Guide

Before you hire an AI automation agency: is the workflow ready?

17 min read

By . Published . Last updated . Reviewed by Mavrin Labs on . How this guide was researched and checked.

A card reading "Before you hire an AI automation agency: is the workflow ready?" beside the Mavrin Labs logo
The card that shows when this guide is shared or linked.

A workflow is ready for automation when you can say what it does, how often it runs, what a good result looks like and who approves the steps that matter. This guide sets out the checks that answer those questions, and what to ask whoever offers to build it. Automate the wrong workflow and you build it twice.

The guide is written for owners and operations leads deciding whether to hire a builder. It covers readiness checks and the questions to put to a firm. Choosing a tool, agreeing the investment and building the workflow itself are covered elsewhere.

What is AI automation, and when is a workflow ready for it?

AI automation is a workflow that runs across the systems you already operate, with steps inside it that read, sort or draft the parts plain rules cannot handle. A workflow is ready when its steps are known, its input is available to software, and someone can say what a correct result looks like.

The stand-alone definition of AI automation sets the term out on its own page, with how it works and where it fits.

Readiness has little to do with how advanced the technology is. Readiness depends on whether the work has a shape. A task that runs forty times a week along familiar steps is a safe candidate, because you can test the build against real examples and measure the result afterwards.

A task that takes a different route every time it runs is not a safe candidate. Nothing stable exists to build against, and nothing tells you whether the finished system is behaving.

If you cannot describe the task in a page, the first piece of work is the writing, and no software will do it for you.

The terms below carry the rest of the guide.

Workflow
One task with a start and a finish, such as moving an enquiry into your customer records.
Baseline
How long the task takes and how often it runs today, recorded before anything changes.
Exception
An item that does not follow the usual steps and needs a person to decide.
Human approval
A step where a person signs off before the system acts, placed where a mistake would be hard to undo.
Language model
The kind of AI that reads and writes text, used for the steps that plain rules cannot handle.

Why it matters: most small firms have not started

Use of AI is still uneven among businesses. A national survey of businesses by the Census Bureau put the share using AI at 19.8% as of May 3, 2026, and found that fewer than one in five firms with four or fewer employees use it (Census Bureau report on business AI use, checked Saturday, October 10, 2026).

Most small firms have not yet started, which makes a readiness check the right first step.

The work a readiness check looks at is a large share of American employment. The Bureau of Labor Statistics estimated 17,753,430 jobs in office and administrative support occupations for May 2025, out of 155,495,730 jobs in all occupations (occupational employment and wage estimates, May 2025, checked Thursday, October 15, 2026).

About one job in nine is this kind of work, which is why the job you name first decides how much the rest is worth.

Which readiness checks can you run this week?

You can run six readiness checks this week without hiring anyone: name one job, watch and time it, inspect the input, define good and bad, place approval deliberately and plan for failure. Together they produce the written record a builder needs, and they often change which task you automate first.

  • Checks on the work itself

    1. Name one job. Pick a single task with a start and a finish, such as filing supplier invoices. A named job can be timed, tested and reviewed, while a goal such as tidying up operations grows until nobody can say whether it worked.
    2. Watch and time it. Follow several real items from arrival to completion with the person who does the task. Record how long each step takes and where items wait, because that is your baseline and where the unmentioned steps show up.
    3. Inspect the input. Look at the actual emails, forms, scans and spreadsheet rows, and note how many arrive in an unexpected shape. Input that software can read, with the key field filled in, makes a build straightforward.
  • Checks on the result and its safeguards

    1. Define good and bad. Write down what correct output looks like and which mistakes are tolerable. Without that line, nobody can test the build, and disagreements surface after launch instead of before it.
    2. Place approval deliberately. Decide which actions a person must approve, such as messaging a customer or paying a bill. Put approval where a mistake would be hard to undo and nowhere else, or the automation becomes a slower manual task.
    3. Plan for failure. Agree who is told when the workflow stops and how the task runs by hand in the meantime. A workflow that fails quietly, with items piling up where nobody looks, is how automation projects lose the hours they saved.
    Name one jobA single task with a start and afinishWatch and time itThe baseline, and the unmentionedstepsInspect the inputThe real emails, forms and rowsDefine good and badWhat a test can pass or failPlace approval deliberatelyOnly where a mistake is hard toundoPlan for failureWho is told, and how the task runsby hand
    The six readiness checks in the order you run them. The filled square is the check that decides where a person approves.
    A personapprovesMessaging acustomerPaying a billAny action that ishard to undoRuns withoutapprovalReading andsorting the inputSteps that areeasy to reverseWork alreadychecked againstthe tests
    Where approval sits: before the actions that are hard to undo, and nowhere else, so the automation does not become a slower manual task.

