A8gent
Agent playbooks

Use-case playbooks for agents people actually ship.

Support, sales, ops, marketing, recruiting. Each playbook covers the workflow, the source data, the human review points, and a safe first version.

ForFounders, support teams, SaaS operatorsSales teams, agencies, consultantsSEO teams, creators, course businessesOperators, founders, project managersAgents, brokerages, property marketersRecruiters, HR teams, founders
All 16Build 11Use 5
Build
Agent playbook

Customer Support AI Agent: How to Set One Up Safely

Design a customer support AI agent that answers common questions, drafts replies, routes edge cases, and keeps humans in control of every escalation.

customer support ai agent
Beginner · 2.6 hr
Build
Agent playbook

Email AI Agent for Sales Outreach: Tools and Cost

Build an email AI agent for sales outreach: which tools do the account research, what a researched account really costs, and when DIY in n8n wins.

email ai agent for sales outreach
Intermediate · 3.4 hr
Build
Agent playbook

AI Agent Content Researcher: Briefs That Get Cited

Put an AI agent to work as a content researcher. Briefs built from keyword clusters, competitor pages and what the AI answer already cites.

ai agent content researcher
Beginner · 4.1 hr
Use
Agent playbook

Operations AI Agent: Automate Status Chasing

Use an operations AI agent to summarize meetings, chase missing inputs, update trackers, and surface blocked work before it costs you a deadline.

operations ai agent
Intermediate · 2.2 hr
Build
Industry agent

Real Estate AI Agent for Lead Follow-Up

Plan a real estate AI agent for lead qualification, listing follow-up, neighborhood research, and showing prep, so no enquiry sits unanswered overnight.

real estate ai agent
Beginner · 2.6 hr
Build
Industry agent

Recruiting AI Agent: Screen Resumes and Outreach

Build a recruiting AI agent that screens resumes, drafts candidate outreach, summarizes interviews, and flags fit gaps before they reach your shortlist.

recruiting ai agent
Intermediate · 3.4 hr
Build
Agent playbook

Email Assistant AI Agent: Inbox Triage Workflow

Build an email assistant AI agent that triages your inbox, drafts replies, pulls missing context, and protects time-sensitive work from getting buried.

email assistant ai agent
Beginner · 4.1 hr
Use
Agent playbook

Meeting Follow-Up AI Agent: Notes to CRM Tasks

A meeting follow-up AI agent is not the transcript. It is the handoff: owners, deadlines, CRM notes and sent emails, plus what the tool stack costs to run.

meeting follow up ai agent
Beginner · 2.2 hr
Build
Industry agent

Ecommerce Support AI Agent for Shopify Stores

Design an ecommerce support AI agent for order status, returns, product questions, subscriptions, and refund routing, with clear limits on what it sends.

ecommerce support ai agent
Intermediate · 2.6 hr
Build
Finance agent

Invoice and AP AI Agent: Automate Approvals

Plan an accounts payable agent that extracts invoice data, checks purchase orders, routes approvals, and prepares payment queues.

invoice ap ai agent
Intermediate · 3.4 hr
Build
Industry agent

Legal Intake AI Agent: Screen and Route Leads

Build a legal intake agent that collects matter details, checks conflicts, summarizes documents, and routes prospects safely.

legal intake ai agent
Advanced · 4.1 hr
Build
Agent playbook

Marketing Campaign AI Agent vs ChatGPT

How an AI marketing agent differs from ChatGPT: it turns campaign goals, audience data, and past performance into launch-ready assets inside guardrails.

ai marketing agent vs chatgpt
Intermediate · 2.2 hr
Use
Agent playbook

Customer Onboarding AI Agent: Setup and Risk Flags

Create an onboarding agent that guides new customers, answers setup questions, checks progress, and flags accounts at risk.

customer onboarding ai agent
Intermediate · 2.6 hr
Use
Agent playbook

Executive Assistant AI Agent: Stack and Pricing

Design an executive assistant AI agent for inbox triage, agenda prep and briefing notes, with the tool stack, its running cost, and where it must not act.

