Format review memoApplies the required review-memo structure.
Choose what AI does. Watch users respond.
Which parts should AI do?
A task can be split into components. An organization chooses which components its AI performs, but users still decide whether to work inside that managed workflow or turn to outside AI.
Can giving AI more to do preserve more reviewer-authored work?
The central tension is simple: more AI can bring more assignments inside the workflow, yet replace more human contribution within each assignment.
Managed AI policyGuided review
illustrative modelInside workflow
Outside
Recovered through conversion+Displaced inside−
Net retained human contributionConversion can win
The decision sequence
One policy choice. Two behavioral margins.
The values are an illustration of the paper’s model, not estimates from legal practice.
- 01Split the target task
Contract review is represented by five named components with different consequences.
- 02Choose the AI bundle
The organization specifies which components the managed AI supplies.
- 03Users respond
Each user–task assignment is completed inside the managed workflow or outside it.
- 04Balance two outcomes
Compare current operating value with retained human contribution that may matter later.
Five selectable AI-support components
These components form the paper’s contract-review illustration, not an exhaustive decomposition of legal work.
Locate contract clauseFinds the contract clause under review.
Retrieve firm ruleRetrieves the firm’s governing review rule.
Retrieve precedentRetrieves a relevant prior case.
Draft replacement wordingProposes replacement contract language.
Interactive workflow lab
See the decision unfold
Start with the three-round challenge, then inspect every component bundle in the sandbox.
Round 1 · Conversion versus displacement
Would you expand the managed workflow?
Non-governed completion—using outside AI or another shortcut—is tempting. Choose a component policy for a team reviewing vendor contracts, then run the illustration.
Your policyGuided review
Conversion–displacement ledger
Where does the contribution change come from?
A nested AI expansion affects human contribution through two opposing channels.
Net retained contribution—Outcome appears after deployment
User–task assignment response
Before deployment
Already governedNewly convertedOutside
Inside organizational workflowGoverned completion
Outside the workflowOutside AI or another shortcut
Deploy the policy to reveal assignment movement
43%expected governed share
Representative display: 28 inside, 36 outside. Each dot is a user–task assignment, not an employee or observation.
View detailed model outcomes
Governed shareAssignments completed inside
Clause finder43%
Guided review—
Current valueImmediate operating performance
Clause finder6.56
Guided review—
Human contributionExpected focal contribution retained
Clause finder0.40
Guided review—
Policy objectiveCurrent value + weighted contribution · higher is better
Clause finder8.16
Guided review—
What the demonstration does—and does not—show
Read the model at the right level
Assignments, not employees
Each dot represents one user–task assignment. Movement means a change in completion mode, not hiring, firing, or worker mobility.
Contribution, not a skill stock
The model tracks human contribution retained in the focal workflow. Its capability interpretation requires that this contribution predict later independent performance.
Illustration, not calibration
Parameters reproduce the paper’s contract-review illustration. They are not estimates, legal advice, or a claim that one bundle fits every setting.
Static result summary
Key model findings
| Question | Model answer | Boundary |
|---|---|---|
| Can a larger AI bundle retain more human contribution? | Yes, exactly when contribution recovered through newly governed assignments exceeds displacement among assignments already governed. | The comparison is a nested bundle expansion. |
| What role do user choices play? | They create the conversion benefit. Fixing participation at a common share removes that benefit in the focal comparison. | The fixed counterfactual holds bundle consequences constant. |
| Why can component identity matter? | Components can reach different participation frictions, so two additions can each be valuable alone yet excessive together. | This is an architectural implication of the same implementation mechanism. |
- Map the components. Identify what the managed AI supplies and what remains for users.
- Estimate implementation. Ask how each bundle changes workflow appeal and support for the remaining work.
- Compare both outcomes. Evaluate current task value and retained focal contribution across plausible scenarios.
It is most useful when use is discretionary, components can be supplied separately, and retained human contribution matters. It is not directly prescriptive when use is mandatory, components are inseparable, outside completion retains comparable contribution, or safety, confidentiality, compliance, and implementation costs bind.
PAPER
Research behind the lab
When More AI Preserves Capability:
Component Selection in Human–AI Workflows
The paper studies a product-architecture decision facing organizations that introduce AI into an existing workflow. Components jointly determine both the division of work and whether users implement it.
AuthorStefanos PoulidisINSEAD · Working paper, July 2026
ModelOrganization first, users secondTwo-stage Stackelberg game
Decision objectFinite component bundleAll 32 modeled managed bundles evaluated
Suggested citation: Poulidis, S. (2026), “When More AI Preserves Capability: Component Selection in Human–AI Workflows,” working paper, INSEAD.