    The approval rule matches published security advice. OWASP's entry on excessive agency, sixth in its 2025 list for language model applications, advises requiring a person to approve high-impact actions before they are taken (OWASP entry on excessive agency, checked Thursday, October 8, 2026). Approval at those points is the control that still holds when the model gets something wrong.

    Timing the task is the check owners most often leave out, and the hour-value tool makes it quick. Enter the weekly hours you recorded, and it shows the hours a year and their worth at the published office wage, so the baseline is on paper before anyone scopes a build.

Which kind of system fits the task?

The kind of system that fits a task depends on how much the task varies and how much judgment it needs. Fixed repeat work suits a plain automation, mixed input suits an AI step, and varied work may suit an AI agent, software that decides the next step within set limits. Some tasks stay with a person.

Which kind of system fits the task?
Kind of system Best for Watch for
Plain automation the same steps every time, with clean input breaks when the input changes shape
Automation with an AI step mixed input that has to be read, sorted or drafted needs review of what the AI step produces
Custom agent work that varies case by case towards one goal harder to test, and needs tight limits
A person, for now judgment calls and rare, high-stakes work write the manual steps down first

The guide to agents that judge each case covers the third row in detail.

Businesses that adopt AI mostly do so to improve how their work runs. In the Census Bureau's report on its 2023 annual business survey, improving the quality or reliability of processes or methods was the most common reason given for adopting AI (45.8%) between 2020 and 2022 (Census Bureau report on technology use, checked Saturday, October 10, 2026).

What does a workflow need before it is automated?

A workflow needs four things before it can be automated dependably. They are a stable sequence of steps, input that software can reach, a field that identifies each item without ambiguity, and permission to read and write in every system involved. Missing any one of them turns a build into guesswork, so check all four.

  • Stable steps, with every branch, exception and its decider written down.
  • Input that arrives where software can reach it.
  • One field that identifies each item in every system.
  • Written permission to read and write in each system.

The sequence has to be stable rather than simple. Branches and exceptions are fine, as long as you can list them.

The identifying field matters more than it looks. If two systems hold the same customer under different spellings, every later step inherits that confusion, and the automation creates duplicates faster than any person could.

Access is the item people discover last. Permission to write into the accounting or scheduling system usually belongs to someone who never attended the scoping meeting.

Ask for the narrowest access that works through each API (the published doorway one program offers so another can read or send its data). The OWASP project on API security ranks broken object level authorization as API1:2023, first of the ten risks in its 2023 edition (the 2023 OWASP API security risks, checked Thursday, October 15, 2026).

How do you hire an AI automation agency to build it?

Judge a firm on how it handles the measurement conversation. Ask how you will agree the result before the build, how the firm records the baseline, what happens to items the system cannot handle, and who you contact when it breaks. The answers separate builders from sellers quickly, often within one meeting.

  • How will the result be agreed, and against what baseline?
  • Which steps does the system take on, and which need approval?
  • What happens to an item the system cannot handle?
  • Who builds it, and who do I contact when it stops?
  • Where do the rules live, so another builder could take over?
  • Does anything in my systems have to be replaced?

Ask the exceptions question twice if you have to, because the plan for exceptions is most of the real engineering.

Two short reads help before the first meeting. The guide on how to hire an AI consultant or automation firm, question by question gives the full question sheet and the answers worth hearing.

A common first question is what does an AI consultant do, which the FAQ answers in plain terms, including what this kind of firm is usually asked to deliver.

  • What should you ask about security and records?

    Ask, too, how the firm handles security as it builds. The secure software development framework from NIST, version 1.1 in its publication SP 800-218, groups its practices into four sets that a buyer can hold any builder to (NIST secure software development framework, checked Thursday, October 8, 2026).

    Ask what record the system keeps of itself. CISA's Secure by Demand guide, published with the FBI, tells a buyer to ask whether security logs come with the product and whether a hosted service retains them for at least 6 months (CISA questions for software buyers, checked Thursday, October 15, 2026).

