50 AI agent prompts for Operations

Copy-paste, beginner-friendly prompts to help you automate reports, SOPs, scheduling, inventory and day-to-day admin. Each one is based on a real-world example and works with ChatGPT, Claude or an AI agent. 32 are free below, and 18 more are in AgentPlaybook Pro.

Free Operations prompts

Here are this week's updates from my team/tools: [paste]. Write a status report: wins, progress vs goals, blockers, risks, and next week's priorities. Flag anything overdue in bold.

I'll describe a process in my own words: [describe]. Turn it into a numbered SOP with: purpose, tools needed, steps, checks, and what to do when something goes wrong. Make it clear enough for an AI agent to follow.

Backlog: [paste items with rough effort]. Team capacity: [hours/person]. Dependencies: [list]. Propose a 2-week plan that fits capacity, ordered by impact, and list what we should NOT do this sprint.

From this meeting transcript [paste], extract: decisions made, action items (owner + due date), open questions, and a 3-line summary I can send to attendees.

I need these documents from each client: [list]. Here's what each client has sent so far: [paste or describe]. For each client, list what's missing, draft a short friendly reminder (email or text) asking only for those items, and flag anyone stuck for more than [X] days so I can call them. Don't send anything; give me drafts.

Build next week's staff schedule. Staff: [paste: name, role, max hours, availability, hours already worked this pay period]. Expected busy times: [paste: sales by day and hour last few weeks, or your forecast]. Rules: [minimum staff per shift by role, opening/closing needs, overtime starts at X hours]. Give me a day-by-day schedule table, total hours per person, anyone at risk of overtime, and any shift that's under- or over-staffed compared with expected sales. Explain any trade-offs in one or two lines.

Here's today's route status: [paste: driver, remaining stops with estimated drive and stop times, driving hours used today, shift hours used, hours used this week]. Our limits: [e.g. 11 hours driving, 14-hour shift window, required breaks, weekly limit]. For each driver, estimate when they'll finish and whether they're likely to break a limit, and by how much. List at-risk drivers first, with the time the problem would start and one fix for each (move a stop to another driver, add a break, reorder stops). Flag any missing or odd data instead of guessing.

Replan today's last-mile deliveries. Current rounds: [paste: route/driver, stops with addresses or zones, time windows, capacity used]. Late or failed orders to place: [paste: order id, zone/address, time window, size/weight, priority]. Rules: [max stops or weight per route, latest delivery time, which zones each driver covers, whether overtime is allowed]. Propose which orders go on which round, any stops to move, and orders that cannot fit (with why). Show a before/after stop list per route and flag anything that needs dispatcher approval. Stay inside the rules; don't invent capacity.

Compare this trade-partner submittal to the project specs for variances. Specs / drawings excerpts: [paste or attach]. Submittal: [paste or attach]. Product / system: [e.g. roofing membrane]. Produce a side-by-side compliance table: requirement from specs (quote + section), what the submittal claims, match/variance/unclear, and risk if mismatched. Cite the exact spec section for every row. List open questions for the APM. Do not approve, reject, or notify the trade partner — leave that decision to me.

You are my purchasing assistant. Every morning, go through my open purchase orders in [ERP / spreadsheet / supplier portal]. For each order due in the next [7] days: 1) check whether the supplier has confirmed the delivery date and quantity; 2) if not, draft a short, polite chase email to [supplier contact] asking them to confirm; 3) if they confirmed something different from what we ordered (date, quantity, or price), flag it. Give me one table: PO number, supplier, what we ordered, what they confirmed, status (confirmed / chased / mismatch), and the draft email. Don't send anything or change any order until I approve it.

You are a data-validation coach for [migration / onboarding / audit] work. Walk me through this validation as a checklist of [N] steps (default 9): for each step, tell me what to check, what good looks like, what files or fields I should upload or point to, and common failure modes. When I paste results or ask a question about the data, answer in plain language; if a SQL query would help, propose the exact query and explain it, but don't run writes. Current step context: [paste]. Client / system notes: [paste]. At any step that would change production data, stop and wait for my approval. End with a short pass/fail summary and open risks.

