Skip to content
← Projects

Demo · demo volumes, synthetic data

AI value report

An agent reads, sorts and files the documents that reach an accounting firm. Each month this report answers the question that matters once it is live: is it worth it? At 3,000 documents a month it gives the team back 54.5 hours, worth €1,428 at Spanish labour cost, for €0.79 of model. The setup pays back in 2 months, and the first year leaves €12,827 net.

Demo: demo volume and synthetic data. Quality and cost measured in 6 recorded runs of the own evaluation; hours at INE labour cost, dollars to euros at the ECB rate of 26/10/2026, Anthropic’s prices.

The problem

AI success stories usually work out savings on other countries’ wages. The model costs the same everywhere; hours don’t. This report values every hour at the labour cost per hour actually worked that INE publishes for Spain: €26.18 (ETCL, Q2 2026).

Hours of work a monthBy hand 75.0 h; with the agent 20.5 h.Hours of work a monthBy hand3,000 × 1.5 min75.0 hWith the agent12.7% by hand + a check on the rest20.5 h54.5 h freed every month.Source: Own evaluation: 6 runs of the document-intakeagent, 26 Oct 2026; method under “One month, in numbers”

What it measures

  • Quality and cost, measured. 6 real runs with Claude Haiku 5.5 and Claude Sonnet 5.5 on 55 synthetic documents: 642 model calls and $0.88 in total. In every run, 54 of 55 documents were read and routed right. This is the own evaluation of the document-intake agent.
  • Time and money, derived. Hours saved, their value, the model cost in euros (ECB rate on the day of the runs), the monthly net and the months to pay back the setup.
  • Every input in plain sight. INE’s hourly cost, the ECB rate, Anthropic’s prices, maxcorti.com’s indicative prices and the method estimates, each with its source and rationale. The client can change the volume, the hourly cost and the model, and the report page recalculates everything.
One month in eurosHours saved €1,428, model €0.79, running €150, net €1,277.One month in eurosHours saved54.5 h × €26.18/h€1,428Model costmeasured per document€0.79Running and improvementindicative price€150Net a month€1,277The model cost barely shows on the chart.Source: Own evaluation (6 runs); INE, ETCL Q2 2026;indicative prices on maxcorti.com

One month, in numbersIndicative

  • 54.5 hours saved: 75 hours by hand against 20.5 with the agent.
  • €1,428 of value at €26.18/h.
  • €0.79 of model a month: 0.06% of the value it saves.
  • After the running fee (€150/month), €1,277 net a month. The €2,500 setup pays back in 2 months.

The maths is conservative: whatever the agent doesn’t file (12.7% in the runs) counts as full manual work, and every filed document adds a 15-second check.

Cumulative position over 12 monthsStarts at −€2,500 and adds €1,277 a month.Cumulative position over 12 months€0−€2,500€12,827024681012monthspaid back in 2.0 monthsSource: indicative prices on maxcorti.com; monthly net fromthe chart above

What decides whether it pays

The model cost barely shows on the chart. What decides is people’s time and how much review work is left. That is why the report measures, every month, the documents that still go through a person.

How it’s built

The report reads the evaluation logs, checks that they agree with each other (calls, tokens and cost) and refuses any data that didn’t come from a real model. The charts come from the same code on the page and in the SVG files. It calls no model.

AI value report architectureCodeEvaluation of the document-intake agent (read-only, fixedversion)Campaignledger: calls,tokens anddollars642 recordedmodelresponsesPer-documentresultsAggregate andcross-checkcalls, tokensand costPublic inputs:INE, ECB,Anthropic,maxcorti.comMethodestimates:minutes perdocument,volumeValue maths: hours saved, € saved, model cost,paybackReport page, ES / EN:recalculates in thebrowserStandalone charts for thecase page and posts
Test data is refused: only responses from a real model count.

With a real firm

The demo documents are synthetic and the volume is a demo value. With a client, minutes per document are measured with their team and the report is recalculated with their figures.

Tell me about a process that eats your time.

In 30 minutes we’ll see whether AI can help. No commitment.