How to Delegate Work to an AI Teammate

Delegate a responsibility, not a vague outcome

To delegate work to an AI teammate, define six things: the continuing responsibility, relevant context, expected output, quality standard, cadence, and decisions requiring human review. Then test the assignment on real work and improve the standing instructions from observed failures.

Do not start with "help me raise money," "find investors," or "manage this deal." Those are goals containing many different jobs and authority levels. Give the teammate one responsibility that can produce a visible artifact.

A strong delegation brief answers:

1. What do you own?
2. What may you use?
3. What should you bring back?
4. What makes the work good?
5. When should it happen?
6. What still needs me?

This guide provides a copyable template and two worked examples for private-capital teams.

The six-part teammate brief

1. Responsibility

Name a continuing job with a clear boundary.

Weak:

Help with our fundraise.

Stronger:

Maintain the review queue for active LP follow-up.

The stronger version does not assume the teammate owns the strategy, relationship, or outcome. It owns the preparation of a particular operating view.

2. Relevant context

Specify the sources or source classes that belong in the job. Include exclusions when they matter.

Example:

Use active LP records, available relationship history, approved meeting notes, open diligence requests, and current official public sources. Do not treat an old portfolio entry or a person's job title as proof of the institution's current mandate.

"Use everything" is rarely a good context instruction. It increases noise and makes permission boundaries difficult to inspect.

3. Desired output

Define the artifact, not simply the topic.

Example:

Return one table with LP, current stage, last meaningful interaction, open question, promised material, responsible owner, supporting source, and proposed next action.

A defined shape makes the work easier to evaluate, compare, and improve.

4. Quality standard

Explain what makes the output useful.

Example:

Be concise. Separate verified facts, user-provided facts, and inferences. Include the observation date for public information. Flag contradictions and missing sources. Do not create a recommendation from fit alone.

Quality is more than tone. It includes evidence, freshness, uncertainty, and decision relevance.

5. Cadence or trigger

State when the responsibility should run.

Examples:

• every Monday before the fundraising review;
• when an LP meeting is added to the calendar;
• when the user assigns a new manager brief;
• after an identified investor submits a question;
• only when manually requested.

Cadence can be manual. A teammate can own a continuing responsibility without running automatically.

6. Human review boundary

Name the decisions and external actions the teammate does not own.

Example:

Do not send messages, promise materials, characterize allocation interest, change the pipeline stage, or infer permission for an introduction. Prepare the evidence and draft for the GP to review.

This is the most important part of the brief when the work affects relationships, disclosures, investments, or commitments.

Copyable AI teammate job-description template

Role name:

Responsibility:
You are responsible for [bounded, continuing job].

Relevant context:
Use [permitted records, documents, communications, tools, and public sources].
Do not use or infer [important exclusions].

Output:
Bring back [specific artifact and fields].
Return it in [conversation, document, table, or supported work surface].

Quality standard:
The work is ready when it [evidence, freshness, completeness, voice, and uncertainty requirements].
If [common failure or missing evidence], mark it as blocked or unverified.

Cadence or trigger:
Do this [schedule, event, or manual assignment].

Human review:
Ask me before [decision, write-back, external message, sharing, or other consequential action].
The person responsible for [relationship, investment, legal, fiduciary, or disclosure decision] remains [named role].

Feedback rule:
Treat corrections about [lasting operating preference] as proposed instruction changes.
Treat corrections about [one artifact] as edits to that work only.

This template is a starting point, not a guarantee. The actual product must enforce the stated access and approval controls. Instructions alone are not a security boundary.

Worked example 1: an LP prospecting teammate

The vague request

Find LPs for our fund.

This request leaves critical decisions undefined:

• Which fund strategy and geography?
• Which allocator types are relevant?
• What counts as current evidence?
• Should the agent search public sources, licensed sources, existing CRM records, or all three?
• How should duplicates and existing relationships be handled?
• Is the output a research list or an outreach queue?

The delegated responsibility

Responsibility: Maintain a research shortlist of institutions and family offices that may fit Fund III's stated strategy. This is a research responsibility, not authority to add an LP to the active pipeline or contact anyone.

Context: Use the approved fund summary, documented allocator criteria, current official public sources, supported connected data, and existing Finta CRM records. Exclude institutions whose public eligibility clearly conflicts with the strategy. Check for aliases and existing records before treating a prospect as new.

Output: Return a maximum of 15 prospects per review. For each, show organization, allocator type, fit rationale, official source, observation date, known restriction, existing CRM status, available relationship evidence, and unresolved question.

