Benchmark · AI document work
The read-to-write ratio of AI agents doing document work
Across 252 real sessions of a document agent, 98.6% of every token went to reading — loading files and re-loading context — and only 1.4% to writing the answer. That is a ~69:1 read-to-write ratio, and it is why memory is the cost lever. Estimate it for your own pipeline below — a read:write analyst will then measure the real numbers on your files.
Your read:write analyst
Describe your document pipeline
A memory-less agent re-reads the whole corpus on each pass, and iterates ~2 turns per pass.
The shape of your bill
Without a rulebook in memory, this pipeline re-reads your corpus ~6× before it writes a word.
Estimated read-to-write ratio ≈ 67 : 1 — on a ~2.1M tokens corpus. With your rules held in memory: ≈ 18 : 1.
Measured by your analyst on your own documents — not estimated here.
ExampleEvery written deliverable is stamped PASS / FAIL against the rules you author — so a cheaper read path never means a sloppier answer.
Drop a representative sample of your documents here (optional) — . Nothing leaves your private workspace.
The finding: reading is ~99% of the work
We measured every token the Workhog document agent processed across 252 sessions and 712 successful calls (2026-06-04 → 2026-07-08). Of 788.1M total tokens, 776.8M (98.6%) were reading and only 11.3M (1.4%) were the written output — a 68.6:1 read-to-write ratio. The single largest slice, 81%, was re-reading context the agent had already seen on earlier turns.
Share of all tokens processed
Reading = fresh input + first-read context + re-read (cached) context. Writing = output.
The full token breakdown
Every token category, with raw counts and its share of the total. Download the full dataset as CSV.
| Token category | What it is | Tokens | % of total | Side |
|---|---|---|---|---|
| Cache reads | Re-reading context the agent already saw on earlier turns | 638,581,677 | 81.0% | reading |
| Cache creation | New context read for the first time (files loaded into the window) | 110,284,391 | 14.0% | reading |
| Fresh input | The prompt itself — the user's message for the turn | 27,892,407 | 3.5% | reading |
| Output | The written answer / deliverable the agent produced | 11,327,594 | 1.4% | writing |
| Reading (total) | Everything loaded IN | 776,758,475 | 98.6% | reading |
How we measured it
- A “call” is one turn of the agent — one request to the model and the response it streams back.
- Reading = fresh input + cache-creation (context read for the first time) + cache-read (context re-read on later turns). Writing = output tokens (the answer the agent produced).
- The sample: 712 successful calls across 252 sessions and 4 workspaces, filtered to general document work (
app_context='general',is_error=false), from the Workhog per-call usage ledger, over 2026-06-04 → 2026-07-08.
Why memory is the cost lever
Reading dominates because document work is multi-turn. On every turn the agent reloads the files and the prior conversation into its context window to reason over the whole corpus. The written answer is tiny next to the pile it was distilled from — so the token volume is overwhelmingly re-reading the same corpus, over and over — and re-reading is the largest line on the bill.
That is the opening. If the corpus is read once — indexed into a compact set of facts plus a rulebook the agent carries forward — then every later turn reads only that rulebook and the deltas: the new or changed documents. Read cost stops scaling with turns × corpus size and stays roughly flat as the matter grows. The 81% re-read slice is precisely the part memory removes.
This is the whole argument for a persistent, indexed workspace over a fresh chat thread: in a chat, the corpus is re-read every turn forever; in a workspace with memory, it is read once and the read cost holds flat as the binder grows.
Read this before you quote the number
These are token shares, not dollar shares. Cached re-reads bill at a fraction of the price of fresh input tokens, so reading's share of the dollar cost is lower than 98.6% — though it is still the largest single line. That gap is exactly why caching and memory are cost levers.
Small sample. 4 workspaces (friendly testers, one of them a founder dogfood account). Treat the ratio as a strong directional signal from real usage, not a population law.
Derived estimate. The figures come from our own per-call usage ledger — an internal accounting of each turn's tokens, not a billing export.
Frequently asked questions
What is the read-to-write ratio of an AI agent doing document work?▼
In this benchmark of 252 real sessions (712 successful calls, 2026-06-04 → 2026-07-08), 98.6% of all tokens went to reading and 1.4% to writing — about 69:1. 81% of tokens were re-reading context already seen on earlier turns.
Why do AI agents read so much more than they write?▼
Document work is multi-turn: on every turn the agent reloads the files and the prior conversation into its context window so it can reason over the whole corpus. A written answer is tiny next to the corpus it was distilled from, so the dominant cost is re-reading the same documents repeatedly.
How does persistent memory reduce the cost?▼
Without memory, the agent re-reads the whole corpus on every stage and every turn, so read cost compounds as the matter grows. With memory, it reads the corpus once to build an index and a rulebook, then reads only that rulebook plus the deltas (new or changed documents) on later turns — keeping read cost roughly flat as the corpus grows.
Is the read-to-write ratio the same as the dollar-cost ratio?▼
No. 98.6% is a share of tokens, not dollars. Cached re-reads are billed at a fraction of fresh input tokens, so reading's share of the dollar cost is lower — though still the largest line. That difference is why caching and memory move the bill.
Want the read:write economics of your own matter?
Workhog is an AI agent that works on your files and holds the rules you author in memory. Set your pipeline in the tool above and hand it your documents — the agent measures the real read:write cost on your corpus, names your biggest cost lever, and keeps a running ledger as the binder grows.
Download the benchmark dataset (CSV)