How much work can AI actually own?
Delegated work will separate the next decade's winners. Autonomous work rate is the share of a workflow AI finishes without a person carrying it.
The paper
The Delegation Frontier.
Contractual makes AI employable. The operating layer for AI-completed work.
By Contractual Research
AI capability is already abundant.
Models draft, analyze, and reason across most desk work. A person still checks the output and enters each case into the business system.
Delegated work compounds. Model quality converges.
Models improve for everyone at the same time. What a company can hand over improves only for that company, and it builds up. Two firms run the same model and finish different shares of the same queue, because only one has written its rules in a form the AI can actually follow.
Six frontier models · 100 cases · four arms · one frozen cohort
Prompting scores 0%. Under Contractual, the work comes back done.
Fig. 1
The Delta Index.1
The surprise
Model intelligence is often not the blocker. Missing context, authority, and reliable acceptance tests are.
Autonomous work rate counts a case only when AI takes it all the way to done, the result checks out in the business system, and nobody had to carry it there.
| Prompting alone | Under Contractual | ||||
|---|---|---|---|---|---|
| Model | A1 alonePrompting | A2 well-prompted | B contract in the prompt | RateContractual | |
| gpt-5.6-terra | 0% | 0% | 0% | 38% | |
| gpt-5.6-luna | 0% | 0% | 0% | 40% | |
| claude-opus-5 | 0% | 0% | 0% | 40% | |
| claude-sonnet-5 | 0% | 0% | 0% | 40% | |
| claude-haiku-4-5 | 0% | 0% | 0% | 40% | |
| claude-fable-5 | 0% | 0% | 0% | 40% | |
Look at the zeros. A prompt cannot promise that anything gets finished, so every model decides the case well and then stops.
- A1 alone
- The prompt someone would actually type. What the task is, what the answer should look like, nothing else.
- No enforcement
- A2 well-prompted
- A careful ops manager's written policy: green, yellow, and red spelled out, with the usual industry limits.
- No enforcement
- B contract in the prompt
- The same operating rules, pasted into the prompt as text.
- No enforcement
- C Contractual
- The same prompt, with every action running through Contractual's runtime.
- Enforced
Delta uses a synthetic 100-case payment-preparation cohort: six models, 100 cases, four arms, live model decisions, an in-memory runtime. 40 of 100 cases qualify for payment preparation, and a person releases funds, so 40% is as high as this cohort can go. Every prompting arm scores zero by definition: autonomous work rate only counts a case that comes with a receipt, and a prompt cannot produce one.
Turn instructions into operating rules.
- 01
Operating rules
Define the work, the exceptions, and what counts as done.
- 02
Green, yellow, red
State which cases AI finishes, which come to a person, and which stay untouched.
- 03
Completion evidence
Confirm the action in the business system and keep the record.
Draw the line case by case.
No two queues hold the same cases. Decide which ones AI finishes, which ones come to your team with the work already gathered, and which ones AI never touches.
Start with one workflow.
Two weeks later you get a decision, measured against work your team already did.
50–100 completed cases2 weeks