Quality: Include a prospect only when a current source supports the relevant mandate or program. Mark "not publicly verified" instead of filling a gap. Do not estimate assets, check size, available capital, or likelihood of investment.

Cadence: Review the shortlist every two weeks or when the GP changes the criteria.

Human review: The GP decides fit, pipeline entry, relationship strategy, and outreach. A relationship path is not consent for an introduction. Nothing is sent from this assignment.

The improved brief converts a broad goal into a testable job. Read AI Prospecting Agents for the full operating model.

Worked example 2: a real estate investor-question teammate

The vague request

Handle questions from investors.

This could accidentally imply authority over disclosures, underwriting claims, suitability, or external communications.

The delegated responsibility

Responsibility: Organize investor questions about the current offering and prepare evidence-linked draft responses for sponsor review.

Context: Use the approved offering materials, current data-room documents, the investor's actual question, prior approved answers, and the CRM relationship record. Do not use draft underwriting scenarios or superseded documents.

Output: For each question, return the question verbatim, relevant approved source, source date or version, draft answer, unresolved issue, responsible internal owner, and recommended review step.

Quality: Quote figures exactly from the approved source. If two documents conflict, stop and identify the conflict. Do not infer tax, legal, performance, liquidity, or suitability conclusions.

Cadence: Run when a supported investor question is assigned or before the weekly open-question review.

Human review: The sponsor or designated professional reviews disclosure and decides whether and how to respond. The teammate does not send the message, update offering terms, or represent that a commitment exists.

The agent now owns coordination. The sponsor retains the relationship and disclosure decision.

Use the SOURCE check before assigning context

Before adding a source to a teammate's job, apply the SOURCE check:

• Scope: Does this source belong to the assigned responsibility?
• Ownership: Who controls it, and who is accountable for its accuracy?
• Use permission: Is the source authorized for this user, workspace, and purpose?
• Recency: When was it created, updated, or observed?
• Conflict: What should happen if it disagrees with another source?
• Evidence: Can the teammate point back to the relevant record or passage?

This prevents "more context" from becoming an excuse for unclear provenance.

Test the assignment before scheduling it

Run the responsibility manually on three representative cases:

1. Normal case: the expected sources are available and agree.
2. Missing case: a required document or fact is absent.
3. Conflict case: two sources disagree or the public information appears stale.

Review whether the teammate:

• chose the relevant sources;
• followed exclusions;
• returned the requested artifact;
• exposed uncertainty;
• stopped at the stated authority boundary;
• made the next human decision easier.

OpenAI's guidance for working with agents recommends making purpose, scope, ownership, coordination, and lifecycle explicit. It also advises evaluating results across multiple cycles, because agent instructions and access often improve over time. That is a sound operational principle even when the product and use case differ.

Do not schedule a flawed brief and hope repetition will fix it. Cadence multiplies both useful work and bad assumptions.

Correct the right layer

Not every correction belongs in the standing instructions.

Do not assume that correcting one output automatically retrains the system or changes future behavior. Verify that the relevant instruction, source, or workflow actually changed.

Common delegation failures

Assigning an outcome the teammate cannot control

"Get us meetings" or "close the fund" depends on other people and many decisions. Assign research, preparation, follow-up review, or another controllable job instead.

Hiding judgment inside a field

A column named "good fit" can conceal a subjective decision. Ask for the evidence and criteria behind fit, then let the accountable person decide.

Confusing a name with specialization

Calling an agent "LP Expert" does not establish expertise. The quality comes from the role definition, context, sources, instructions, evaluation, and controls.

Treating instructions as permissions

"Do not read this folder" is weaker than an access control that prevents reading it. Use instructions for behavior and actual controls for access.

Automating before the artifact is useful

First make the output worth reviewing. Then decide whether a recurring trigger would help.

Turn the brief into a working relationship

Choose one bounded responsibility and write the six-part brief. Run it on normal, missing, and conflicting cases. Keep the artifact visible, improve the standing instructions deliberately, and preserve human control over relationships and decisions.

Next, learn how AI Agent Routines give recurring work a cadence, or use the AI Teammates for Private Capital guide to choose a different role.

Explore Aurora when you are ready to connect the teammate's work to the relationship context already in Finta.

Research checked September 27, 2026. The templates and examples are operational education, not legal, compliance, investment, tax, broker-dealer, fiduciary, or fundraising-outcome advice. Verify product controls and your organization's policies before assigning sensitive work.