The best AI assistant for Fantasy Premier League: what to look for in 2026

What an AI FPL assistant actually has to do

The phrase "AI assistant for Fantasy Premier League" now covers three very different products. The first is a chatbot wrapper: a large language model with a friendly persona that talks confidently about players but has no live data and no maths behind it. The second is a spreadsheet with a web front end: real numbers, usually a single expected-points column, with no sense of how uncertain that number is. The third — the only category worth paying for — is a simulation engine with a conversational layer on top, where the assistant can explain a projection because it can read the model that produced it.

The test is simple. Ask the tool why a player is ranked where he is. A wrapper will produce plausible prose. A spreadsheet will produce nothing. A real engine will tell you the player's start probability, his expected minutes, the fixture-adjusted attacking rates behind the projection, and the range of outcomes the simulation produced.

Minutes are the whole game

Every Fantasy Premier League point is conditional on a player being on the pitch. A striker with elite per-90 numbers and a 55 per cent chance of starting is worth less than a competent one who plays every minute, and no amount of expected-goals sophistication rescues a tool that gets the team sheet wrong.

A serious model therefore starts with a minutes distribution per player rather than a binary starts-or-doesn't flag: probability of starting, probability of appearing at all, and the spread of minutes given an appearance. It then reconciles those probabilities at team level, because exactly eleven players start and exactly 990 outfield-plus-keeper minutes are available. Models that treat each player's start probability independently routinely project a squad with fourteen starters, which quietly inflates everyone.

One number is not an answer

Expected points is an average. A player projected at 5.0 points will frequently return 2 and occasionally return 14. Two players with identical 5.0 projections can have completely different risk profiles: a defender on a clean-sheet-heavy path with a hard floor, and a rotation-risk forward whose average is carried entirely by a small chance of a hat-trick.

That distinction decides the armband. If you lead your mini-league, you want the highest floor. If you are chasing, you want the highest ceiling. An assistant that only reports the mean cannot tell you which is which, so insist on a floor and ceiling — typically the 10th and 90th percentile of the simulated outcomes — alongside the average.

Does it publish its own errors?

Any tool can look clever in preview. The question is what it says after the gameweek. A credible engine backtests itself automatically: mean absolute error overall and by position, minutes error, start-prediction accuracy, and captain efficiency — the recommended captain's actual haul divided by the best possible haul that week.

Crucially, those numbers should be published for every completed gameweek, including the bad ones. A tool that only surfaces its wins is marketing, not modelling. FPLHERMES exposes its full audit in the Model Lab page for exactly this reason.

A practical checklist

• —Minutes modelled as a distribution, reconciled to eleven starters per team
• —Availability scaled continuously from official flags and injury reporting, not just a warning icon
• —Fixture-by-fixture simulation rather than a season-average multiplier
• —A floor and a ceiling reported next to every expected-points figure
• —Separate safe and ceiling captaincy shortlists
• —A transfer planner that respects your current squad, bank and free transfers instead of rebuilding from scratch
• —Mini-league logic based on effective ownership inside your league, not global ownership
• —A published, per-gameweek accuracy record including losses
• —An explainable interface: you can ask why, and get the actual inputs back

Where FPLHERMES sits

FPLHERMES was built around that checklist. It simulates every upcoming fixture 500 times, sampling minutes, team goals from a Poisson process driven by attack and defence strength, then allocating goals, assists, saves, cards and bonus points inside each simulated match. The average is the expected-points figure; the spread produces the standard deviation, floor and ceiling.

The conversational layer is not decoration. HERMES has read access to the same tables the engine writes to, so when you ask why a player is ranked seventh it answers from the model's own numbers rather than from general football knowledge.