SystemTypical behaviorUseful forControl question
AI assistantResponds to a prompt and returns text, analysis, or a draft.
Summaries, drafts, extraction, questions, and first-pass analysis.
Who verifies the answer before it is used?
Rule-based automationFollows predefined triggers and steps.
Routing, reminders, status changes, and repeatable system actions.
What happens when inputs do not match the rule?
AI agentPlans or selects steps, uses tools or systems, and works toward a goal.
Multistep workflows with variable inputs and clear boundaries.
What may it access or change, and where must it stop?
Multi-agent systemCoordinates specialized agents across parts of a workflow.
Larger domains with distinct tasks, permissions, and owners.
Who owns the combined outcome and resolves conflicts?
A chatbot with a new label is not automatically agentic. Require a concrete description of the workflow, tools, permissions, memory, decision rights, outputs, and stopping conditions.
These pilot designs adapt the workflows in McKinsey's March 4, 2026 analysis, with narrower permissions and explicit review points. They are not descriptions of a single shipped product, firsthand tests, or promised results.
01 / DOMAIN
Maintenance request triage
Prepare a work order for human review.
- Inputs
- A resident's request, unit ID, approved work-order history, and the property's escalation policy.
- Allowed actions
- Retrieve the relevant history, group the reported symptoms, draft a work order and resident update, and request approval. In this pilot, do not dispatch vendors or operate building systems.
- Output
- A proposed routing decision with its source records, missing information, and a draft response.
- Stop and escalate
- Escalate emergencies, uncertain unit identity, conflicting records, or denied tool access. Never delay the existing emergency process for an AI review.
People retain authority over
A property manager or maintenance lead decides urgency, access, dispatch, spending, and completion.
02 / DOMAIN
Leasing inquiry follow-up
Prepare an approved answer and tour options.
- Inputs
- The inquiry, current approved unit availability, published terms, and a permitted tour calendar.
- Allowed actions
- Retrieve the relevant facts, identify available tour slots, draft a reply, and route it to staff. Do not invent availability, approve applicants, or change terms.
- Output
- A draft reply and proposed appointment, with the availability timestamp and any unanswered question.
- Stop and escalate
- Escalate stale prices, calendar conflicts, accommodation requests, screening questions, or requests outside the approved information.
People retain authority over
The leasing team approves outbound messages and bookings and retains responsibility for applicant treatment, fair housing, and commitments.
03 / DOMAIN
Asset reporting preparation
Assemble a source-linked review packet.
- Inputs
- Approved lease records, operating statements, the prior report, and the team's reporting template.
- Allowed actions
- Retrieve the records, extract relevant terms, compare reporting periods, flag inconsistent figures, and draft the narrative. Keep access read-only during the pilot.
- Output
- A draft asset report with source references, calculation inputs, and unresolved discrepancies.
- Stop and escalate
- Escalate missing periods, conflicting lease amendments, uncertain amounts, or an unsupported explanation for a variance.
People retain authority over
An asset manager validates the sources and calculations. People retain valuation, capital-allocation, credit, and investment decisions.
04 / DOMAIN
Construction document coordination
Route a complete request to the project owner.
- Inputs
- A request for information, permitted project documents, revision dates, and the review-owner list.
- Allowed actions
- Find the relevant records, check document versions, identify missing inputs, draft the request, and propose a reviewer. Do not approve a specification or change order.
- Output
- A review packet with document references, outstanding questions, a proposed owner, and an action log.
- Stop and escalate
- Escalate conflicting revisions, safety questions, unclear scope, missing authority, or a proposed budget change.
People retain authority over
Qualified project staff approve technical answers, procurement, contracts, budgets, and field decisions.
STEP 01
Choose one bounded outcome
Select a domain with repeat volume, visible handoffs, and an accountable owner. Avoid a company-wide autonomous agent.
STEP 02
Map every system and action
List what the agent reads, writes, sends, schedules, changes, and remembers. Assign a permission and owner to each.
STEP 03
Establish the baseline
Measure cycle time, staff time, errors, rework, escalations, service outcomes, and review burden.
STEP 04
Start with full review
Review every output and proposed action. Narrow the sample only when evidence supports it and consequences remain reversible.
STEP 05
Test exceptions
Use missing data, conflicting records, unavailable tools, unusual requests, and permission failures. Confirm the agent stops.
STEP 06
Expand by evidence
Compare the pilot with the baseline. Count monitoring as an ongoing operating cost, not a one-time project.
Example pilot for work-order drafts
Use historical non-emergency tickets from one property, with personal data minimized. Compare the agent's draft routing and missing-data flags with the maintenance lead's decisions. Review every output and keep vendor dispatch, messages, spending, and building access disabled. Measure preparation time, incorrect routing, missed exceptions, and reviewer correction time. Only consider live use after the owner sets acceptance criteria and confirms that escalation works. This is an illustrative test design, not a reported customer result.
MIT Sloan warns that reclaimed time does not automatically equal labor-cost savings and that monitoring becomes a permanent operating expense. Read the analysis.