workflows

How to map a business process before automating it

Map the real workflow, exceptions, owners and human approvals with a practical packet before choosing an automation tool.

Before automating a business process, define its trigger and observable result, observe real executions, record every step with its owner, input and output, and separate the common path from exceptions. Simplify work that adds no value. Automate only the part that is stable, verifiable and reversible; keep sensitive decisions and external effects under appropriate human approval.

This guide helps you turn one real workflow into a simple, auditable specification before choosing software. It is not a BPMN course and it does not certify that a process is ready. Its original tool—the Reality-to-Automation Packet—is an editable editorial framework for finding what you know, what you assume and what still needs investigation.

If you have not chosen a candidate yet, first use our guide to choose the first process to automate. If weekly reporting is your candidate, the weekly client-reporting workflow provides a related implementation example.

Map reality, not the ideal procedure

A process map is useful only when it represents what people and systems actually do. A policy may say that a request moves directly from intake to approval. In practice, someone may correct the spreadsheet, ask for missing information in a chat, wait two days and then copy the result into another system. Those transfers, waits and corrections are part of the current process.

Do not begin by drawing the process you wish existed. Begin with evidence of the process that exists now:

  • completed forms, records or approved documents;
  • timestamps and status histories that you are authorized to inspect;
  • the people who perform and receive each step;
  • known rework and exception paths;
  • system outputs and evidence that a step finished.

Microsoft’s process-mapping documentation distinguishes activities, transitions, start and end points, variants, frequency and duration. Those concepts are enough for this exercise; you do not need a complex notation. The NIST AI Risk Management Framework likewise emphasizes context, purpose, scope, roles, impacts and human oversight before an AI system is deployed.

The map will probably expose uncertainty. Preserve it. Write unknown rather than filling a gap with a confident guess.

Choose one process and draw a boundary

Map one process at a time. “Client management” is too broad. “Prepare a weekly client-report draft after the reporting period closes” is bounded enough to observe and test.

Define:

  1. Trigger: the observable event that starts the process.
  2. Result: the output or state that proves the process ended.
  3. Start and end: the first included action and the last included action.
  4. Owner: the person accountable for the result, even when several people contribute.
  5. Beneficiary: the person or group that receives value from the result.
  6. Frequency: what has been observed, or unknown when it has not been measured.

A boundary prevents the map from quietly absorbing adjacent work. It also reveals external effects. Preparing a draft and sending it to a client are not the same process risk: the second changes what someone outside the business receives.

Observe executions instead of relying on memory

When practical, observe or reconstruct three recent executions as an initial heuristic. Three is not a statistically sufficient sample and does not prove that you found every variant. It is simply a useful starting point for comparing the formal procedure with recent work.

For each execution, ask:

  • What started it?
  • Which source supplied each input?
  • Who performed the next action?
  • Where did the work wait?
  • Which decision changed the path?
  • What was corrected or repeated?
  • How did the next person know the step was complete?
  • What evidence shows the final result was accepted?

Use existing, authorized records whenever they are sufficient. Do not record screens, meetings or employees merely because observation would be convenient. Before any recording or detailed monitoring, evaluate consent, organizational policy, credentials, confidentiality and whether a less intrusive method can answer the question.

If you cannot observe three executions, record the number you could inspect and the limitation. Never invent a frequency from memory.

The Reality-to-Automation Packet

Copy the complete packet below into a document or plain-text file. Fill it with the people who perform the work. The packet is a Pulse Work Lab editorial tool—not a standard, certification or scientifically validated method.

reality_to_automation_packet:
  packet_version: "1.0"
  mapped_on: "YYYY-MM-DD"
  mapped_by: "name or role"
  evidence_window: "dates inspected"
  executions_observed_or_reconstructed: "number or unknown"
  known_limitations:
    - "what could not be verified"

  boundary_card:
    process_name: ""
    trigger: "observable event that starts the process"
    observable_result: "output or state that proves completion"
    start: "first included action"
    end: "last included action"
    before_out_of_scope: "work immediately before this boundary"
    after_out_of_scope: "work immediately after this boundary"
    owner: "person or role accountable for the result"
    customer_or_beneficiary: ""
    observed_frequency: "value, period and evidence, or unknown"

