What to ask an AI automation provider before you sign (12 questions)

Before signing with an automation or AI company there are six answers you need in writing: who owns the code, where the workflows run and whose account that is, whose name the WhatsApp number sits under inside Meta, who pays for the AI API and with which key, what happens the day you stop working together, and how the result is going to be measured. If the provider answers without hesitating and puts it in the proposal, good sign. If the answer is that this gets sorted out later, the problem is not today: it shows up fourteen months in, when you want to change providers and find out you own nothing.

The 12 questions, ordered by how much it hurts to get them wrong

1. Who owns the code and how do you hand it over to me?
2. Where do the workflows run and whose name is that account under?
3. Whose name does the WhatsApp number stay under inside Meta?
4. Who pays for the AI API and whose name is the key under?
5. What happens the day we stop working together?
6. What SLA do you offer and what happens if you miss it?
7. How are we going to know whether this worked?
8. What does the bot do when it does not know the answer?
9. Where are my customers' conversations stored?
10. How much does it cost to change something three months after go-live?
11. What happens to the bill if volume multiplies by five?
12. What are you recommending I do not automate?

The first four are about ownership and they are the ones that leave you trapped. Questions 5 to 8 are about operations. Questions 9 to 12 are about judgment, and they are the ones that separate a provider from a salesperson. The six answers at the top of this article are questions 1, 2, 3, 4, 5 and 7: the ones no proposal should leave out, whatever the size of the project.

1. Who owns the code and how do you hand it over to me?

Good answer: the code, the workflows and the documentation are yours from day one and are delivered exported into a folder or repository of yours, not just running on the provider's server. In n8n that means the JSON of each workflow, which you download from the workflow's own menu with the Download option.

The detail almost nobody asks for: that JSON carries the logic but not the credential values. Add the list of configured credentials, which account each one uses and who the account holder is. Without that you have the blueprint of the house, not the keys.

Red flag: that the code is their intellectual property. It is legitimate if they sell a product with a monthly license, but then the proposal should say so and the setup should be priced accordingly.

2. Where do the workflows run and whose name is that account under?

Good answer: one of two, stated clearly. Either they run on a server contracted in your name (Hetzner, DigitalOcean, Hostinger, AWS) or in an n8n Cloud account you own, with the provider coming in with permissions; that VPS costs between USD 6 and USD 20 per month and you pay for it. Or they run on the provider's infrastructure inside a subscription that includes hosting, backups and monitoring, which for most SMBs works out cheaper and calmer. Neither one is wrong; what cannot happen is that you do not know which one you are buying. Ask literally: if next month I want to move this, what do I have to do and how long does it take?

Red flag: that they do not know which server and which country your automation runs in, or that the hosting sits in the personal account of someone on the team instead of the company's.

3. Whose name does the WhatsApp number stay under inside Meta?

Good answer: the number and the WhatsApp Business Account stay inside your Business Manager, and the provider gets access as a partner. It is the reverse of what happens by default when someone is in a hurry: creating everything in their own Business Manager is about twenty minutes easier on setup day, and it is the decision that does not undo itself later.

It weighs more than the code question because workflows get rebuilt in weeks and the number your customers have had saved for six years does not. Transferring a WABA between Business Managers is requested from WhatsApp Manager, but it requires the current owner to approve it and the process is measured in business days. If the relationship ended badly, you depend on someone who no longer takes your calls.

Red flag: being told that they take care of that. Ask to see the partners screen of your own Business Manager with the provider listed there, not as the owner.

4. Who pays for the AI API and whose name is the key under?

Good answer: the OpenAI, Anthropic or Google key is in your account, with your card and with a hard monthly limit configured. For an SMB with around 1,000 conversations a month, a cap of USD 50 to 80 leaves plenty of room and covers you against the only scenario that is genuinely scary: a workflow that goes into a loop on a Saturday night and burns three months' worth in eight hours. If the provider uses their key and bills it inside the monthly fee that is not wrong either, but they have to be able to tell you at month end whether the AI spend was USD 12 or USD 90.

