AI trained on your data to make split second decisions.

TrueUp learns from your own records, then makes the everyday calls for you: whether a bill is right, what a job should cost, which records belong together. Every answer shows the rows it came from, so nothing is made up and nothing needs rechecking.

💸 Charged twice📦 Billed but never delivered🏷️ Price higher than agreed 🔢 Wrong quantity🧮 Totals that don't add up

🔒 Free account · nothing to install · your files stay private to your team

supplier_statement.csv ⇄ receiving_log.csv

152 of 164 charges matched a delivery. 25 things need a look:

Billed twice

$10,464.00 — same item and amount as row 22

supplier_statement.csv · row 30

Never delivered

$8,879.76 — no delivery matches this charge

supplier_statement.csv · row 76

Price changed

$292.57 a unit billed vs $270.90 ordered

statement row 131 ⇄ receiving row 126

Short shipment

Billed for 54, received 48

statement row 107 ⇄ receiving row 99

These are real TrueUp results on a supplier's statement and receiving log. Try TrueUp on sample files after a free sign-up.

How it works

Trained on your data. Deciding in a fraction of a millisecond. Better with every outcome.

TrueUp isn't a chatbot. It's a decision engine you train on your own records, then call every time a decision needs making.

Trains on your dataGive it your past records: statements, invoices, finished jobs, matches you've already made. TrueUp learns how your business decides, which clues matter and how much, instead of guessing from the internet.

Decides in under a millisecondA trained model answers without writing a word: 0.14 ms per decision, about 6,800 decisions a second. Through the API, the whole round trip is well under a second.

Gets better on its ownReinforcement learning loops score every outcome and feed it back. Good calls are reinforced, and misses and your corrections are learned from, so each decision beats the last.

Five decision formats, one APIOr plug in any AI assistant through MCP.

DecideThe next move, chosen from the options you allow. Learns from every outcome./v1/agents/:name/decide

EstimateA number: what a job should cost, from your past jobs./v1/estimate

MatchWhich records are the same thing, across two lists./v1/match

ReconcileDoes the bill agree with what arrived, and what's owed back./v1/reconcile

AuditIs this document right: totals, math and missing pieces./v1/audit

Cheaper than Jev. Faster and 1,000× cheaper than an LLM.

Cost of one million decisions

TrueUp$5

Jevabout $21

Claude Fable 5.1about $5,500

Time to an answer

TrueUp0.27 s

Claude Fable 5.16.45 s

Costs for a short 500-token question with a 10-token answer at published API prices: Jev $0.042 per million input tokens, output free; Claude Fable 5.1 $10 in and $50 out per million. Longer prompts cost them more. TrueUp: Starter plan rate, $4 on Pro. TrueUp time is the median full API round trip over the internet (the engine itself decides in 0.14 ms, over 17,262 decisions in 20 games of Maze Chase). Claude Fable 5.1: time to the first answer token at its fastest setting, as measured by Artificial Analysis.

100%

of planted bad charges caught on held-out test data, with no false alarms

99%

of documents it flags really are wrong (FinBalance, a public accounting benchmark)

2.7%

average error pricing 20 new jobs from past estimates

14 s

to match 10,000 × 10,000 records on an ordinary laptop CPU

Measured on data TrueUp never saw while it was being built. How we measured, and every limit →

Using TrueUp

Three steps. No setup, no rules to write.

You don't tell TrueUp which column is which or what counts as a match. It works that out from your files — then shows its work.

1

Add your filesSpreadsheet exports (CSV), PDF statements and invoices, text documents, or a zip of folders. Or start with the samples.

2

TrueUp learns themIt figures out what each file is, picks the checks that fit, and lines the records up — usually in seconds.

3

See what's off, and askEvery finding in plain English. Ask “what was billed but never delivered?” and get the rows.

📍

Every answer shows its sourceEach finding names the file and the row it came from, so you can check it in seconds.

🧮

The math has to add upTotals are checked to the cent. If the numbers don't balance, TrueUp stops rather than guess.

🙋

Unsure goes to youWhen TrueUp isn't sure, it says so and asks you — it never quietly decides for you.

🔒

Your files stay putIt runs on your computer. Nothing is uploaded, and there's no account to create.

Try it in a minute.

Start with the sample files: a supplier statement, a receiving log, invoices and bank statements with mistakes hidden in them.

Watch TrueUp learn, live.

TrueUp plays Maze Chase, a retro maze game, and gets better with every game. See each decision, how fast it made it, and the score climb.

Watch it play →

Use cases

What people use TrueUp for

Anywhere two records should agree — or a document's numbers should add up — and checking by hand takes too long. Each card says what it's been tested on and where it falls short.

Get started

Three steps, about a minute

Use it in the app, or let your AI assistant drive it for you.

1

Add the files that should agree — spreadsheet exports (CSV or TSV), PDF statements and invoices, JSON, plain-text documents, or a ZIP of them. Or start with the sample files.