  current_process_evidence_map:
    - step: 1
      responsible: "person or role"
      input: "required information or item"
      input_source: "authorized source"
      source_use_authorized: "yes, no or unknown"
      action: "what actually happens"
      system_or_channel: ""
      decision: "rule or judgment, or none"
      output: ""
      completion_evidence: "record, status, document or check"
      active_time: "observed value or unknown"
      waiting_time: "observed value or unknown"
      handoff_to: "next person, role or step"
      rework_observed: "what repeats and why, or none"

  exception_register:
    - condition: "what departs from the common path"
      observed_frequency: "count and evidence window, or unknown"
      detected_by: "signal, validation or person"
      decision_owner: "person or role"
      recovery: "current recovery steps"
      reversible: "yes, no or unknown — explain"
      external_effect: "none or describe"
      approval_before_effect: "named role or not applicable"
      protected_data_involved: "personal, sensitive, confidential, credential, or none"
      credential_or_secret_involved: "yes, no or unknown — never include the secret"
      evidence_consulted: "authorized record or signal used"

  readiness_gate:
    clear_start_and_end: "yes | no | unknown"
    acceptable_result_is_verifiable: "yes | no | unknown"
    accountable_owner_exists: "yes | no | unknown"
    inputs_and_sources_are_known: "yes | no | unknown"
    required_fields_are_defined: "yes | no | unknown"
    repeatable_decisions_have_rules: "yes | no | unknown"
    common_exceptions_have_known_treatment: "yes | no | unknown"
    error_can_be_detected_before_relevant_harm: "yes | no | unknown"
    safe_stop_exists: "yes | no | unknown"
    external_effects_require_approval_when_needed: "yes | no | unknown"
    data_use_is_authorized: "yes | no | unknown"
    baseline_evidence_exists: "yes | no | unknown"
    investigations_required:
      - "every unknown becomes an investigation"
    blocking_conditions:
      - "unauthorized data, missing containment or other blocker"

  automation_handoff_contract:
    included_path:
      - "stable path included in the proposed test"
    excluded_paths:
      - "exception or consequence outside the test"
    required_fields:
      - field: ""
        authorized_source: ""
        validation: ""
    deterministic_rules:
      - "if/then rule that does not require interpretation"
    ai_may_assist_with:
      - "bounded drafting, classification or extraction task"
    ai_does_not_decide:
      - "sensitive judgment, unsupported interpretation or external delivery"
    human_approvals:
      - before: "external or sensitive effect"
        approver: "named role"
        checks: "what must be verified"
    safe_stop:
      - "condition that stops the run before an external effect"
    log_or_evidence:
      - "minimal, authorized evidence retained for review"
    minimum_test:
      data: "fictitious, redacted or otherwise authorized"
      cases:
        - "common path"
        - "known exception"
        - "missing or conflicting input"
      pass_condition: "observable and reviewable result"
      fail_condition: "unsafe, unsupported or unverifiable result"
      comparison_baseline: "current method and evidence available"
      review_owner: "named accountable role"

The sections below explain how to fill it without turning uncertainty into false precision.

Build the evidence map step by step

Use one row for each meaningful handoff, action or decision. Start with verbs: receive, check, correct, approve, copy, calculate, draft, review, send or archive.

For every step, identify:

  • responsible: who performs it;
  • input and source: what the person needs and where it comes from;
  • action: what actually happens;
  • system or channel: where it happens;
  • decision: which condition changes the path;
  • output: what the step produces;
  • completion evidence: how another person can verify it finished;
  • active and waiting time: observed values when available, otherwise unknown.

Do not combine “analyst prepares, manager reviews and client receives” into a single step. Each transfer changes responsibility and may introduce waiting, permission or error risk.

Record decisions, handoffs, waits and rework

A straight-line diagram often hides most of the operating cost. Mark four things explicitly:

  1. Decisions: Which condition selects the next path? Is it a written rule or personal judgment?
  2. Handoffs: Who gives what to whom, through which authorized channel?
  3. Waiting: Is the work waiting for a date, an input, capacity or approval?
  4. Rework: Which previous step repeats, why and how often was that observed?

Frequency and duration can help identify common variants and bottlenecks, but only when they come from evidence. If timestamps are incomplete, say so. Do not estimate them merely to make the map look complete.

Make a separate exception pass

After the common path is mapped, inspect each execution again only for exceptions. This second pass reduces the chance that unusual work disappears inside a generic step.

For each exception, capture the condition, observed frequency, detection signal, decision owner, recovery path, reversibility and any external effect. Examples include:

  • a required metric is missing;
  • two approved sources disagree;
  • the usual reviewer is unavailable;
  • a record arrives after the reporting period closes;
  • a draft contains an interpretation unsupported by the source.