Red flag: AI included with no declared cap and no way to see consumption. At some point there is an uncomfortable conversation about that number, and you will have it with all the information on the other side of the table.

5. What happens the day we stop working together?

Good answer: concrete steps and deadlines. 30 days' notice from either party, delivery of the workflow exports and the knowledge base in an open format (JSON, CSV), transfer of ownership or of access where it applies, and a handover call with the incoming team. It should be in the contract, not in a goodwill email. Ask it with this exact sentence: if next month I hire someone else, what do I take with me and what stays?

Red flag: discomfort. Anyone who tenses up at the exit question has a retention model built on friction. Second red flag: a mandatory 12- or 24-month commitment on a monthly service. If the service holds up on its own, it does not need a lock.

6. What SLA do you offer and what happens if you miss it?

Good answer: hours and channel. Full bot outage, first response in under 1 hour within the on-call window; a query or minor tweak, response in under 4 business hours; the channel an email address or a group with more than one person in it. And what happens when it is not met: a credit, a free month, something with a financial consequence. An SLA without a penalty is not an SLA, it is an intention, and the penalty does not need to be big to change the order in which you get attended to when there are three incidents at once. Ask also who answers: if it is one person's name and that person takes their holidays in January, your bot takes its holidays in January.

Red flag: phrases that cannot be verified. That they reply fast, that they are always there. Ask for that to be translated into hours and into a channel with more than one person.

7. How are we going to know whether this worked?

Good answer: a provider who asks you for two weeks of baseline data before starting. How many queries came in, how many were answered, how fast, how many ended in an appointment, a quote or a sale. Without that prior number, any later result is a well-written opinion.

Then, three or four metrics on a dashboard, and a Looker Studio or a Metabase connected to the conversation database is enough: percentage of conversations resolved without human intervention, which on repetitive queries usually starts in a range of 40 to 70 percent; time to first response outside business hours; appointments or leads generated; and how many of those leads reached the CRM with name, phone and reason for the query filled in, which is where a bot that answers well can still leave the manual work untouched.

Red flag: that the proposed metric is messages sent or conversations handled. They go up on their own over time and they do not distinguish between handling well and simply handling.

8. What does the bot do when it does not know the answer?

Good answer: it hands off through three different paths. When the confidence of the answer is low, when the customer writes a keyword (human, agent, I want to talk to someone) and when a topic from the blacklist you defined comes up: complaints, cancellations, anything involving money. The conversation lands in a shared inbox, whether Chatwoot or the inbox of whatever tool they use, and whoever picks it up sees the full history instead of starting from scratch. Outside business hours, the right answer is not for it to improvise: it is to say when a person will reply and leave the case queued.

Red flag: being told that the bot answers everything. A bot that never hands off will at some point invent a price, a deadline or an availability. In a clinic or a real estate agency that is a wrongly booked appointment or a reservation that does not exist.

9. Where are my customers' conversations stored?

Good answer: the provider tells you where each thing lives. The conversations in a specific Postgres database, at a specific provider and in a specific region; the knowledge base somewhere specific; who on their team has access and how it is revoked when someone leaves. If your business handles health or financial data, this answer decides whether you can hire them at all.

A point that often gets confused: when the automation uses the OpenAI or Anthropic APIs, those providers do not by default take the content of the calls to train models. That covers the model, not the platform in between. A copy of the conversations still gets written into the automation's database, and what matters is who owns that database.

Red flag: that the answer is that it is all in the cloud. Whose cloud, in what country, and who has the password.

10. How much does it cost to change something three months after go-live?

Good answer: a number of change hours included per month (2 to 4 is the usual in subscription plans) and a declared hourly rate for anything beyond that, which in LATAM moves between USD 40 and USD 80 depending on seniority. And an explicit distinction between a tweak and development: changing a piece of text, adding a question to the knowledge base or moving an opening hour falls inside the included hours; integrating a system that was not there before is quoted separately. It is the question that moves the most money over the medium term and the one almost nobody asks in the first meeting.