You will operate our EHR on a sandboxed desktop (computer use). Task: collect payment for patient [MRN / name], amount [X], method [card on file / send link / note only]. Runbook: [paste click path or screenshots of the billing screens]. Rules: verify the on-screen MRN matches before typing amounts; if more than one patient matches, stop; if a dialog is unexpected, screenshot and pause for me; never clear a field that already has a real default value without confirming; never submit a charge above [limit] without approval. When done, report: screens visited, amount posted or link sent, confirmation number, and any warning you saw. Prefer exporting a receipt/report file over reading clinical codes from pixels.

You are an ops email router for [desk / team]. Allowed categories and routes: [paste e.g. trade confirm, KYC chase, pricing request, complaint, spam/other]. For each email below, return: category, urgency (high/medium/low), missing fields the sender must provide, enrichment notes (account/ticket IDs if present), and the recommended route or queue. Draft a short clarification reply only when required fields are missing. Never invent balances, trade details, or approvals. Don't send mail or move tickets until I approve the batch. Emails:
[paste]

You are a field-service briefing agent for [equipment type, e.g. elevators / HVAC / industrial machines]. Assemble a pre-visit brief from: alert/telemetry [paste], unit ID and model [paste], service history [paste], contract/SLA notes [paste], and relevant manual or past-case excerpts [paste]. Produce: (1) likely root causes ranked with evidence, (2) recommended first checks on site, (3) parts/tools to bring, (4) safety and entitlement constraints, (5) a 60-second voice briefing the tech can play in the truck. Flag anything that needs a Digital Ops / expert human before dispatch. Do not invent part numbers or torque specs not in the sources.

You are a construction change-order agent for [subcontractor / trade]. Job and subcontract references: [paste]. Incoming evidence (SMS/email/PDF/portal text): [paste]. Jurisdiction notice/lien rules summary: [paste or 'use checklist below']. Produce: (1) classify as CO vs RFI vs other, (2) extract amounts, dates, references, (3) cost + schedule exposure score with assumptions, (4) draft GC notice in a professional tone, (5) calendar of statutory deadlines with owner/legal escalation triggers, (6) PM approval checklist. Never file a lien, post to a portal, or email the GC until a human PM/legal approves.

You help our fleet desk draft repair estimates. Vehicle: [VIN, make, model, mileage, current location]. Maintenance history: [paste last services and open issues]. Driver report: [paste the call notes or message describing the problem]. Approved labor rates and parts markup rules: [paste]. Produce: (1) a plain-language diagnosis of the most likely issue, (2) a line-item estimate with labor hours, parts, and totals, (3) what is uncertain and what a tech should verify on arrival, (4) a short SMS I can send the driver with next steps. Mark any line that needs a human estimator's approval before we quote the customer. If the report is too vague, ask for the three missing facts only.

Help me run a staged feature launch from Slack. Feature flag name: [name]. Rollout plan: [e.g. 5% internal → 10% → 25% → 50% → 100%]. Success metrics: [list]. Kill criteria: [e.g. error rate +X% or conversion -Y%]. Write: (1) the setup checklist before we flip the flag (dashboards, alerts, AI context), (2) the exact questions I should ask the analytics agent at smoke-test and each checkpoint, (3) a decision table for ramp / hold / roll back, (4) a post-launch readout prompt that asks for metrics moved plus qualitative themes from feedback. Keep every question short enough to paste into Slack.

You are my insurance policy checker. Here is what we requested or quoted for [client name]: [paste quote, binder, or application details]. Here is the policy the carrier issued: [paste policy declarations / schedule text]. Here is what's in our agency system: [paste current record]. Compare all three line by line for: named insured, addresses, policy dates, each coverage, limits, deductibles, endorsements, premium, and locations or vehicles listed. Give me a table with columns: item, what we asked for, what the policy says, what our system says, match yes/no, suggested correction. Then list anything that could leave the client underinsured or expose us to an errors-and-omissions problem, and draft a short note to the carrier asking them to fix the mismatches. Only use the documents I pasted and flag anything you can't find instead of guessing.