2

Let it learn — TrueUp proposes the checks that fit your files and works out which columns and rows line up. No rules to write, no examples to label.

3

Read what doesn't match — every finding comes with the row it came from, and anything TrueUp isn't sure about is set aside for you. Save what it learned and check next month's files in about a second.

Create a free account →

Connect your own software: create an API key in your dashboard, then send TrueUp two files and read every finding.

curl https://trueup-cloud.merchantprotocol.workers.dev/v1/reconcile \
  -H "Authorization: Bearer $TRUEUP_API_KEY" \
  -F left=@statement.csv \
  -F right=@receiving.csv

The same checks as the app, metered on your plan. API docs →

Documentation

Everything, explained

What TrueUp is for, how to use it, how to connect your AI assistant, and exactly how well it works. Read it right here.

Loading…

FAQ

Questions people ask

Where do my files go?

Into your team's private storage in TrueUp's hosted service, where only your team can see them. TrueUp doesn't send them to any AI provider. You can delete a file or a whole project whenever you like.

Do I need an AI subscription, API keys, or a powerful computer?

No. TrueUp runs on our servers, so any browser will do, and there's no AI subscription to buy. API keys are only for connecting your own software. Matching a list of 2,500 against 17,500 takes about 25 seconds.

What kinds of files can I use?

Spreadsheet exports as CSV or TSV, PDF statements and invoices, JSON, plain-text documents (TXT, Markdown), and ZIP files of any of those — a zip of folders works for a batch of documents.

What about Excel files, PDFs and scanned paper?

PDFs, yes: the statements and invoices you download from your bank, accounting system or a supplier's portal, and scanned ones too. TrueUp rebuilds each table from the page, and every finding names the page and line it came from. A scan is read by OCR and its rows are checked by their own arithmetic (running balances, quantity × price), and TrueUp tells you about any that don't add up. A downloaded PDF is still best: nothing has to be read from an image. Handwriting isn't read. Excel isn't read directly: choose File → Save As → CSV.

How accurate is it?

On held-out tests it caught every planted bad charge with no false alarms; on FinBalance, a public accounting benchmark, 99% of what it flags is really wrong and it catches 75% of the problems; matching real product catalogs it scores 0.68–0.91 out of 1 (F1) without being given a single labeled example. Every number, and how to rerun it →

What happens when TrueUp isn't sure?

It says so. Uncertain findings are marked unsure and listed separately for you to review — it never quietly picks an answer for you. In your project, TrueUp asks you “are these the same?” one question at a time, and the next time it learns it uses your answers.

Do I have to set it up or teach it first?

No labeled examples and no rules. It works out which columns mean what, which rows belong together and how your documents add up from the files themselves. Every time it learns, it saves what it learned as a small model file you can reuse.

Can I train it once and use it every month?

Yes. Learn from files you trust, then check next month's files with that saved model — in your project (Check new files). Nothing is learned again, it's faster, and it works on batches too small to learn from on their own: a single invoice, a handful of rows, one new job request.

How is this different from Excel or my accounting software?

Excel lookups need identical values; TrueUp matches records that are written differently (“GYP BD TYPE X 5/8 4X12” and “5/8 in type X fire rated drywall”), with dates a few days apart. Accounting software records transactions; TrueUp checks them against each other and tells you why something looks wrong. Use it alongside the tools you already have.

Does it fix the mistakes for me?

No — it finds them and shows you the proof so you can decide: dispute a charge, chase a delivery, correct a record. It doesn't change your files or your books.

Can one payment cover several invoices?

Not yet. Matching is one-to-one: a payment that covers several invoices (or an invoice paid in parts) is flagged for you to look at, not split automatically.

Can it work with my other tools?

Yes — through MCP: any AI assistant that supports it connects to your team with an API key, then adds files, runs the checks and reads the findings. Your own software can use the TrueUp API the same way. Set it up → For anything else, contact Merchant Protocol.

Can my AI assistant use it?

Yes. TrueUp is an MCP server: connect your assistant with an API key and it can add files, run a reconcile, match or audit, and read the findings. Set it up →

What does it cost?

Free to start: the free plan needs no card. Paid plans add more analyses and API calls — see pricing.

Find out what doesn't match — in about a minute.

Drop in two files that should agree, or start with the samples. You'll see every difference and the row it came from.

AI trained on your data to make split second decisions.

Trained on your data. Deciding in a fraction of a millisecond. Better with every outcome.

Three steps. No setup, no rules to write.

Try it in a minute.

Watch TrueUp learn, live.

What people use TrueUp for

Three steps, about a minute

Everything, explained

Questions people ask

Find out what doesn't match — in about a minute.

What would you like to check?

Add your files

Here's what I'll check

Pick a public dataset

All set — here's what I found

Every check, in detail

Ask anything about your files

Check new files with a saved model

1 Pick a saved model

2 Add the new files