executive assistant ai agent
Intermediate · 3.4 hr
Build
Analytics agent

Data Analyst Reporting AI Agent: Weekly Updates

Build a reporting agent that answers metric questions, explains changes, drafts weekly updates, and keeps data interpretation grounded.

data analyst reporting ai agent
Advanced · 4.1 hr
Use
Agency agent

Agency Client Reporting AI Agent: Draft Reports

Create an agent that compiles campaign results, highlights client-ready insights, drafts reports, and prepares account managers.

agency client reporting ai agent
Intermediate · 2.2 hr

What this collection is for

Sixteen playbooks, each covering one job an agent can realistically hold down: support triage, sales prospecting, invoice chasing, recruiting screens, client reporting and so on. A playbook is the workflow as it happens today, the source data the agent needs, the points where a human has to sign off, and the smallest safe first version. It is the thinking, not the artifact and not the code.

This is the wrong hub if you want something to copy and paste - that is templates and guides. It is also the wrong hub if the process you have in mind has no stable rules yet. Write it down by hand for a month first; the process documenter makes that less painful than it sounds.

How to choose

Pick by the shape of the job, not by which agent sounds most impressive.

  1. Where does the work arrive? If it arrives as inbound questions, read the customer support agent, or the ecommerce support agent if orders and returns are the bulk of it. If it arrives as email, the email assistant and the executive assistant are the pair to compare. If it arrives by phone or web form in a regulated setting, start at legal intake.
  2. Is the job to chase, or to answer? Chasing work - sales prospecting, invoice and AP, meeting follow-up, customer onboarding - is mostly cadence and stopping rules, and it is the safest category to start with because a missed send is recoverable. Answering work needs retrieval over your own material and a much tighter guardrail story.
  3. Is the output a document or an action? Documents - reporting, client reporting, content research, campaign work - fail visibly and cheaply, so ship those first. Actions that move money, book time or contact a candidate need an approval step from day one: see recruiting and operations.
  4. Is your industry named? If you sell property, the real estate agent playbook is written for it, and automation by industry covers the rest in owner language, including the compliance notes that apply to each trade.
  5. Does it repeat weekly? If the honest answer is no, close the playbook. Repetition predicts payoff far better than how annoying a task feels, and work that runs twice a month rarely clears the build cost. The ROI calculator will tell you in under a minute.

Where this sits

Playbooks are stack-agnostic on purpose: they say what to build, not what to build it on. Platforms answers that, models answers which LLM runs underneath, and compare breaks a two-way tie. Templates are the documents a playbook asks you to produce, such as the support SOP or the workflow brief. Build guides are the fully wired version of a handful of these workflows, and learn is where the vocabulary comes from.

What to do first

Choose the playbook matching a task your team repeated this week, then read its source-data section before anything else - most agent projects fail there rather than in the model. If the records are spread across three systems or living in somebody's inbox, fix that first and check the integration checker for what actually connects. Then ship the smallest version with a human approving each action, and only widen once you have seen the exceptions. If you want the whole thing scoped for you instead, the free audit ends in a written roadmap you keep.

How to read a playbook

One agent that holds down a real job beats a grand autonomous system that never ships

Each playbook covers one job an agent can actually keep: support triage, sales follow-up, ops reporting, recruiting screens. It is not a list of features. It is the workflow, the source data the agent needs, the points where a human has to sign off, and the smallest safe version to ship first.

Read the source-data section before anything else. Most agent projects fail there, not in the model. If the data an agent needs to do the job is messy or locked away, the playbook tells you to fix that first.

Every playbook ends with a first version you can build in a week. That is deliberate. The goal is one narrow agent in production earning trust, then expanding from there.

FAQ

Questions people ask

Still stuck? Ask us directly and we will answer.

What is an agent playbook?

A step-by-step blueprint for one use case: the workflow, the data it needs, the human review points, and a safe first version to build.

Which playbook should I start with?

Pick the one that matches a task your team already repeats every week. Repetition is the strongest signal that automation will pay off.

Do these work with any platform?

Yes. The playbooks are stack-agnostic. Pair one with the stack picker tool to choose the right builder.

How is this different from a template?

A template is the artifact you download. A playbook is the thinking around it: what to build, why, and where it breaks.

How is this different from the platforms directory?