You are a retail field coach helping me, an area manager for [number] stores at [brand]. For each store, here is the data: sales vs target for [period]: [paste]; head-office tasks and whether they were completed (e.g. floor sets, promos, markdowns): [paste]; messages and whether staff read them: [paste]; training completion by topic: [paste]. Step 1: rank the stores from furthest behind to furthest ahead. Step 2: for each store that's behind, point out the likely causes by matching gaps in tasks, messages or training to the products or categories where sales are weak, and say how confident you are. Step 3: give me 2-3 specific actions per store and a talking point for my next visit or call with the store manager. Step 4: list anything the data can't explain that I should ask about in person. Use only my data and don't invent reasons.

You are assisting a claims adjuster. Claim type: [auto / home / commercial / other]. Policy summary: [coverages, limits, deductibles, key exclusions]. Claim file materials: [paste notes, FNOL, repair estimates, photos descriptions, police/medical summaries]. Step 1: list the hard facts with a source line for each (date of loss, parties, damages claimed, prior claims mentioned). Step 2: flag coverage questions and missing documents. Step 3: suggest next steps for the adjuster (request docs, inspect, reserve change, deny/partial pay) with a one-line reason each. Step 4: draft a short adjuster note they can paste into the claim system. Use only the file I gave you, mark gaps as 'needs info', and do not approve or deny the claim yourself.

You are our team's daily intake reviewer. Source: [connected shared inbox, Slack channel, or paste today's requests]. Our intake rules: [paste routing guide: request types, who handles what, required info]. Every run: Step 1: list each new request with sender, request type, deadline (if stated) and a link or reference to the original message. Step 2: put anything due within [2] business days at the top. Step 3: flag requests missing required information and say exactly what is missing. Step 4: suggest who on the team should pick up each one, based only on our rules. Step 5: end with a 3-line summary (total requests, urgent ones, ones missing info). Never reply to senders or assign work yourself; if a request doesn't match our rules, mark it 'needs a human decision'.

You are the first responder on our shared inbox [address]. Our rules: [paste: what we handle, tone, pricing/policy facts, who owns what]. For each new email: Step 1: label it (question, order/account issue, sales lead, complaint, spam, other). Step 2: if it's a simple question fully answered by our rules, draft a short reply in our tone. Step 3: if it involves money, refunds, legal, an angry customer or anything not covered by our rules, don't reply; leave a one-line internal note saying who should take it and why. Step 4: at the end, post a summary: emails handled, drafts waiting for approval, and ones handed to a human. Never invent prices, dates or promises, and keep every reply in the original thread.

You're my service-planning team. Work in four short passes, then combine them. Jobs coming in: [paste list, e.g. customer, item, problem, due date, priority]. Our locations or bays and their capacity this week: [paste, e.g. Shop A 3 techs, can't do electrical]. Parts on hand and supplier lead times: [paste]. Pass 1 (jobs): rank jobs by urgency and value to the customer. Pass 2 (work needed): for each job, estimate hours and the parts and skills it needs. Pass 3 (capacity): match jobs to the location or tech that can do them soonest. Pass 4 (parts): flag any job blocked by a missing part and say what to order today. Then give me one plan: a day-by-day schedule, jobs that will miss their due date and what to tell those customers, and the 3 changes that would free up the most capacity. If something changes [e.g. a tech is out], redo only the affected parts of the plan.

Help me set up a repair assistant my technicians can ask from their phones. We service [type of equipment, e.g. HVAC units / car-wash machines / restaurant equipment]. I'll give you: the manuals and parts diagrams for these models [attach PDFs or list them], and our past repair notes [paste or attach, e.g. date, customer, machine model, fault, what fixed it]. First, organize the repair notes into a table by machine model and fault, and point out notes that are too vague to be useful. Then answer technician questions this way: 1) the 3 most likely causes, ranked; 2) what we did last time on this exact model or machine, with the date; 3) the manual and page to check; 4) safety warnings from the manual. If the answer isn't in the manuals or notes, say so instead of guessing. Test it now with these 3 real questions from my team: [paste questions]. Finally, suggest the simplest way to put this on our phones with [tool, e.g. a Claude Project / Custom GPT / NotebookLM] and how technicians can mark an answer right or wrong.

