shifts and nine rules. A rota that breaks one of them is a shift nobody can open, or a 17-year-old working past the hour the law allows.

    A cafe manager’s task. Scheduling.

    10 of 12 models, every time32 of 36 answers passed every check

    Most often wrong, in 4 of 36 answers: Lets nobody open the morning after they closed.

    Wrong at least once: GPT-6 Luna, Gemini 4 Argon.

    Read the test

How a test is run

Every test follows the same four steps, and each step leaves something you can read: the task, the documents, the checks and every model’s result. The models get one message and no tools, the way a person pastes a job into a chat window. Nothing is marked by opinion, mine or a model’s.

  1. A real task, invented documents

    The task comes from a job someone does. I write new documents for it, with invented names and figures, so nothing private is published and no model has seen them before.

  2. Checks written before the run

    For each task I write down what a right answer must contain, as rules a program applies. The rules are tried on a known right answer and on known wrong ones before any model is asked.

  3. Every model, three times

    Each model gets the same message three times, with its maker's default settings. Asking again matters: a model that is right twice and wrong once cannot be left alone with the work.

  4. Everything published

    The page of each test prints the task, the documents, the result of every run and one wrong answer word for word. The answers of the whole run are one file anyone can download.

What these tests do not show

A test like this answers one narrow question: given this task and these documents in one message, does the model return a result a person could use without correcting it? It says nothing about a whole job, and it is true for the day it was run. These are its limits, stated plainly.

  • One message, no tools. The models work from one message, as in a chat window. A system built for the job, with tools and its own checks, can do better. These tests show the starting point.
  • Text only. The documents are given as text. Reading a scan or a photograph is a different test.
  • Three runs is a small sample. It shows whether a model is consistent on a task. It does not measure its exact error rate.
  • Invented documents. They are written to be like real ones, traps included. They are not your documents.
  • A date, not a verdict. Models are replaced every few weeks. Each page says the day its test was run.
  • Open data. The answers of the whole run are one file, under CC BY 4.0.

Send a task from your job

Describe something you do at work and what a right result looks like. I read every task myself. If it can be tested, it gets a page here like the ones above.

  1. You describe the task. No documents, and nothing confidential.
  2. I rebuild it with invented names and figures.
  3. Every current model does it three times.
  4. The result is published here and I send you the link.
The human check appears here.

Free. Your address is used to answer you and for nothing else. Not every task can be tested, and I tell you when yours cannot.

Questions about the tests

Can AI do my job?
A job is many tasks, and the answer is different for each of them. These tests take one task at a time. In November 2026, of six tasks tested, every model was right every time on four; on the others at least one model made an error a person would have had to catch. Send the task you care about and I will test it.
Which AI models are tested?
The twelve models of the November 2026 run are Claude Fable 5.2, Claude Sonnet 5.5, Claude Haiku 5.5, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Luna, Gemini 4 Argon, Gemini 4 Hydrogen, Grok 4.7, DeepSeek V4.1 Flash, Mistral Large 4 and Qwen4 Max. They are current models from seven makers, a flagship and a smaller model where the maker sells both, each asked through its API with the maker's default settings.
How do you decide whether an answer is right?
By rules written before the run. Each task has a list of checks, such as a date, a total or a name that has to be exactly right, and a program applies them to every answer. An answer passes only when every check passes. No person and no model marks an answer by opinion.
What happens when a test itself is wrong?
I read every failed answer before a run is published. If a careful person could have given that answer from the text alone, the fault is the test's: I fix the wording, set the task's answers aside and ask again. An answer my own runner cut short is asked again too. In the November 2026 run that set aside 89 answers, and none of them is counted.
Can I send a task that involves confidential documents?
Describe the task and leave the documents out. I rebuild every case with invented names, companies and figures before any model sees it, so nothing you send is published and nothing private reaches a model.
Does it cost anything?
No. Sending a task is free and so is reading the results. I pay for the runs: the November 2026 run cost $2.71 for 216 answers.
Can I quote or reuse the results?
Yes. The results and the answers are published under the Creative Commons Attribution 4.0 licence: use them with a link to this page. The file with every answer of the run, word for word, is linked from each test.