The playbooks say what to build. The platforms directory says what to build it on, and the models directory says which LLM does the reasoning underneath. Read a playbook first: choosing a platform before you have named the workflow is the most expensive mistake in this space.

How many playbooks are there?

Sixteen, covering support, sales, ops, marketing, recruiting, finance, legal intake, real estate and reporting. If your job is not on the list, the closest playbook by shape is usually more useful than the closest by industry name.

Buy them all

The full Operators course for $347. Less than 3 hours of contractor time.

Operators tier. Lifetime access. Every new chapter free. Refund in 7 days if we're wrong.

How to read a playbook

One agent that holds down a real job beats a system that never ships

Each playbook covers a single job an agent can genuinely keep: support triage, sales follow-up, ops reporting, recruiting screens, invoice chasing. It is not a feature list. It is the workflow as it happens today, the source data the agent needs, the points where a human has to sign off, and the smallest safe version worth shipping first.

The framing is deliberate. An agent doing one narrow job well is the cheapest hire a business will make, and it earns the trust needed to expand. A grand autonomous system that touches everything is the version that gets built for three months and quietly retired.

What a playbook contains, section by section

A playbook here is not a product page for an agent. It is the operating document you would need to brief a builder, and it is written so that an operator who knows the process but has never automated anything can read it end to end and recognise their own week in it.

Each one runs through the same seven parts: the workflow as it happens today, including the steps people are embarrassed about; the source data the agent needs and where it currently lives; the trigger that starts a run; the actions the agent is permitted to take on its own; the points where a human has to approve before anything reaches a customer; the failure modes and what the agent should do when it hits one; and the smallest safe version worth shipping first.

The section people skip is the source data, and it is the one that decides the project. Most agent builds do not fail at the model. They fail because the records were inconsistent, the customer's name lives in three different formats across two systems, or the thing the agent needs to read is a PDF attachment somebody forwards manually. A playbook that sends you back to fix that first has done its job even if you never build the agent.

Bottom line: The source-data section is the part of a playbook worth reading first: agent projects fail far more often on inconsistent records than on model quality.

Five archetypes cover most of what businesses actually deploy

The sixteen playbooks under this hub are specific jobs, but underneath them sit five recurring shapes. Recognising which shape your problem is makes the build far more predictable, because the shapes fail in different ways and need different amounts of human oversight.

Inbox triage

Reads what arrives, classifies it, routes it, and drafts a reply for a human to send. This is the highest-volume archetype and the safest first build, because the human is already in the loop by design: the agent proposes, a person sends. It works on support queues, shared sales inboxes and general enquiry addresses. It is the shape behind the support, ecommerce support and email assistant playbooks.

Where it breaks: threads with long history, attachments that carry the real content, and anything where the correct answer depends on a system the agent cannot read. Triage that classifies well but drafts badly is still worth having, and it is fine to ship the classification and leave the drafting manual.

Receptionist and intake

Answers, qualifies, captures the details and books or escalates. This is the archetype with the clearest business case because the alternative is a missed call, and it is also the one with the least tolerance for error, because it is speaking to a customer live with no human in between. Legal intake is the strictest version: the qualifying questions are constrained by what a non-lawyer is permitted to ask.

Where it breaks: accents and background noise, callers who interrupt, and any question outside the script. A receptionist agent needs an escalation path to a human that works on the first attempt, not the third.

Scheduler and reminder

Owns the calendar side: confirming, reminding, rescheduling, filling gaps from a waitlist. Mechanically the simplest archetype and often the one with the fastest payback, because the value is in preventing no-shows rather than in anything clever. It is also the least model-dependent; much of it is ordinary automation with a small amount of language on top.

Where it breaks: two-way calendar sync, timezone handling, and stop handling. If someone replies STOP to a reminder, that must be honoured everywhere immediately, and this is a compliance obligation rather than a nicety.

Research and enrichment

Takes a thin record, goes and finds the rest, and writes it back. Prospect enrichment, candidate screening, competitor monitoring and content research are all this shape. It is the archetype with the highest tolerance for imperfection, because a wrong field in a research row is cheap to correct and nothing customer-facing depends on it.