Help me hand my contracting company's inbox and back office to a few AI helpers, one job each. We do [type of work, e.g. commercial interiors / roofing / electrical]. Here are [20-40] recent emails: [paste sender, subject and body]. Sort each one into: bid invite, active job, supplier invoice, timesheet or payroll change, customer or GC question, junk. For bid invites, add a row to a bid board table: project, GC or client, bid due date, scope, location, documents link, and 'go / no-go / needs me'. For invoices, log vendor, job, invoice number, amount and due date, and flag any that don't match a job on my list: [paste active jobs]. For timesheet changes, list worker, date, old and new hours, and who asked. Don't pay, reply to or delete anything; mark it for me. Then suggest how to split this into 3-5 separate helpers using [Grok Bot / ChatGPT / Claude / Zapier], each with its own daily log and a weekly count of what it did, and which one should start with read-only access.

Act as a project controls assistant for our construction project, [project name]. Here is our schedule with milestones, start and finish dates and percent complete: [paste or attach]. Here are recent field reports, RFIs, submittals, change orders and delivery updates: [paste or attach]. For each milestone in the next [4-8] weeks: 1) rate the risk of missing it as low, medium or high; 2) list the specific field issues, late items or open questions driving that risk, quoting the source document; 3) propose 2-3 practical mitigation steps (resequence work, expedite a delivery, add a crew, chase an RFI answer) with who should own each one; 4) note what extra information would change your rating. Put it in a table sorted by risk, then write a short summary I can share at our weekly coordination meeting. Don't change the schedule; just recommend.

Help me set up a before-and-after routine for my field technicians at my [type of business, e.g. HVAC / elevator / appliance repair / IT support] company. Part 1, pre-job briefing: using this job ticket [paste], the equipment record [paste make, model, location] and past service notes [paste], write a short phone-friendly briefing: what and where the equipment is, the reported problem, likely causes based on the history, tools and parts to bring, and any safety notes. Part 2, spoken debrief: act as a debrief interviewer. Ask me one question at a time about the visit (what I found, what I did, parts used, time spent, anything still at risk, tips for the next tech). If my answer is vague, ask a follow-up. At the end, turn my answers into a clean service report and a 2-3 line 'note for the next technician'. Then suggest how to run this with [ChatGPT voice mode / Grok Bot / a voice agent tool] and save the reports to [our job software / Google Sheets].

You prepare the document pack for a new insurance claim. The loss report: [paste]. The policy file: [paste or list documents with their effective dates: declarations pages, endorsements, forms]. 1) Pull out the policy number, insured name, date of loss, type of loss and policy type (homeowners, renters, landlord, condo). 2) Pick the declarations page that was in force on the date of loss and every endorsement and form effective on that date. Show the effective date next to each one. 3) Give the pack's table of contents in order: claim summary, declarations page, endorsements, forms. 4) List anything missing or unclear (e.g. no declarations page for that date, policy number doesn't match) under 'NEEDS A PERSON'. 5) Write a 2-line note to the claims team saying the pack is ready or what's blocking it. Use only the documents I gave you and don't decide coverage.

I'm getting quotes for [project, e.g. kitchen renovation at my home / office fit-out]. Here is what I want done: [describe the scope in plain words, or paste it]. Here are the contractor quotes: [attach PDFs or paste each quote with the contractor's name]. 1) Build one master list of every work item across all quotes, grouped by trade (demolition, plumbing, electrical, carpentry, finishes, permits, cleanup, etc.). 2) Make a table with one row per item and one column per contractor showing the price, or 'NOT INCLUDED' / 'UNCLEAR'. 3) Flag items missing from any quote and items where one price is more than 30% above or below the others. 4) Note differences in materials, brands, quantities, timeline, payment terms, warranty and what happens with extra costs. 5) Give each contractor's total adjusted for missing items, using the average of the other quotes for anything they left out, and clearly label this an estimate. 6) Write a short list of questions to send each contractor before I decide. Don't recommend one; just make the gaps clear.

