Hiring an AI automation agency can save your team hundreds of hours a year, or it can leave you with a pile of fragile workflows nobody understands, running on accounts you do not control. The difference is rarely the technology. It is the questions you ask before you sign.
We are on both sides of this. We run our own businesses on AI automation, using Claude connected to email, accounting, e-commerce and ad platforms, packaged skills for repeat work, and an n8n instance for scheduled workflows. We also build these systems for clients. This is the checklist we would want a buyer to use on us.
Start with your own homework
Before you talk to any agency, write down three things:
- The processes that hurt. Not "we want AI." Specific tasks: reconciling Shopify payouts to QuickBooks, answering the same ten customer questions, building the Monday sales report, qualifying inbound leads.
- How long they take today. Rough hours per week and who does them.
- What good looks like. Faster, fewer errors, available after hours, or simply off a senior person's plate.
An agency that does not ask you for this in the first meeting is selling a product, not solving your problem. If you are not sure where to start, our AI readiness assessment checklist is a good warm-up.
Questions about approach
What would you not automate?
This is the best first question. A credible agency will name tasks that should stay manual, or that need a human approval step, or that are not worth automating at your volume. If every answer is "we can automate that," be careful.
Which tool would you use for each part, and why?
Good automation usually mixes tools. Fixed, scheduled, high-volume steps belong in a workflow platform such as n8n. Steps that need judgment, such as reading an email and deciding what to do, belong with an AI model like Claude. Ask the agency to explain the split for your process. We walk through how we make that call in n8n vs Claude for workflow automation.
Be wary of an agency that uses one tool for everything, especially if they have a commercial reason to.
Where does AI make decisions, and where does a human approve?
Get this in writing for every workflow. Anything that sends to customers, moves money, publishes content or deletes data should have a clear rule: either a human approves it, or there is a documented reason it does not need approval and a way to audit it afterward.
Can you show me something similar you have built?
Ask for a live demo or a recorded walkthrough of a comparable system, with client details removed. Slides are not enough. You want to see what happens when an input is messy, an API fails or the AI is unsure.
Questions about ownership and access
These are the questions most buyers skip and most regret.
Whose accounts will this run on?
The workflows, AI subscriptions, API keys and hosting should be in your company's name, on accounts you control, with the agency added as a user. If the agency runs everything on its own accounts, you are renting your operations from them, and leaving becomes expensive.
Who owns what you build?
The contract should say you own the workflows, prompts, skills, code and documentation produced for you. Agencies may reasonably keep ownership of their general methods and reusable components, but you need a perpetual right to use and modify everything running in your business.
What access do you need, and can it be scoped down?
An automation agency will need access to some of your systems. Ask for a list. Then ask whether each one can be read-only, limited to specific folders or accounts, or time-limited. When AI connects to your systems through MCP servers, the permission scopes are set at connection time, so this is a concrete conversation, not a vague one. Our guide to what MCP servers are explains how that works.
How do we revoke access if we part ways?
There should be a simple offboarding list: remove the agency's user accounts, rotate any keys or tokens they had, and disconnect any connectors set up under their login. If they cannot describe this in two minutes, they have not thought about it.
Questions about security and data
Where does our data go?
Ask which AI providers and services will process your data, under what terms, and whether those terms allow training on it. Business plans from major AI providers generally offer stronger commitments than consumer plans. Get the specific answer for the services they plan to use.
How do you handle prompt injection?
When an AI reads outside content, such as inbound emails, web pages or uploaded files, that content can contain hidden instructions. A competent agency will explain how they limit what the AI can do after reading untrusted input, usually with narrow permissions and human approval on outbound actions.
Do you use third-party skills, templates or connectors?
Pre-built components save time but carry risk. Anthropic's own guidance for Claude skills is to use only ones from trusted sources and to audit any you did not write. Ask the agency what they reuse, where it came from, and who reviewed it.
Questions about maintenance
Automation is not a one-time project. APIs change, vendors rename fields, your process evolves and AI models are updated.
What happens when something breaks?
Ask how failures are detected. In a workflow platform, that means error alerts that tell a named person what failed and link to the failed run. Ask who gets the alert, how fast they respond, and whether that is included in the fee.
How will my team change things without you?
You should be able to make simple changes yourself. Ask for written documentation of every workflow, plain-language descriptions of every AI skill or prompt, and a short training session. If only the agency can change anything, you are locked in.
What is the ongoing cost, all in?
Get three numbers: the agency's fee, the software subscriptions you will pay directly, and AI usage costs. Usage-based costs can grow with volume, so ask for an estimate at your current volume and at double it. Prices for these tools change often, so the agency should show current pricing, not a figure from last year.
Questions about results
How will we measure success?
Agree on a baseline before work starts: hours spent, error rate, response time, or whatever matters for that process. Then agree when you will measure again. Give it enough time. A few days of results is not a meaningful read; a month or a quarter usually is.
What does the first 30 days look like?
A good agency starts small. One or two processes, running in parallel with the manual version, reviewed by the person who does the work today. Expansion comes after the first ones are stable.
What if it does not work?
Ask about the structure. Phased work with a clear stop point after the first phase protects both sides. Large upfront fees for a long roadmap shift the risk to you.
Red flags
- Guaranteed percentage savings before they have seen your processes.
- No questions about who does the work today.
- Everything runs on their accounts.
- No written answer on data handling.
- No plan for maintenance or failures.
- A demo that only shows the happy path.
A short scorecard
| Area | What a good answer includes |
|---|---|
| Approach | Names what not to automate, explains tool choices |
| Approvals | Written rules for sends, payments, deletes |
| Ownership | Your accounts, your IP, documented handoff |
| Access | Scoped, listed, revocable |
| Security | Clear data path, injection controls, vetted components |
| Maintenance | Alerts, named owner, documentation, training |
| Results | Baseline, measurement date, phased scope |
This is how we structure our own AI workflow automation engagements: your accounts, written approval rules, documentation your team can use, and a small first phase that has to prove itself before anything grows.
Frequently asked questions
How much does an AI automation agency cost?
It varies widely with scope, from a fixed fee for a single workflow to monthly retainers for ongoing build and support. Ask for the agency fee, software subscriptions and AI usage costs separately so you can compare proposals on the same basis.
Should we hire an agency or build in-house?
If you have someone technical with time to own it, in-house can work well, and an agency can still help with the initial design. If not, an agency gets you running faster. Either way, insist on documentation so the knowledge stays in your company.
How long does a typical first project take?
A single well-defined process can often be built and running in parallel within a few weeks. Larger programs take longer. Be skeptical of timelines quoted before the agency has seen how the work is done today.
What should be in the contract?
At minimum: ownership of deliverables, accounts held in your name, a list of systems accessed and how access is removed, data handling terms, support response expectations, and a phased scope with a clear exit after the first phase.
Will AI automation replace my staff?
In most mid-sized businesses we see it removing repetitive parts of jobs rather than whole roles. The people who did the manual work are usually the best reviewers of the automated version, so involve them from the start.