Cloud Nerd

Field guide

Where can we use AI in our business? A comprehensive audit checklist

Use this checklist to walk any company through the question "where can AI or agents remove human work?" It is organized so a leadership team can score it in a single sitting.

August 15, 2026

Contents

The operating principle

AI agents exist to remove human work. A human should only be in the loop to approve escalations. Agents do the rest, driven by rules and scheduled runs.

For every item on this checklist, ask four questions:

  1. Is a human doing this today?
  2. Can it be described as rules, thresholds, or a schedule?
  3. What is the escalation condition that pulls a human back in?
  4. What system of record does the agent read from and write to?

Part 1: The atomic action inventory

Almost every task a human or an agent performs inside software reduces to one of these primitives. There are not hundreds. Score each row for the volume your company does per week and whether an agent could do it instead of a person.

Record actions

  • Create a record
  • Read or look up a record
  • Update a field on a record
  • Delete or archive a record
  • Merge duplicate records
  • Relate or link two records
  • Tag, categorize, or label a record
  • Assign a record to an owner or queue

Communication actions

  • Send an email
  • Reply to an email
  • Send a text or SMS
  • Send a chat or DM (Slack, Teams, WhatsApp)
  • Place an outbound call or click to dial
  • Answer or route an inbound call
  • Post to a channel, group, or feed
  • Schedule or reschedule a meeting

Document and content actions

  • Draft a document (memo, email, proposal, spec)
  • Summarize a document or thread
  • Translate content between languages
  • Extract data from a document (PDF, image, form)
  • Fill a template with data
  • Convert between formats (PDF, DOCX, CSV, image)

Data and workflow actions

  • Enrich a record with external data
  • Match, dedupe, or reconcile records across systems
  • Move data between systems (sync, import, export)
  • Run a calculation, score, or ranking
  • Validate data against rules
  • Trigger a workflow, approval, or handoff

Decision and approval actions

  • Approve or reject a request
  • Escalate to a human
  • Flag for review
  • Route based on rules

File and asset actions

  • Upload or download a file
  • Search, retrieve, and cite from a file store
  • Generate an image, chart, or slide
  • Publish or share an asset

Part 2: Back-office motions by function

Roughly 90% of what makes companies operate is the same across industries. Score each function for the motions listed. For each yes, capture the current system of record, the rules, and the escalation condition.

Executive and administrative

  • Calendar management, scheduling, and rescheduling
  • Meeting prep briefs (agenda, attendee background, prior notes)
  • Meeting notes, action items, and follow-up assignment
  • Inbox triage, categorization, and routing
  • Draft replies for routine email
  • Travel booking and expense capture
  • Document filing and folder organization
  • Task list creation and reminder follow-up

Sales

  • Prospect research and account briefs
  • List building and enrichment
  • Personalized outbound email and sequences
  • Inbound lead qualification and routing
  • Meeting scheduling and confirmation
  • Call recording, transcription, and summary
  • CRM data hygiene (activity logging, field completion, dedupe)
  • Pipeline updates and forecasting rollups
  • Proposal and quote generation
  • Contract redlining and comparison
  • Win/loss note capture and tagging
  • Renewal and expansion signal detection

Marketing

  • Content ideation and outlining
  • First draft blog, email, and social copy
  • Repurposing long form content into variants by channel, persona, and industry
  • Newsletter assembly
  • SEO keyword research and briefs
  • Ad copy variant generation and testing
  • Image and creative generation
  • Landing page copy and A/B variants
  • Campaign performance summaries
  • Attribution and pipeline contribution rollups
  • Competitive intel monitoring
  • Review and mention monitoring

Customer support

  • Ticket triage, classification, and routing
  • First response drafting
  • Knowledge base article suggestion and drafting
  • Escalation detection based on sentiment or SLA risk
  • Customer health scoring
  • Refund, credit, and adjustment processing under threshold
  • Post-resolution survey follow-up
  • Support ticket to product feedback loop

Finance and accounting

  • Invoice generation and delivery
  • Invoice receipt, coding, and approval routing
  • Bill payment scheduling
  • Expense report review and policy check
  • Bank and credit card reconciliation
  • Journal entry drafting from source documents
  • AR follow-up and dunning sequences
  • Revenue recognition entries for standard contracts
  • Month-end close checklist execution
  • Budget vs actual variance summaries
  • Cash flow forecasting updates
  • Sales tax, VAT, and 1099 preparation

Human resources and people ops

  • Job description drafting
  • Resume screening against a scorecard
  • Interview scheduling and coordination
  • Reference check outreach
  • Offer letter and onboarding packet generation
  • New hire IT provisioning ticketing
  • Onboarding checklist tracking and reminders
  • PTO and leave request routing
  • Policy Q&A from the employee handbook
  • Performance review reminders and template prefill
  • Offboarding checklist and access revocation

IT and internal operations

  • Access request routing and provisioning
  • Password reset and account lockout resolution
  • Software license usage and renewal alerts
  • Endpoint compliance and patch reminders
  • Vendor onboarding and security review intake
  • Runbook execution for known incidents
  • Ticket categorization and assignment
  • Documentation drafting from resolved tickets
  • NDA and standard agreement generation
  • Contract intake, clause extraction, and comparison to playbook
  • Redline suggestion against approved fallbacks
  • Signature routing and status tracking
  • Policy review reminders
  • Compliance evidence collection and file placement
  • Data subject request intake and routing
  • Meeting brief prep for regulatory or counsel calls

Product and engineering

  • Backlog triage and tagging
  • Duplicate bug detection
  • Release notes drafting from merged tickets
  • Documentation drafting from PRs
  • Standup and status summaries
  • Customer feedback tagging and routing to backlog
  • Roadmap update rollups
  • On-call incident summary and post-mortem draft

Data and reporting

  • Recurring dashboard refresh and narrative
  • Weekly and monthly business review packet
  • Anomaly detection alerts on core metrics
  • Ad hoc question answering from the warehouse
  • Data quality monitoring and owner notification
  • Report distribution to stakeholders

Part 3: The agent design pattern

For each item flagged as an automation candidate, define these six fields. This is the intake form for the AI backlog.

The six-field intake

  • Trigger. Event, schedule, or inbound message that starts the run.
  • Inputs. Records, files, or fields the agent reads.
  • Rules. The decision logic in plain language, including thresholds.
  • Actions. Which primitives from Part 1 the agent performs.
  • Escalation condition. The exact rule that hands control to a human.
  • System of record. Where the agent writes the outcome and the audit trail.

Part 4: Prioritization

Score each candidate on three axes from 1 to 5.

AxisScore 1Score 5
VolumeA human rarely does thisA human does this many times a week
Rule clarityThe decision cannot be written downThe decision is clean written rules
Blast radius of an errorAn error is expensive or hard to reverseAn error is cheap and easy to undo

Priority = Volume × Rule clarity ÷ Blast radius. Low blast radius means it is safer to automate. High blast radius means keep a human tighter in the loop.

Start at the top of that list. Ship one agent. Instrument the escalation rate. Tune the rules. Move to the next.

Part 5: Governance

Before the first agent goes live, agree on these in writing.

  • Who owns each agent (a named human, not a team).
  • What the agent is allowed to do without human approval.
  • What the agent must escalate.
  • How the audit trail is preserved (which record, which field, which log).
  • How performance is measured (run count, escalation rate, error rate, cycle time saved).
  • How the agent gets retired or replaced.

The point of the audit is not to install AI everywhere. It is to identify every place a human is doing work that a rules-driven agent could do on a schedule, and to move the human to approving escalations.