You are my field schedule checker. Here is tomorrow's schedule: [paste or connect the calendar: person, appointment type (sales visit / install / service), start and end time, customer address]. Here is the list of jobs that should be scheduled tomorrow: [paste from the CRM or order list]. Each person starts from: [home / shop address]. Allow [15] minutes of buffer between stops. Do this: 1) For each person, list their stops in order and estimate the drive time between each (use maps if you can; otherwise say it's an estimate). 2) Flag any gap shorter than drive time plus buffer, any appointment too short for its job type (sales visit [60] min, install [3] hours), and any overlap. 3) Compare against the job list and flag jobs that are missing from the calendar or booked for the wrong person or day. 4) Suggest a revised order or time for each problem, keeping customer time windows where they're stated. 5) Give me a short list per person that I can post in our team chat. Don't change the calendar or message any customer; I'll approve the fixes first.

You are my automation checker. Here are my scheduled automations and their recent run history: [paste or export for each workflow: name, what it's supposed to produce (e.g. new leads, emails sent, orders synced, posts published), schedule, and for each run: date/time, status, number of items produced]. Here is what 'normal' looks like for each one: [e.g. 'usually 20-50 new leads a day'] (if you don't know, say so and I'll infer it from history). Do this: 1) For each workflow, list runs that finished as 'success' but produced zero items, or far fewer than normal. 2) Find streaks: how many runs and days in a row it produced nothing, and when that started. 3) Give the most likely causes to check (an input list or date range that ran out, a filter or search that stopped matching, an expired login, a source page that changed, a step that was disabled) without claiming you know which it is. 4) For each workflow, suggest one simple guard: e.g. 'if items produced = 0 for 2 runs in a row, send me a message', and where to add it. 5) Give me a one-line health status per workflow (OK / check / broken). Don't change any workflow yourself.

Here is the accepted purchase agreement and any addenda: [upload or paste]. Acceptance (effective) date: [date]. My side: [buyer / seller]. List every deadline and time period in the documents: earnest money, inspection, repair request and response, financing/loan commitment, appraisal, title review, HOA documents, insurance, final walkthrough, closing and possession. For each give: the deadline as a calendar date (show how you counted: calendar vs business days, and what the contract says about weekends and holidays), the clause or paragraph number, who is responsible (me, client, other agent, lender, title/escrow), and what happens if it's missed. Put them in date order as a table, then write a short email to my client listing their own next 3 actions. If a clause is unclear or a date can't be calculated, say so instead of guessing. I will check every date against the contract.