Where it breaks: confident invention. An enrichment agent that cannot find the answer will supply a plausible one unless it is explicitly permitted to return nothing. Every enrichment build needs an empty value to be an acceptable output and a confidence field the human can sort by.

Follow-up and chase

Notices that something has gone quiet and does something about it: an unanswered quote, an overdue invoice, a lapsed customer, a meeting with no notes sent. This archetype has the most direct relationship to revenue and the strongest case for automating, because the work is genuinely never done consistently by hand.

Where it breaks: knowing when to stop. A chase sequence that does not detect a reply, a payment or a cancellation keeps going and does real damage to the relationship. The stop conditions matter more than the message.

Bottom line: Inbox triage is the safest first build because a human already sends the reply; follow-up has the strongest revenue case; a receptionist has the least room for error because it is talking to a customer live.

What one realistically saves, and how to work that out honestly

We do not publish a saving figure per archetype, and you should be suspicious of anyone who does. The number is entirely determined by your volume, and a page that quotes hours saved without knowing your volume has made it up. What we can give you is the arithmetic, which is short enough to do on the back of an envelope.

Take the number of times the workflow happens in a week. Multiply by the minutes it takes a person to do it once, honestly measured rather than remembered. That is your weekly hours. Multiply by your loaded hourly cost for whoever does it now, and by fifty. Then subtract the part the agent will not do: the exceptions, the review, and the cases it escalates. On a first build that residual is commonly a third or more of the total, and pretending otherwise is how projects lose credibility in month two.

The second half of the case is usually not labour at all. A missed call that becomes a booking, a quote that gets followed up before the customer buys elsewhere, an invoice paid three weeks sooner: these are revenue and cashflow effects rather than hours, and for owner-operated businesses they are frequently larger than the time saving. They are also harder to attribute, so measure the baseline before you switch anything on. Once the agent is running, nobody can tell you what the missed-call rate used to be.

The ROI calculator does this arithmetic and the cost pages give the build and running bands, so the two together produce a payback window you can defend to a partner or a bookkeeper without inventing anything.

Bottom line: Runs per week times honest minutes per run, minus the third or so the agent will escalate, is the only saving figure worth quoting, and you have to measure the baseline before you switch anything on.

Which of these fail without a human in the loop

Full autonomy is not the goal and it is rarely the right destination. The useful question is narrower: for this specific action, what does a wrong answer cost, and can it be undone? Sort the agent's actions by that and the approval points place themselves.

Actions that should stay behind a human approval on any first build: anything that sends a message to a customer in your name, anything that moves money or commits to a price, anything that writes to a system of record you cannot easily reverse, and anything a regulator would take an interest in. Legal intake, invoice and accounts payable work, and recruiting screens sit in this category permanently rather than temporarily; the approval step there is the design, not a training wheel.

Actions that are safe to automate early: classification, routing, drafting, enrichment into a staging field, internal notifications, and anything whose worst case is that a person has to redo it. These are where the volume is, so the common shape of a good first build is fully automatic up to the point of contact and human-approved at the point of contact.

The approval step also earns its keep as an instrument. Every time a human overrides the agent, you have found an exception nobody mentioned during scoping. Two or three weeks of overrides is the cheapest specification you will ever get, and it is what makes the reporting meaningful once the approval step comes off. Removing it before that is how an automation gets switched off after one embarrassing message.

Bottom line: Automate up to the point of customer contact and keep a human at it: the overrides in the first weeks are the exception list nobody gave you during scoping.

How to actually run one, from scoping to the point it is boring

The first job is choosing, and the criterion is repetition rather than annoyance. Pick the thing that happens several times a week, has a stable rule you could write on a card, and has an obvious place for a person to check the output. If two candidates qualify, take the one where you can measure the before state today.

Then spend a week doing nothing but the data. Get the records into one place, agree what a duplicate looks like, and write down what the agent should do when a field is empty. This is unglamorous and it is where a two-week build turns into a two-month one if it is skipped.

Build the smallest version with the human approving every action. Run it that way until the overrides stop teaching you anything, which usually takes a few weeks rather than a few days. Widen it one action at a time, and only widen the actions whose worst case you would accept without a person looking.

