A permission-first field guide


The best first agent is not the one that can do the most. It is the one with a clear job, the minimum necessary permissions, a named human checkpoint, and a result you can measure against the way the work gets done today.

  • Updated July 13, 2026
  • Evidence-backed guide
  • 15-minute read

The permission envelope

Expand only from evidence →

  1. 0

    Advise

    Reason from supplied context; no connected tools

    A person decides whether to use the answer

  2. 1

    Read

    Search approved, read-only sources

    No records, messages, or files can be changed

  3. 2

    Prepare

    Draft, classify, extract, or assemble a proposed action

    A person reviews the complete output

  4. 3

    Approve to act

    Use named tools only after explicit approval

    Each external or consequential action is confirmed

  5. 4

    Bounded action

    Act inside strict limits, logs, and stop conditions

    Exceptions and thresholds force a handoff

  6. 5

    Open-ended

    Broad tools, objectives, or authority

    Not a sensible first small-business pilot

Use-case suitability matrix

These are design patterns, not documented GetEducated.ai client outcomes or performance benchmarks. Adapt them to your data, policy, risk, and existing process.

JobWhere it fitsLevelHuman gateMeasureStop when

Inbox triage

Useful when messages require context and routing1–2Person sends replies and handles sensitive messagesRouting corrections and response timeMissing context, sensitive data, or uncertain recipient

Lead research

Useful when the research path changes by company1–2Person controls outreach and CRM stageAccepted research briefs and time per qualified leadConflicting identity, weak sources, or restricted data

Support drafting

Useful for retrieving policy and preparing a response1–2Person approves refunds, account changes, and exceptionsCorrection rate and resolution timePolicy conflict, complaint escalation, or account action

Meeting follow-up

Often an AI-assisted workflow; agent complexity may not be needed0–2Person verifies commitments and sendsCorrections and time to approved follow-upUnclear owner, deadline, or consent to record

Vendor research

Useful when evidence gathering requires adaptive checks1–2Person selects, contracts, and purchasesEvidence completeness and review timeMissing primary evidence or material risk

Content research brief

Useful for sourcing and organizing competing evidence1–2Person fact-checks claims and publishesSource acceptance and revision rateUnsupported claim, weak source, or rights concern

Bookkeeping or payments

Poor first autonomous agent; use extraction or reconciliation support0–2Person posts entries, moves money, and approves paymentsCorrection rate and review timeAny payment, tax judgment, or unresolved mismatch

Hiring screening

High-impact work; do not delegate the decision0–1People define criteria and make every employment decisionEvidence organization quality, not automated selectionProtected data, inferred traits, ranking, or rejection

Agent fit score

Give each factor 0, 1, or 2 points: no, partly, or yes. A higher score suggests the job may benefit from adaptive behavior. It does not grant permission. If a mistake can materially affect rights, money, safety, access, or employment, lower the autonomy and require accountable human review.

Score one real workflow
  1. 01

    Uncertainty: Does the next useful step change with context?

    If no, fixed automation may be the better architecture.

  2. 02

    Repetition: Does the job recur often enough to evaluate?

    One-off work rarely produces a reliable pilot.

  3. 03

    Tool need: Must the system choose among sources or tools?

    A single model task is not automatically an agent.

  4. 04

    Reviewability: Can a person inspect the evidence and proposed action?

    Hidden reasoning is not a substitute for reviewable output.

  5. 05

    Reversibility: Can mistakes be contained or undone?

    Irreversible work requires a lower permission level.

  6. 06

    Measurement: Is there a baseline, success metric, and stop condition?

    Without these, a demonstration can masquerade as value.

A controlled pilot

Step 1

Name one job

Write the trigger, inputs, required output, owner, and current baseline.

Step 2

Set the envelope

Choose the lowest permission level that can complete representative work.

Step 3

Create the evaluation set

Include normal, incomplete, unusual, and adversarial examples from approved data.

Step 4

Run in shadow mode

Let the agent prepare work while the existing process remains authoritative.

Step 5

Review the evidence

Track task success, corrections, failures, latency, operating cost, and human review time.

Step 6

Keep, narrow, or stop

Increase authority only when observed results support a specific, bounded change.

Direct answers

An AI agent is software in which a model directs at least part of a workflow, such as choosing a tool, gathering information, or adapting the next step within defined boundaries. A model that performs one fixed task inside a predefined workflow is not necessarily an agent.

Choose a recurring, reviewable, reversible information task where the path genuinely changes with context. Read-only research, triage, and draft preparation are stronger first pilots than autonomous payments, hiring decisions, refunds, deletions, or external communications.

Start at the lowest level that can produce useful evidence: advise, read approved sources, or prepare a draft. Keep consequential actions behind explicit human approval. Broader autonomy should be earned from documented performance on representative cases, not assumed from a demonstration.

Compare it with the current process using task success, human corrections, failure rate, latency, model and tool cost, review time, and the business metric the job is meant to improve. Define a stop condition before the pilot begins.

Keep people accountable for high-impact decisions involving employment, credit, healthcare, legal rights, financial transfers, access permissions, safety, or similarly consequential outcomes. AI may help organize approved evidence, but a high fit score should never override the impact of a mistake.

Use normal automation when triggers, rules, and actions are predictable. Consider an agent when the job requires interpreting changing context, selecting among tools, or adapting the next step. Use the simplest architecture that can complete the work reliably.

Primary sources

The permission ladder and suitability matrix are GetEducated.ai decision tools. The supporting risk, oversight, adoption, and agent-design principles are grounded in the primary sources listed here.

Choose the first job

Audit the work before choosing the tool. If you want implementation support after the decision is clear, explore Builder.

Audit my first agent workflowSee Builder pricing