  • How do you judge a firm's approach?

    Compare approaches rather than reputations. Some firms sell a platform and fit your work to it, some build on the systems you already operate, and some only advise. None of those is wrong, but they lead to different outcomes when your process changes, so settle which one matches how often your work changes.

    Mavrin Labs publishes its own AI consulting process, from the first conversation to the review date, as one yardstick for comparing proposals: any firm should be able to show its stages as plainly.

    The measurable results method sets out how Mavrin Labs agrees a target and a review date before a build, and the AI automation service is the one that matches a repeat task with a measurable baseline.

When is hiring anyone the wrong move?

Hiring anyone is the wrong move when the work is not ready to describe, runs too rarely to repay a build, or is already handled by a setting in a tool you own. In those cases the better first step is writing the task down, switching on what you have, or leaving the work with a person.

  • The task changes shape every time it runs.
  • It runs a few times a year, so setup never pays back.
  • A setting in a tool you already use would do it.
  • Nobody can yet say what a correct result looks like.
  • The people who do the work have not been asked.

Whether a task runs too rarely to repay a build can be tested with a rule rather than a feeling.

Count the hours the task takes in a year, with the hour-value tool if it helps, and set them beside the hours your own team will spend on the build, its tests and its upkeep in that year. If the first figure is not clearly the larger, leave the task with a person.

The readiness checks above are still worth running in each case. They take a week of attention, and the written record they produce is what any later builder, inside the business or outside it, will need first.

What mistakes undo automation projects?

Four mistakes often undo automation projects before the build is finished. They are automating a whole department, building before measuring, putting approval on every step, and telling nobody when the workflow stops. Each has a plain correction, set out below, and each is easier to avoid at the start than to repair after launch.

Automating a department
A goal the size of a department cannot be tested. Pick one job with a start and a finish, and automate that first.
Building before measuring
Without a baseline nobody can say whether the build worked. Time the task as it runs today before anything changes.
Approval on every step
Approval everywhere makes the automation slower than the manual task. Place it only where a mistake would be hard to undo.
Nobody told when it stops
A quiet failure loses the hours the automation saved. Name the person who is alerted and the manual fallback.

How has guidance on AI risk developed?

Guidance on AI risk has moved from general principles to practical frameworks a small business can borrow. NIST released its AI risk management framework on January 26, 2023 for voluntary use, added a profile for generative AI on July 26, 2024, and says the framework is now being revised (NIST AI risk management framework, checked Thursday, October 8, 2026).

The framework's four functions, govern, map, measure and manage, make a useful checklist for deciding what you will watch after launch. Neither the framework nor any other reference tells you which task to automate.

Read beside the readiness checks, the risk view changes the order of the list. A task whose mistakes are easy to undo needs approval at fewer points, can be loosened sooner and returns its hours earlier.

A task where every output is hard to take back keeps a person at each step. Among tasks that pass the checks, start with the one whose errors are easiest to catch and reverse.

Sources behind this guide

For systems that use a large language model (the kind of AI model behind chat assistants, trained on large amounts of text to write language), OWASP's list of the top risks for those applications describes failure modes in plain engineering terms.

The OWASP list's 2025 edition lists prompt injection (hidden or crafted instructions in text that steer a model with permission to act) first (OWASP Top 10 for LLM applications, checked Thursday, October 8, 2026).

The NIST frameworks, the three OWASP references, the two Census Bureau surveys, the employment estimates from the Bureau of Labor Statistics and the CISA buying guide are cited in the sections above. They are useful for deciding where a person stays in control, which is the readiness question most often skipped.

Each claim and its source

The table below pairs each claim the guide makes from outside evidence with that evidence and its publisher.