Then decide who owns it. An agent nobody owns stops working within a quarter and nobody notices for a month. Ownership means someone reads the failure log weekly, someone gets an alert when a run does not fire, and someone is accountable for the plain-language report on what ran and what it saved. That reporting habit is the difference between an automation that survives a busy season and one that quietly gets bypassed.

Bottom line: Pick on repetition, spend the first week on the data, ship with a human approving every action, then widen one action at a time and name an owner before you consider it done.

The same agent, different industry, different constraints

The archetypes are stable across industries but the constraints are not, and the constraints are usually what decides whether a build is a fortnight or a quarter. A reminder agent for a salon and a reminder agent for a dental practice are the same shape and a different project, because one of them is touching patient records and inherits a whole compliance posture with them.

Regulated intake changes the questions the agent is permitted to ask and who has to see the answer. Practices with clinical software often have limited or awkward API access, which moves the work from writing prompts to getting data out at all. Trades businesses tend to have their records inside field service software and their real bottleneck on the phone, so the receptionist and follow-up archetypes dominate. Professional services usually have the cleanest data and the strictest rules about what may leave the building.

The industry pages translate each playbook into one trade's vocabulary and carry the compliance notes that apply to it, and the automation cost pages price a specific archetype for a specific industry rather than in the abstract. If you know your trade, those are the faster entry point than the generic playbook, and they link back here for the mechanics.

Bottom line: The archetype is the same across industries; what changes is data access and what the agent is permitted to say, which is usually what determines the timeline.

The five archetypes, what they touch, and where the human stays
ArchetypeWhat it touchesWhere the human staysPlaybooks in this shape
Inbox triageShared inbox, helpdesk, CRMSending anything a customer will readCustomer support, ecommerce support, email assistant
Receptionist and intakePhone system, calendar, intake formEscalation on anything outside the scriptLegal intake, real estate, customer onboarding
Scheduler and reminderCalendar, messaging, waitlistRescheduling exceptions and stop requestsOperations, customer onboarding, real estate
Research and enrichmentCRM, web sources, applicant recordsAnything acted on before a person reads itSales prospecting, recruiting, content research, data analyst
Follow-up and chaseCRM, accounting system, calendarTone, discounts, and when to stopInvoice and AP, marketing campaign, meeting follow-up, agency reporting

How to pick and read one

  1. 1

    Pick the job your team already repeats weekly

    Repetition is the strongest predictor that automation pays off, stronger than how annoying the task feels. Work that runs twice a month rarely clears the build cost no matter how tedious it is, and that is a legitimate reason to leave it alone.

  2. 2

    Read the source-data section before anything else

    Most agent projects fail there rather than in the model. If the records the agent needs are inconsistent, spread across three systems, or locked inside somebody's inbox, the playbook will tell you to fix that first. Skipping ahead is how a two-week build becomes a two-month one.

  3. 3

    Ship the smallest version, then widen it

    Every playbook ends with a first version you can have running in about a week, usually with a human approving each action. That approval step is not training wheels; it is how you find the exceptions that no one mentioned during scoping, and it is what makes the reporting meaningful once you remove it.

A good fit if
  • Operators who know the process and need the automation shape.
  • Consultants scoping a build for a client in an unfamiliar industry.
Probably not for you if
  • You want copy-paste code. A playbook is the thinking; the templates are the artifacts.
  • Your process has no stable rules yet. Write it down by hand for a month first.

Where to go next

Book the free AI automation audit

A 30-minute call and a written opportunity report inside 48 hours. You keep the roadmap whether or not you hire us.

See the done-for-you service

We scope it, build it, hand it over with documentation, and send a recurring plain-language report on what ran and what it saved.

Automation by industry

The same playbooks translated into the vocabulary of one trade, with the compliance notes that apply to it.

What AI automation actually costs

Build bands, monthly running costs, and the salary comparison that decides most of these projects.

Customer support agent

The inbox triage archetype in full: classification, drafting, escalation, and where the human stays.

Invoice and AP agent

The follow-up archetype applied to cashflow, with approval rules for anything that touches money.

Sales prospecting agent

Enrichment done safely, including permission to return nothing rather than invent a plausible field.

Legal intake agent

The intake archetype where the approval step is permanent rather than a phase of the rollout.

