How to choose an automation agency: what to ask first

Account and workflow ownership, the process inventory, error alerts, handover documents, data protection and a pilot: what to settle with an automation agency before you sign.

· 13 min readAI

Short answer

When choosing an automation agency, insist that the tool accounts, workflows and credentials sit in your name. The work should start with a process inventory, and every workflow should have an error alert, a log and a written guide. Before a long contract, ask for a short, measured pilot on one process.

One Monday morning the orders stop arriving in the accounting software. An agency built the workflow months ago; the person who built it no longer works there. The team first hunts for the tool the workflow lives in, then for the email address that tool was registered with, and last of all for whoever has the password. Most of the questions to ask when you choose an automation agency exist to spare you that morning. This article covers what is specific to automation work: ownership, process, error alerts, handover documents, data protection, AI steps, tool choice, fee models and the pilot. The general questions that apply to any agency are collected in how to choose a digital agency; we don’t repeat them here. We are an agency that does this work too, so we wrote the guide in a way that lets you read our own proposal by the same measure.

Who will hold the accounts, workflows and credentials?

An automation has three parts: the tool account (n8n, Make or Zapier), the workflow itself, and the credentials the workflow uses to connect to your other systems, meaning API keys and authorisations. If all three sit in the agency’s account, you have nothing that runs on the day you part ways.

A healthy arrangement looks like this:

  • The tool account is opened with your company email and you pay the subscription. The agency is a user invited to that account.
  • The CRM, accounting, email and marketplace connections are authorised from your own accounts. Where possible, a separate user or a separate API key is created for the automation, so a workflow doesn’t stop when an employee leaves and the key can be revoked on its own.
  • The AI provider’s API account belongs to you as well.
  • An exported copy of the workflows is handed to you regularly.

The last item is not enough by itself. According to n8n’s documentation, an exported workflow is a JSON file that includes the names and IDs of the credentials. The file carries the logic of the workflow; to run it again, the accounts it connects to have to be yours too. If the agency says “let’s build everything in our account, it’s more practical”, ask for a written description of what will be transferred, and how, if you leave.

Process first, tool second

Be careful if the first meeting goes straight to a tool name or a ready-made package. A good start is an inventory: who does what, how often, in how many minutes, and what happens when it goes wrong. In our own work that inventory takes one week and decides which task is automated first. We described the method step by step in where to start with automation; if you draw up your own inventory before meeting agencies, their proposals become easier to compare.

It is also a good sign when an agency is able to say: “Don’t automate this one.” A process that hasn’t settled and changes every week, or a job of a few minutes done a few times a month, is rarely worth the maintenance. To see which tasks typically get automated, look at the workflows in our business process automation examples.

Who hears about it when a workflow stops?

Automations break: an API changes, a password is renewed, a record arrives in a shape nobody expected. What matters is who finds out, and how quickly.

The tools offer ready-made ways to handle this, but none of them sets it up for you. According to n8n’s error handling documentation, you build a separate error workflow that starts with the Error Trigger node and select it in the settings of the workflow it should watch. The documentation adds that the same error workflow can serve several workflows and send a notification by email or Slack. Make’s help centre defines five error handlers and explains how each one ends the run: some skip the faulty record and carry on, others stop the run. Which one is chosen at which step is a design decision, and the agency should be able to explain it to you.

Ask the agency:

  • Does every workflow have an error notification, and which named person does it go to?
  • Where are the run logs kept, and for how long?
  • What prevents the same record from being processed twice?
  • Is there a written note on how to do the job by hand when the workflow is down?
  • How often are the running workflows reviewed?

“Just write to us if something goes wrong” doesn’t count as an answer; when a workflow stops silently, your customer is usually the first to notice.

What will you hold at handover?

A running workflow should not be the only thing handed over. You should have the process inventory, a page for each workflow describing what it does and which systems it touches, and a list of the accounts and connections in use. You should also have exported copies of the workflows and a user guide your team can read. The test of the guide is simple: could a developer who has never met the agency read a workflow from these documents and change it?

Where does your data go?

If a workflow carries customer names, phone numbers, orders or health information, every stop the data passes through matters. Ask the agency for a data map: which data passes through which tool, in which country it is processed, where it is stored and for how long.