Each claim and its source
Claim Evidence Source
Few very small firms use AI yet AI use stood at 19.8% of businesses nationally as of May 3, 2026, and below one in five firms with four or fewer employees (AI use figures, checked Saturday, October 10, 2026) Census Bureau
Firms adopt AI mainly to improve their processes Improving the quality or reliability of processes or methods was the most common reason for adopting AI (45.8%) between 2020 and 2022 (2023 survey report, checked Saturday, October 10, 2026) Census Bureau
A person should approve high-impact actions The entry on excessive agency advises approval before high-impact actions are taken OWASP entry on excessive agency
Untrusted text can steer a language model The current list of language model risks places prompt injection first OWASP list of LLM risks
Repeat clerical work is a large share of employment Office and administrative support occupations are about one job in nine nationally BLS occupational employment estimates
Permission is the first API risk to design against Broken object level authorization heads the current list of API security risks OWASP API security risks, 2023 edition
A buyer may ask what record a product keeps of itself Secure by Demand tells buyers to ask whether security logs come with the product CISA questions for software buyers

Key takeaways

  • A workflow is ready when its steps, input and correct result are known.
  • Run the six checks yourself before you hire anyone.
  • Record a baseline before anything changes.
  • Put human approval only where a mistake would be hard to undo.
  • Ask any firm about exceptions, baselines and who answers when it breaks.

If you are still deciding whether to build, buy, connect or automate, the guide to choosing among the four paths works through that choice. If the work changes with every item, the guide to custom AI agents covers that case, and the checklist for joining systems covers the systems a workflow touches.

Readers who want to meet in person can check the list of areas served.

Handled well, repeat work stops consuming the week, and the hours it was taking return to the people you already employ.

Frequently asked questions

How long should a readiness review take before a build starts?

A readiness review takes as long as it takes to watch the work happen and write down what you saw. For most small-business tasks, that means following a handful of real items from arrival to finish with the person who does the job. If you cannot yet write the task down that way, the review is not finished.

What if the task is not documented anywhere?

An undocumented task is normal and not a reason to stop. Write the steps down as you watch them, in the order they happen, and expect steps nobody mentioned, such as a quick check in a second system. Those hidden steps are usually where the hours go. The written record is the first deliverable, whoever builds it.

Does my data need to be clean before automating anything?

Your data does not need to be clean everywhere, but it does need to be dependable in the fields the workflow reads and writes. Pick the task first, then look only at the records it touches. Check that the field identifying each customer or order is filled in and unique, and that dates and amounts follow one format.

Which tasks should I leave alone for now?

Leave alone any task that changes shape every time it runs. The same goes for work that depends on a judgment you would struggle to explain to new staff, work that happens too rarely to repay the setup, and work where an expensive mistake would go unnoticed. Those tasks are better served by making the manual version clearer first.

What should a proposal from an automation firm contain?

A proposal should name the task, its baseline, the agreed result and the review date. It should say which steps the system takes on, which steps a person approves, and what happens to an item the system cannot handle. A proposal that describes technology without naming a result is not yet a proposal you can judge.

What are the warning signs when choosing a builder?

The clearest warning sign is a promise of results made before anyone has looked at your work. Others are a refusal to record a baseline, no agreed review date, a plan that replaces systems your staff rely on, and a proposal that will not say where a person stays in charge.

Who should own an automation after it goes live?

One named person inside your business should own each live workflow. Ownership means that person is told when the system stops, knows how to run the task by hand in the meantime, and decides when a rule needs to change. The builder keeps the technical responsibility agreed in writing, but the owner is why a quiet failure gets noticed early.

The Mavrin Labs mark

Who writes these guides

Client work at Mavrin Labs is delivered by a team in mixed roles, led by our founder. AI tools help with research and drafting, and a person at Mavrin Labs checks each fact against its named source before a guide is published.

Enquiry form

Start with one workflow

Tell Mavrin Labs what is stuck and which systems it touches. A person at the firm reads every enquiry and replies by email.

Start a conversation

Fields marked with an asterisk are the ones a reply needs.

So a reply can address you properly.

The address the reply goes to.

Optional. Give one only if you would rather talk than write; the reply goes by email either way.

Optional, and useful when it says what the work has to fit around.

The result you want, and when you would want it working. Two or three sentences are plenty.

The task as it actually happens, roughly how often it runs, and which systems it touches. Two links at most, please.

Optional.

What you send goes to the Mavrin Labs inbox and is used to answer you, and a phone number, if you give one, is used for the same thing and nothing else.

A person reads it and replies by email.

If the form does not send, write to info@mavrinlabs.com instead; it reaches the same inbox.

Which hours would you take back first?

Name the one workflow that eats the most of your week. Mavrin Labs reads it, and the reply comes from the people who would build the fix.

Start a conversation