Value Modelling — See The Expert Hours & Dollars Alfrada Saves You

July 14, 2026

Alfrada now models the value of the work it does for you, in the only units that mean anything: expert hours saved and dollars. Open Settings → Account → Hours Saved and you'll see, day by day and conversation by conversation, how many billable hours of a human professional's time your agent replaced, what that time is worth at market rates, and whether you walked away satisfied.

What you can do

• See hours saved, value saved, and a compression multiple across Today / 7d / 30d / 90d windows — totals up top, a per-day breakdown, and a per-conversation ledger underneath.
• Read the per-conversation receipts. Every analyzed conversation-day shows the inferred profession (one of 15 job types, from Software engineering to Legal to Founder & operator), estimated human billable hours, the 50%-flexed credit, measured agent runtime, hours saved, a one-line summary of the work, and a satisfaction chip (satisfied / partial / unsatisfied / unclear) judged from how the day actually ended.
• Edit the money math. Each profession has a default $/hr rate (e.g. Legal $250, Consulting & strategy $180, Software engineering $95). Open Hourly rates, set them to your market, and every figure reprices instantly. Uncategorized work is priced at the median of your effective rates.
• Backfill history on demand. The nightly batch (02:00 UTC by default) analyzes every closed UTC day; the refresh button scans up to 90 days back. Today's numbers are provisional and recompute once the day closes.
• Keep it private and local-first. The analysis is done by your local Gemma (gemma4:e4b) — transcripts are never sent to a cloud model for scoring. Tool inputs/outputs and binary payloads are stripped before analysis; only the prose is read. While your desktop is open it owns the analysis; the cloud only picks up the nightly batch when your machine is away, and results sync both ways.

The framework: task-level value at expert wage rates (Anthropic's method)

The model behind these numbers is deliberately not homegrown. It follows the methodology Anthropic established with the Anthropic Economic Index and refined in Estimating AI productivity gains from Claude conversations — the closest thing the industry has to a standard for pricing AI-performed work:

1. Value the task, not the tokens. Anthropic maps conversations to real occupational tasks (via the O*NET taxonomy) rather than counting messages or tokens. We do the same: each conversation-day is classified into a profession whose work it actually resembles.
2. Price at the wage of the human who would otherwise do it. The Economic Index anchors task value to the hourly wages of the workers who perform that task (BLS wage data). Our per-profession $/hr defaults play that role — and because your market may not be the US average, they're editable.
3. Estimate counterfactual human time, then discount it. Anthropic estimates how long the task would take a human expert without AI, and validates those model-made estimates against ground-truth completion times — finding they need calibration, not blind trust. We adopt the same posture: Gemma is instructed to report pure human billable hours with no AI discount, and the 50% AI-flex handicap is then applied deterministically in code. Claiming half is the conservative stance; the compression multiple you see is computed off the flexed number, not the raw estimate.
4. Charge the AI's own time against the credit. Hours saved is flexed hours − measured agent runtime (from real production telemetry), never the gross estimate.
5. Track success, not just volume. Anthropic's economic primitives include task success as a first-class measurement. Our satisfaction label — judged from the closing exchanges of each day — is that primitive, so inflated hours on work you had to fight with are visible for what they are.
6. Preserve privacy while measuring. Anthropic's analysis runs through Clio, a privacy-preserving pipeline that never exposes raw conversations. Our equivalent is stronger for you personally: the analyst is a local model on your own hardware, and tool I/O never reaches it.

This is the standard to aim at because it prices AI output in units markets already price — expert labor — using a method that has been published, validated against ground truth, and adopted as the reference point by economists studying AI. A token counter tells you what the AI did; this tells you what it was worth.

Where this shows up

• You've been running Alfrada hard for a month and someone asks what it's actually worth. You open Hours Saved, set the 30-day window, and read off a dollar figure grounded in your own conversations — not a vibe.
• A legal-research day and a quick personal errand both happened yesterday. The ledger prices the first at Legal rates and the second near zero as Personal & life, instead of pretending all usage is equal.
• You're a contractor who bills at $140/hr, not the $95 default. You edit one rate and every historical figure reprices.

Try it

This one lives in the app, not the chat box. To open it:

• Go to Settings → Account → Hours Saved to see totals and the per-day / per-conversation ledger. Switch between the Today / 7d / 30d / 90d windows at the top. (When Alfrada has already saved you measurable value, a Hours Saved shortcut also appears on the new-chat screen.)
• Hit Refresh to backfill on demand — it scans up to 90 days back.
• Open Hourly rates, set (for example) Software engineering to $140/hr and Legal to $300/hr, and every figure reprices instantly.

Heads up

• Estimates, conservatively framed. Billable hours are a local model's judgment of your transcripts, then halved by the AI-flex handicap. Treat the output as a defensible lower-bound model, not an invoice.
• Closed days are final; today is provisional. Rows computed mid-day recompute after the day closes (UTC). On-demand refreshes of today are rate-limited to one per 15 minutes.
• Substance gate. Trivial days are skipped — a conversation-day needs roughly 1,500 generated tokens, a 700+ character transcript, or tool activity before it's analyzed.
• Very large days are sampled. Million-token days are analyzed from a head+tail sample of the transcript, with the true day scale passed alongside so the estimate covers the full day, not just the visible sample.