An exception with unknown treatment is not an invitation for AI to improvise. It is an investigation or a stop condition.

Mark data, credentials and third parties

For every input and tool, distinguish:

  • authorized source;
  • specific purpose;
  • minimum fields needed;
  • personal, sensitive or confidential data;
  • credentials and secrets;
  • people and providers with access;
  • retention and deletion expectations;
  • external effect created by the output.

Never paste credentials or secrets into the packet. Do not send personal, sensitive or confidential information to an AI tool without applicable authorization and assessment. Use fictitious or properly protected data in examples and early tests.

Recording a process can itself create a new data flow. Screenshots, screen recordings and meeting transcripts may expose client names, employee information, private messages or access tokens. Evaluate consent and policy first, minimize the capture and prefer existing authorized evidence when it is enough.

The ANPD’s guidance on generative AI highlights risks involving personal data and the importance of necessity and transparency. Applicable legal and organizational requirements vary. This guide is educational and operational, not legal advice or a substitute for internal governance.

Simplify before automating

Once the current process is visible, inspect each step in this order:

  1. Eliminate: Does the step produce evidence, quality or value that someone actually needs?
  2. Combine: Can two handoffs become one controlled action?
  3. Clarify: Can an ambiguous judgment become an explicit rule or approval criterion?
  4. Standardize: Can the input, field name or output format become consistent?

Do not automate waste simply because it is repeatable. If every execution requires someone to repair a malformed input, standardizing the input may be the best intervention. If two teams collect the same field, removing duplication may be safer than synchronizing it with an integration.

Document the proposed change separately from the current-state map. Otherwise, reviewers will not know whether a step is observed or merely desired.

Classify each remaining step

Use four practical classes:

Human work

Use when the step depends on accountable judgment, negotiation, contextual knowledge or a sensitive decision. Decisions involving money, contracts, people, reputation or external communication may require human control even if software prepares supporting information.

Deterministic rule

Use when the same known condition reliably produces the same action: required-field checks, date comparisons, approved calculations and routing rules. Conventional automation is often more appropriate than generative AI here.

AI assistance

Use for bounded work such as preparing a draft, extracting candidate fields or classifying text when the result can be reviewed before consequence. Define permitted sources, unsupported-output handling and the person who checks it.

Automation with an external effect

Use when the system sends, publishes, changes, purchases, commits or otherwise affects another person or system. These steps need explicit permissions, logging, safe stops and appropriate approval. Preparing a client email is not the same as sending it.

Classification is not permanent. Start with the narrowest useful assistance and widen the action surface only when evidence and controls justify it.

Apply the readiness gate

Answer yes, no or unknown for all twelve questions:

  1. Are the start and end unambiguous?
  2. Can an acceptable result be verified?
  3. Is there an accountable owner?
  4. Are inputs and their sources known?
  5. Are required fields defined?
  6. Do repeatable decisions have documented rules?
  7. Do common exceptions have known treatment?
  8. Can an error be detected before relevant harm?
  9. Is there a safe stop?
  10. Do external effects require approval when necessary?
  11. Is the use of the data authorized?
  12. Is there basic evidence for a before-and-after comparison?

This gate does not produce a universal score, certification or automatic approval. Treat each unknown as an investigation. Keep an unknown frequency as unknown; never estimate it silently.

Unauthorized data use or the absence of containment blocks a pilot with real data. An unstable process should be stabilized before automation. A no may lead to redesign, a narrower assistance-only scope or a decision not to automate.

Write the handoff contract

The handoff contract translates the evidence map into a bounded proposal. It must state:

  • which stable path is included;
  • which paths and consequences are excluded;
  • required fields and authorized sources;
  • deterministic rules;
  • the bounded tasks where AI may assist;
  • who approves what, and what they check;
  • conditions that stop the run safely;
  • minimal logs or evidence that support review without retaining unnecessary data;
  • a minimum test using fictitious, redacted or otherwise authorized data.

Test at least a common path, a known exception and missing or conflicting input. A test passes only when its output is observable, reviewable and contained. It fails when the system proceeds with an unsupported claim, bypasses approval or cannot explain why it stopped.