As a general principle, you are the data controller for your own customers’ data; the agency and the tools it uses process that data on your behalf. Article 12 of KVKK (Türkiye’s Personal Data Protection Law, Law No. 6698) says that where data is processed by someone else on the controller’s behalf, the controller is jointly responsible with them for the security measures. Processing the data on a server abroad is subject to the separate conditions in Article 9 of the Law. GDPR draws a similar line between controller and processor. So there should be a data processing agreement between you and the agency that sets out the roles and the measures, and the agency should be able to name each sub-processor it uses. This paragraph is not legal advice; draw up the texts with your legal adviser.

If there is an AI step

In a proposal that includes AI, you want to see three things.

Where the model is used. When the input is regular and the rules are clear, a rule-based workflow is more predictable. A model is added to the individual steps that rules can’t cover, such as free text. “AI takes care of everything” doesn’t describe a design.

Where the human check is. In steps where a mistake is expensive, such as messages to customers, price and stock changes or payments, the final approval stays with a person. We don’t build assistants that answer your customers with no human check at all; if an agency says it will, ask who is responsible for a wrong answer.

What happens when the answer isn’t in the documents. An assistant that answers from your own documents should hand the matter over to someone on your team when it can’t find the answer, and should not guess. Ask to see how this was tested with real customer questions.

Read in the provider’s own documentation how it uses the data you send. OpenAI’s data controls page, for example, says that data sent to the API is not used to train models unless you explicitly opt in, while inputs and outputs may be retained for up to 30 days to identify abuse. Every provider’s terms are different; you should hear which provider the agency chose, why, and that it knows these terms.

Tool choice and lock-in

Each of the three common tools does good work; they differ in the unit they charge by, where the data is processed and who is left with the maintenance. The detailed comparison is in n8n vs Make vs Zapier. In a selection meeting, watch for these points:

  • Does the agency know only one tool, or does it choose by the job? It should be able to explain the choice in terms of your workflow’s volume and data.
  • Running n8n on your own server keeps the data with you, but it also gives someone the job of updates, backups and security. Who that person is should be written down.
  • Workflows don’t move from one tool to another by themselves. Changing tool means rebuilding the workflow. That makes the document describing the workflow’s logic a more lasting asset than the tool.
  • If the agency proposes a middle layer it wrote itself and only it can access, ask who will hold the source code and the hosting.

Which fee model is the proposal written in?

Automation work has three cost items, and they should appear on separate lines in the proposal (we compare agencies’ general fee models in digital agency pricing):

  • Project fee: paid once for a build with a written scope. How many workflows, which systems, which tests and which documents are included should be stated plainly.
  • Monthly maintenance: paid for following error notifications, reviewing logs and making small changes. What counts as maintenance and what counts as new work, and how quickly an error gets a response, should be defined.
  • Tool and API usage: the automation tool’s subscription and the AI provider’s usage charges. These should be paid straight to the provider, from your own account.

If the third item dissolves into the agency’s invoice, you can’t see what you are paying for, and the account is most likely in the agency’s name. Because the tools count usage in different units, such as tasks, operations or executions, the agency should be able to give you a rough usage estimate based on your workflow’s monthly volume.

A pilot before a long contract

There is no need to sign a twelve-month contract before you have seen a single workflow run. Ask for a narrow pilot: one trigger, one result. In our automation and AI service the order is this:

  • Week 1: the process inventory. Who does what, how often, in how many minutes.
  • Weeks 2 and 3: a pilot on the one process that takes the most time. The workflow first runs side by side with the manual job, then takes over the real work.
  • End of the month: the pilot is compared with the manual version of the job by the same measures: minutes per week, number of errors, time to completion.
  • After that: the result becomes a written decision to continue. The decision can be “continue”, “fix and continue” or “stop”; if the result holds up, the same model moves to the next process, and the error notifications and logs become permanent.

Saying “stop” at the end of the pilot should carry no penalty; the workflow and the documents built during the pilot should stay with you either way.