More in AgentPlaybook Pro

• Inventory reorder assistantPredict when to reorder based on sales history.
• Appointment intake & bookingAnswer booking questions and schedule 24/7, like HSS's scheduling agent.
• Daily open-issue sweepReview open issues every morning, flag the stale or urgent ones, and route them to the right person, like Poka's scheduled shop-floor agent.
• Missed clock-outs and unusual voids checkScan yesterday's time clock and POS data for missed clock-outs and voids that look out of the ordinary, like the configurable agent templates Toast is testing with restaurants.
• Set up extraction for a new document layoutWhen a document arrives in a layout your pipeline has never seen, have an agent study samples, write the field mapping and tests, and hand it to a human to review, like the BUILDSTR agent that cut DashFuel's onboarding of new bill-of-lading formats from about 90 minutes to 5.
• Assemble a prior-auth clinical packetPull chart context, answer payer clinical questions, and draft a prior-authorization packet for staff review before submission, like MIT Health using Notable's Sidekick AI co-pilot to cut authorization handling time by about 45% across 750+ Carelon submissions.
• Triage apartment maintenance requestsTurn resident maintenance calls or messages into prioritized work orders with urgency, unit context, and a calm reply draft, like The Ardizzone Group using EliseAI Maintenance to answer thousands of calls, auto-submit work orders, and cut median completion time from 29 to 17 hours.
• Self-heal a broken retailer portal scraperWhen a hand-coded portal scraper breaks after a UI change, have a computer-use agent diagnose the new screens, patch the automation, and re-cache the happy path—like the CPG data platforms a16z documented running 15–20M portal interactions a month with agents as a self-healing fallback.
• Give a team agent its own Workspace identitySpin up a coworker agent with its own email, calendar, Drive, and directory presence so the team can @mention it in Chat or Docs like a teammate—like Google's Gemini agent pattern for roles such as an events coordinator that drafts launch-readiness docs back into the group.
• Draft clinical notes from an ambient visitListen to (or take a transcript of) a patient encounter and draft the structured clinical note inside the EHR workflow so clinicians finish charts in minutes—like Muleshoe Area Medical Center with Dragon Copilot, cutting note time from 20+ minutes to about 6–10.
• Spot agents stuck on the same stepRun a watchdog that checks what each of your agents is focused on and for how long, and flags the one that's been looping for hours, like Felix Tay, who built a live view of his 89 agents after finding some stuck on the same step for 12 hours.
• Ask "Can I make payroll?" and get a straight cash answerTalk to an AI business assistant in plain English, like "Can I make payroll?" or "Who owes me money?", and it pulls your numbers, works out the answer and pauses before anything is sent or paid. Anthropic's free Small Business plugin for Claude packages 44 skills like this across cash, sales, marketing and hiring.
• Turn the day's client Slack messages into ready-to-assign Asana tasksAn agent that sits in your client and team Slack channels drafts properly structured tasks (subtasks, dependencies, due dates, standard time estimates) from the day's messages, DMs them to the account lead for approval, then pushes them to Asana. One agency owner says this made task creation 10x faster.
• Give an agent a task budget to buy the research and images it needsInstead of setting up an account and API key with every data or AI provider, a business gives an agent one job and a spending limit; the agent plans which paid services it needs, gets the plan approved, then pays each provider as it goes without going over budget. Mastercard demoed this with the marketing agency Hype at Cloudflare Connect on Oct 20, 2026 (Agent Pay for Machines, built on the x402 protocol): an agent planned a protein-bar launch campaign and made nine payments to two providers under one approved budget, and developer early access is open.
• Let technicians ask for a part in plain words and have an agent find the exact part number, check stock and order it if missingField technicians lose time on phone calls and emails just to find out which part fits a machine, whether it's in the warehouse and who will order it, and late-night breakdowns mean waking someone up. An AI agent can turn a request like 'I need a spark plug for asset 2073' into the exact part number, check stock and location, and either issue it to the work order or raise a purchase, with the person confirming each step. Diginomica reported on Oct 20, 2026 that Kodiak Gas Services uses an IFS Loops materials replenishment agent that does this, cutting each transaction from 15-20 minutes to two or three and removing the need for someone to answer the phone at 2 or 3 a.m.; every action is human-initiated with an 'Are you sure?' confirmation, and Kodiak is now building agents for sourcing, AP invoicing and job intake.
• Turn a one-line "I need..." request into a complete, policy-compliant purchase requestStaff often submit purchase requests with missing suppliers, wrong cost codes or off-contract items, and approvers bounce them back. An AI assistant can take a short plain-language request, match it to your approved suppliers, catalog and past orders, and fill in a complete request for the employee to confirm before it goes for approval. Precoro announced its Intake Agent on Oct 22, 2026, which builds a compliant requisition from a brief description using the company's contracted suppliers, catalogs and past purchases, with the requester confirming every request before submission.
• Run a product recall trace drill in minutes: from raw ingredient lot to every customer shipmentFood, cosmetics and other regulated makers must be able to show where a bad batch came from and where every unit went, and mock recalls often mean someone digging through spreadsheets for most of an hour. An AI assistant connected to your production and shipping records can trace a lot backward to its suppliers and forward to every customer and remaining stock, then draft the recall contact list. Goldman Sachs reported on Oct 22, 2026 that McClure's Pickles in Detroit cut a recall trace from 30 to 45 minutes to under two minutes after building AI into its ERP system, with a dashboard showing where batches shipped, what's left in the warehouse and batch test results.
• Decide which AI decisions can run on their own and which go to a person: set a confidence cut-off from your own past examples and route the unsure ones to reviewWhen an AI sorts emails, tickets, documents or orders, the hard part is knowing when to trust it. Instead of checking everything or nothing, you can have the AI give a confidence level with each decision, test that against examples you've already labelled, and pick a cut-off where the confident ones go straight through and the unsure ones land with a person. OpenAI made its Decisions API generally available (Oct 22, 2026) for fast choice, score and yes/no answers with confidence values; Box said it used it to classify certificate-of-insurance documents about 5x faster with the same accuracy, automating routine cases and sending lower-confidence ones to reviewers, and Sierra said it picks next actions in live customer chats at under 280 ms.

See Pro pricing →

Guides that include Operations prompts

Other categories