Filled example: weekly client-report preparation

The following example is entirely fictitious. It continues the weekly-report scenario used elsewhere on Pulse Work Lab and does not represent a real client, employee or result.

reality_to_automation_packet:
  packet_version: "1.0"
  mapped_on: "2026-09-28"
  mapped_by: "fictional operations analyst"
  evidence_window: "three fictional weekly cycles"
  executions_observed_or_reconstructed: 3
  known_limitations:
    - "Three examples are a starting heuristic, not a sufficient statistical sample"
    - "Holiday-week behavior was not observed"

  boundary_card:
    process_name: "Prepare the weekly report draft for fictional client Aurora"
    trigger: "Friday 14:00 with the reporting period closed"
    observable_result: "Draft report ready for account-manager review"
    start: "Check that the reporting period and approved sources are closed"
    end: "Make the factual draft available for internal review"
    before_out_of_scope: "Produce and approve metrics in their source systems"
    after_out_of_scope: "Interpret unsupported variations, approve and deliver to the client"
    owner: "Account manager"
    customer_or_beneficiary: "Account team and fictional client Aurora"
    observed_frequency: "Once per week in all three fictional cycles"

  current_process_evidence_map:
    - step: 1
      responsible: "Operations analyst"
      input: "Reporting period and list of approved sources"
      input_source: "Fictitious internal calendar and register"
      source_use_authorized: "yes"
      action: "Verify that the reporting period is closed"
      system_or_channel: "Fictitious internal board"
      decision: "Is the reporting period closed? Continue only when its status is closed"
      output: "Confirmed period or stopped process"
      completion_evidence: "Closed status recorded"
      active_time: "unknown"
      waiting_time: "unknown"
      handoff_to: "Step 2"
      rework_observed: "None in the fictitious executions"
    - step: 2
      responsible: "Operations analyst"
      input: "Approved metrics for the period"
      input_source: "Fictitious exports from authorized sources"
      source_use_authorized: "yes"
      action: "Collect required fields without changing source values"
      system_or_channel: "Fictitious internal folder"
      decision: "Are all required fields present and consistent? Stop if a field is missing or sources disagree"
      output: "Checked metrics set"
      completion_evidence: "Completed required-fields checklist"
      active_time: "unknown"
      waiting_time: "unknown"
      handoff_to: "Step 3"
      rework_observed: "Return to the source when a discrepancy exists"
    - step: 3
      responsible: "AI drafting assistant under internal operation"
      input: "Checked metrics and approved template"
      input_source: "Output of step 2 and fictitious editorial template"
      source_use_authorized: "yes"
      action: "Prepare a factual summary and flag gaps"
      system_or_channel: "Approved test environment with fictitious data"
      decision: "No business decision; only flag gaps, never invent explanations and never change metrics"
      output: "Factual draft"
      completion_evidence: "Document saved with linked sources"
      active_time: "unknown"
      waiting_time: "unknown"
      handoff_to: "Step 4"
      rework_observed: "Regenerate only after correcting the identified input"
    - step: 4
      responsible: "Account manager"
      input: "Draft and source checklist"
      input_source: "Output of step 3"
      source_use_authorized: "yes"
      action: "Check numbers, support for claims and contextual suitability"
      system_or_channel: "Fictitious internal document"
      decision: "Approve, return for correction or stop; no external delivery without explicit approval"
      output: "Internally approved or returned draft"
      completion_evidence: "Review status and reviewer identity recorded"
      active_time: "unknown"
      waiting_time: "unknown"
      handoff_to: "External delivery process, outside scope"
      rework_observed: "Return when a claim lacks support"