Meeting follow-up agent

A narrow, low-risk first build with a checkable output, which makes it a good place to start.

Receptionist cost: dental practices

The receptionist archetype priced for one industry, with the compliance constraints included.

Receptionist cost: HVAC companies

The same archetype where the bottleneck is the phone and the records live in field service software.

Receptionist cost: law firms

Intake priced where what the agent may ask is constrained before the build starts.

Automation for HVAC and plumbing

These archetypes in trade vocabulary, with the workflows that actually pay off first.

Automation for dental practices

Recall, reminders and intake where clinical software and patient records set the constraints.

ROI calculator

Runs per week times minutes per run, converted into annual hours and a payback window.

Risk scorer

Sorts the agent's actions by what a wrong answer costs, which is how the approval points get placed.

Build guides

When the playbook has settled the shape, these carry the wiring with the configuration left in.

Common questions

What is the difference between an agent playbook and a build guide?

A playbook is the thinking: the workflow as it happens today, the data it needs, the approval points, and the smallest safe first version. It is platform-neutral on purpose, because the shape of the job does not change with the tool. A build guide is the wiring for one specific workflow on one specific stack, with the configuration left in. Read the playbook to decide what to build and where the human stays; read the guide when you are ready to make it run.

Which agent should a business build first?

Almost always the one whose work already repeats several times a week and whose output a person can check before it reaches a customer. In practice that means inbox triage or a reminder and follow-up agent rather than anything autonomous. Repetition predicts payback better than how tedious the task feels, and a checkable output means the first weeks of overrides teach you the exceptions instead of costing you a customer.

How much does one of these agents cost to build and run?

It splits into a build cost and a monthly running cost, and the two behave differently: the build is a one-off that scales with how messy your data is, while the running cost scales with volume and is usually dominated by the platform rather than the model. The cost pages carry the current bands, and the automation cost pages price specific archetypes for specific industries. Anyone quoting a figure without asking about your volume and your systems is guessing.

Do these agents replace staff?

In a small business they usually absorb the part of a role nobody enjoys and nobody does consistently: chasing, confirming, sorting, retyping. The honest framing is that an agent is the cheapest hire a business will make, not that it is a replacement for a person. Where a role is genuinely made redundant, that is a decision the owner makes with full information, and it should never be the sales pitch for the project.

How long does an agent take to build?

The build itself is usually the short part. Getting the source data consistent and agreeing what the agent does when a field is empty is where the time goes, and it is the difference between a fortnight and a quarter. A narrow first version with a human approving each action is realistically a couple of weeks once the data is in order. Widening it happens over the following month as the overrides tell you what was missing from the scope.

What happens when the agent gets something wrong?

It should be designed to be wrong safely rather than to never be wrong. That means an approval step in front of anything a customer sees, an explicit permission to return nothing when it does not know, a failure log a person actually reads, and an alert when a run does not fire at all. Silent failure is the dangerous mode: an agent that stops working is far more expensive than one that makes a visible mistake, because nobody notices for weeks.

Can one agent do several of these jobs?

It can, and it usually should not on a first build. A single agent with one job has an output you can judge, a failure mode you can name, and an owner who understands it. A system that touches five jobs is difficult to evaluate and tends to be abandoned rather than fixed when one part misbehaves. Ship one job, get it boring, then add the second as a separate flow that shares the same data.

Do I need a specific platform to follow these playbooks?

No. Every playbook here is deliberately platform-neutral because the workflow, the data and the approval points are the same whichever tool you build on. Once you know the shape, the platform directory and the comparisons narrow the tooling, and the choice usually turns on how the vendor bills rather than on features. Plenty of these agents run on a no-code builder with a small amount of language work on top.

What if my process is not written down anywhere?

Then write it down by hand for a few weeks before automating anything. An agent encodes rules, so a process with no stable rules will produce an automation that is wrong in a new way every week. Recording what actually happens, including the exceptions people handle silently, is the single most useful preparation for a build, and it frequently exposes a fix that needs no automation at all.

A playbook tells you what to build. If you would rather not be the one building it, we will do the whole thing and hand it over documented, with a recurring report on what ran and what it saved.