Red flags

  • A proposal with no inventory. A package that arrives before anyone has listened to your process is most likely a copy of a workflow built for another company.
  • Accounts in the agency’s name. If the tool, the connections and the AI account are with the agency, leaving means rebuilding.
  • No error alerts. “It has never broken so far” doesn’t count as monitoring.
  • Documentation left for later. A guide left for later usually never gets written.
  • The promise of an unattended assistant. An AI that answers customers with no check makes its mistakes in front of the customer too.
  • Promises with no measure. Lines like “we’ll halve your team’s workload” are a guess if nobody has measured the starting point.
  • No answers to data questions. An agency that can’t tell you in which country and with which provider the data is processed doesn’t know it on your behalf either.

Questions for the first meeting

The table below sets the twelve questions next to what a good answer sounds like, by the measure used in this article.

Question What a good answer sounds like
1. In whose name will the automation tool account be opened, and who pays the subscription? The account is opened with your company email and you pay the subscription; the agency is an invited user.
2. Which users and which access levels will the connections be set up with? They are authorised from your own accounts; where possible the automation gets a separate user or API key.
3. Do you start with a process inventory? Can I see an example? Yes; before any tool is named comes who does what, how often and in how many minutes.
4. By which measures will we judge the first workflow, and who measures the starting point? Minutes per week, number of errors and time to completion; the starting point is measured before the job is automated.
5. When a workflow fails, who is told, and through which channel? Every workflow has an error notification; it goes to a named person by email or Slack.
6. Where are the run logs kept, and who checks them, how often? Where the logs are and how long they are kept is known; reviewing them is part of monthly maintenance.
7. Which documents will I receive at handover? The process inventory, a page for each workflow, the list of accounts and connections, exported copies and a user guide.
8. Which tools will my data pass through, and in which country will it be processed? Do you sign a data processing agreement? You are shown a data map and a list of sub-processors; a data processing agreement is signed.
9. At which step will AI be used, and where is the human approval? The model runs only in the steps rules can’t cover; where a mistake is expensive, the final approval stays with a person.
10. Are project, maintenance and tool costs on separate lines in the proposal? The three items are listed separately; tool and API usage is paid straight to the provider, from your own account.
11. Do you run a pilot before a long contract, and how is the decision made at the end? A narrow pilot on one process; it ends in a written decision: continue, fix and continue, or stop.
12. If we part ways, how will the workflows, accounts and documents be handed over? The accounts are already yours and the agency’s access is removed; the contract says how and within what time the handover happens.

Ask for the answers in writing and put the agencies’ answers side by side.

Common questions

What is the difference between an automation agency and a freelancer?

Either can know the tools well. The difference usually shows in continuity: is there a second person who gets the alert when a workflow stops, are the documents written, who carries on when the person who built it can’t be reached. Put the questions in this article to both; accounts in your name and a written handover apply whoever you work with.

Can our own team build the automation?

Yes, especially workflows with one trigger and a few steps. Most of the tools can be learned by someone without a technical background. Outside help starts to make sense when a workflow touches money or customers, when data is transformed between several systems, or when error handling and logging are needed. Workflows you build yourselves still need an owner, an error alert and a short note.

Will our automations keep running if we leave the agency?

They will if the tool account, the connections and the subscription are in your name; removing the agency’s user access is enough. If the account belongs to the agency, the workflows have to be exported and rebuilt in your account and the connections authorised again. Have the contract say how and within what time that will be done, before you sign.

An automation inventory you can make today

Before the first meeting, set aside half an hour and write every automation running in your company today into a table. Note which tool it lives in, which email address the account was opened with, who pays the subscription, who is told when it fails. Every cell that is empty, or that says “not known”, “the agency” or “nobody”, is an item to raise with the agency. If you have no automations yet, ask the agency to fill in the same table for the first workflow it would build.

Automation Which tool Account opened with which email Who pays the subscription Who is told when it fails
Example: form to CRM record n8n The company’s shared address The company The sales lead
Example: order and stock sync Not known The agency’s address Inside the agency’s invoice Nobody

Sources

This article is general information, not legal, financial or technical advice. Platforms, their rules and the law change over time, so some details may have changed since 10 November 2026 or contain small errors. Before you make a decision for your business, check the current official sources or write to us.

Tell us about
your project.

Fill in the form or write to us directly, and we will agree on a day to talk.

Your first meeting is with Deniz Sezer, founder of Detmore.

First meeting free · Reply within one working day · Written proposal in 2 to 3 days · NDA ready

What do you need? pick as many as you like

Name, email and phone are required.

Contact

We reply within one working day.