  exception_register:
    - condition: "Reporting period is incomplete"
      observed_frequency: "unknown"
      detected_by: "Period status differs from closed"
      decision_owner: "Account manager"
      recovery: "Wait for closure or cancel that week's run"
      reversible: "yes — fully reversible before delivery"
      external_effect: "none"
      approval_before_effect: "Account manager"
      protected_data_involved: "none in the fictitious example"
      credential_or_secret_involved: "none in the test packet"
      evidence_consulted: "Fictitious period status"
    - condition: "Required metric is missing"
      observed_frequency: "unknown"
      detected_by: "Incomplete required-fields checklist"
      decision_owner: "Operations analyst escalates to the account manager"
      recovery: "Request correction at the source; do not fill by estimation"
      reversible: "yes — fully reversible before delivery"
      external_effect: "none"
      approval_before_effect: "Account manager"
      protected_data_involved: "none in the fictitious example"
      credential_or_secret_involved: "none"
      evidence_consulted: "Fictitious checklist"
    - condition: "Approved sources show different values"
      observed_frequency: "unknown"
      detected_by: "Comparison of identifier, period and value"
      decision_owner: "Metric owner and account manager"
      recovery: "Stop until the authorized source is identified and corrected"
      reversible: "yes — fully reversible before delivery"
      external_effect: "none while the stop is respected"
      approval_before_effect: "Account manager"
      protected_data_involved: "none in the fictitious example"
      credential_or_secret_involved: "none"
      evidence_consulted: "Divergent fictitious exports"
    - condition: "Draft contains an interpretation unsupported by sources"
      observed_frequency: "unknown"
      detected_by: "Review compares each claim with linked sources"
      decision_owner: "Account manager"
      recovery: "Remove the interpretation or add authorized human context"
      reversible: "yes before delivery; undetermined after delivery"
      external_effect: "Reputational risk if delivered"
      approval_before_effect: "Mandatory account-manager review"
      protected_data_involved: "none in this fictitious example"
      credential_or_secret_involved: "none"
      evidence_consulted: "Draft and linked fictitious sources"

  readiness_gate:
    clear_start_and_end: "yes"
    acceptable_result_is_verifiable: "yes"
    accountable_owner_exists: "yes"
    inputs_and_sources_are_known: "yes"
    required_fields_are_defined: "yes"
    repeatable_decisions_have_rules: "yes"
    common_exceptions_have_known_treatment: "yes — the four documented cases stop, recover or return for review"
    error_can_be_detected_before_relevant_harm: "yes — named review occurs before external delivery"
    safe_stop_exists: "yes"
    external_effects_require_approval_when_needed: "yes"
    data_use_is_authorized: "unknown — a real organization must verify this before using real data"
    baseline_evidence_exists: "unknown — active and review time have not been measured"
    investigations_required:
      - "Confirm authorization, access, retention and permitted tools for the real data"
      - "Measure current preparation and review evidence without intrusive monitoring"
      - "Observe whether additional variants appear beyond the three fictitious cycles"
    blocking_conditions:
      - "Do not test with real data until authorization and tool handling are confirmed"
      - "Do not create an external effect without named account-manager approval"

  automation_handoff_contract:
    included_path:
      - "Collect approved metrics"
      - "Flag missing fields"
      - "Draft factual summary"
      - "Prepare document for review"
    excluded_paths:
      - "Invent missing explanations"
      - "Change source metrics"
      - "Send report to client"
    required_fields:
      - field: "client"
        authorized_source: "fictitious internal register"
        validation: "must match the requested account"
      - field: "reporting period"
        authorized_source: "approved metrics sheet"
        validation: "must be closed and match the requested week"
      - field: "metric name, value and unit"
        authorized_source: "approved metrics sheet"
        validation: "all required labels and values must be present"
      - field: "project updates"
        authorized_source: "project-board export"
        validation: "must fall within the reporting period"
      - field: "approved source and approval status"
        authorized_source: "fictitious internal register and source exports"
        validation: "source must be authorized and status approved"
    deterministic_rules:
      - "Stop when the reporting period is incomplete"
      - "Stop when a required metric is missing"
      - "Stop when approved sources disagree"
    ai_may_assist_with:
      - "Draft a factual summary from the validated, authorized input set"
      - "Flag missing fields and candidate unsupported interpretations for review"
    ai_does_not_decide:
      - "Cause of a variation without evidence"
      - "Change to a source value or source selection"
      - "Suitability of an interpretation for the client relationship"
      - "External delivery"
    human_approvals:
      - before: "Any external delivery"
        approver: "Account manager"
        checks: "metrics, reporting period, factual support, tone, recipients and confidentiality"
      - before: "Any interpretation not directly supported by the approved inputs"
        approver: "Account manager"
        checks: "supporting evidence and whether the interpretation should be removed"
    safe_stop:
      - "Reporting period is incomplete"
      - "Required metric is missing"
      - "Sources disagree"
      - "Output contains an unsupported claim"
      - "Authorized source is unavailable"
    log_or_evidence:
      - "Version of the inputs used"
      - "Missing fields and discrepancies without sensitive content"
      - "Draft version"
      - "Reviewer identity and decision"
    minimum_test:
      data: "Fictitious records matching the required schema"
      cases:
        - "Complete and consistent weekly inputs"
        - "Missing required metric"
        - "Conflicting values across approved sources"
        - "Unsupported claim in the draft"
      pass_condition: "No metric is changed, every gap is flagged and no external delivery occurs"
      fail_condition: "The process invents data, modifies a source value, misses a required stop or proceeds to external delivery"
      comparison_baseline: "Current manual preparation and review evidence, still to be collected"
      review_owner: "Fictitious account manager"

This example deliberately leaves two gate items as unknown. A polished template must not hide missing authorization or missing baseline evidence. The proper next step is investigation with fictitious data—not a claim that the workflow is ready.

When the process is not ready

Do not move to a real-data automation pilot when:

  • the process changes substantially from one execution to the next;
  • nobody is accountable for the result;
  • the acceptable output cannot be checked;
  • required inputs or their sources are unknown;
  • common exceptions depend on undocumented knowledge;
  • error detection happens only after meaningful harm;
  • there is no safe stop or recovery path;
  • the data use, provider access or retention is not authorized;
  • the proposed first version can send, publish, purchase or commit without appropriate approval;
  • the organization cannot retain enough minimal evidence to investigate a failure.

The answer is not always “never automate.” It may be “simplify first,” “standardize the input,” “write the decision rule,” “collect baseline evidence,” “test assistance only” or “keep this step human.”

Your next step

Choose one bounded process and reconstruct up to three recent executions when feasible. Fill the boundary card and evidence map before discussing vendors. Then run the exception pass, mark protected data and apply the readiness gate.

If an answer is unknown, assign an investigation rather than a score. If real-data authorization or containment is missing, keep the minimum test fictitious. Only after the stable included path is clear should you finish the automation handoff contract and evaluate a tool.

The next useful connections are our guide to choose the first process to automate and the worked weekly client-reporting workflow.

Frequently asked questions

What is the difference between a process map and an automation specification?

A process map documents the current work, including owners, inputs, decisions, waits and exceptions. An automation specification selects a stable subset, defines rules, permissions, approvals, stops and tests. The Reality-to-Automation Packet keeps both connected without pretending they are the same artifact.

Do I need BPMN software to map a process?

No. A text table or the YAML packet in this article is sufficient when it makes the trigger, actions, decisions, transfers and result clear. BPMN offers a formal notation for events, activities, flows and gateways, but the simplest notation that the responsible people can review is often the better starting point.

How many executions should I observe?

When practical, three recent executions are a useful initial heuristic for exposing variation. They are not statistically sufficient and do not guarantee that every exception was found. Record the actual number inspected, its evidence window and its limitations.

What if the written procedure and real work disagree?

Record both separately. Map the observed current state, cite the written procedure as a requirement and investigate the difference with the process owner. Do not silently redraw reality to match the document or assume that the observed workaround is authorized.

Should unknown frequencies be estimated?

No. Keep them as unknown and define how they could be measured. An unverified estimate may make a rare exception look common—or hide a common failure behind false precision.

Can AI observe employees to create the map?

Do not introduce recording, screenshots, transcription or monitoring without evaluating consent, organizational policy, credentials, privacy and necessity. Prefer existing authorized records or a facilitated reconstruction when they can answer the question with less intrusion.

When can AI be included in the process?

After its purpose, permitted inputs, review mechanism and safe stop are defined. Begin with a bounded assistance task whose output is reviewed before consequence. Do not allow the model to invent missing information or decide how to handle an undocumented exception.

Who should approve external actions?

A named, accountable role with the authority and context to inspect the result. The approval must state what that person checks. A generic “human in the loop” label is not a control if nobody performs a defined review before the external effect.

Does passing the readiness gate prove the automation will work?

No. The gate exposes missing foundations; it is not a certification or prediction of performance, savings or return on investment. A bounded test is still required, and the result applies only to the tested process, inputs and controls.

Once the map supplies a defensible baseline, use the automation ROI calculator to estimate cost, capacity and payback with those documented premises.

Use the mapped decision points and external effects to define meaningful human approval. Turn known exceptions, the safe stop and recovery evidence into an automation failure contingency plan.

For concrete applications of the map, document the commercial states in a proposal and follow-up workflow or the inputs, decisions and exceptions in a message-to-task intake flow.

When the map exposes competing spreadsheets, unclear handoffs or an unstable source of truth, use the spreadsheet-or-CRM decision packet before choosing a new system.

Method note

These sources support the process, risk, privacy and oversight principles. They do not validate the Reality-to-Automation Packet, establish a universal readiness threshold or guarantee savings, error reduction or return on investment.

Sources