Red flag: that changes will be looked at as they come up. Translated: every request is going to be a negotiation, you are going to stop asking, and the bot stays frozen in its month 1 version while your business keeps changing.

11. What happens to the bill if volume multiplies by five?

Good answer: the provider separates the three bills and tells you which one grows with volume. Theirs generally does not, or it goes up in tiers declared in advance. Meta's does, because it charges per conversation initiated. AI does too, because it charges per token. A good provider gives you the cost per unit and not just the total: how much one more conversation costs. If today you handle 800 conversations a month and in high season you reach 4,000, the Meta and AI spend multiplies in that proportion while the provider's work barely changes.

Red flag: an all-inclusive, unlimited price with no fine print. Either it has a cap they did not tell you about, or it already assumes the worst month of the year and you are paying that worst month for the other eleven.

12. What are you recommending I do not automate?

Good answer: a short, concrete list of things that in your operation are better left as they are. What usually goes in there: negotiating a large quote, the first contact with a customer who is already angry, any process that still changes every week, and everything you do four times a month that takes ten minutes each time: that is 40 minutes a month against a flow that has to be built, tested and then maintained every month. At StriqTech this part takes up half of the 15-minute audit, and not out of modesty: every process we take out of scope is one less flow to maintain, and maintenance is exactly what gets billed month after month.

Red flag: that you throw any idea at them and they say yes to everything. They are not advising you, they are quoting you.

What answering questions 1, 2 and 3 badly really costs

A scenario built with typical ranges from the LATAM market, not with data from a specific client. A dental practice with three chairs, around 900 WhatsApp conversations a month, a bot that answers questions about health insurance plans and prices, books against Google Calendar and sends reminders 24 hours in advance. They signed for USD 900 of setup and USD 220 per month. At 14 months they want to change providers because requests take weeks.

Scenario A, they own nothing. The workflows live in the provider's n8n, the WABA is in the provider's Business Manager and the AI key is the provider's.

• Rebuilding the five flows with another team: USD 800 to 1,500 of new setup, because there is nothing to reuse.
• Recovering the number: a transfer request that depends on the current owner approving it. Business days if they cooperate; if they do not cooperate, the only real way out is to activate another number and tell the entire patient base.
• Conversation history: it stays on the other side if nobody asked for an export.
• Meanwhile the front desk goes back to answering by hand: between 8 and 12 extra hours a week for 2 to 4 weeks.

Scenario B, they own everything. Same bot and same money invested, but the server is in their name, the WABA in their Business Manager and the OpenAI key in their account.

• They give administrator access to the new team on the same n8n and the same Business Manager.
• The new team spends between 4 and 8 hours understanding what is there: USD 150 to 400.
• Days without WhatsApp: zero. The bot keeps running throughout the transition.

The difference between A and B, in this scenario, ranges from USD 400 to USD 1,350 depending on how much there is to reuse, plus 2 to 4 weeks of degraded operation on side A and none on side B. And it is not decided the day you want to leave: it is decided in the meeting where you sign, with three questions that take four minutes.

If the project is small, ask only four of these questions

Running all twelve in audit mode on someone charging you USD 400 of setup is disproportionate and it will leave you out of good small providers, which do exist. If the setup does not reach USD 1,000, stick to four: who owns the code (1), whose name the WhatsApp number is under (3), what happens if we stop working together (5) and what I should not automate (12). The other eight start to make sense when the setup goes above USD 1,500, when the automation touches billing or scheduling, or when two days of downtime cost you sales.

When the right answer is not to hire anyone

Three situations where none of these questions helps you, because the project should not exist yet:

• Low, scattered volume. Fewer than 50 queries a month and all of them different from each other. There is no repetition to capture, so the bot costs more than the hours it saves. What is missing there is demand, not automation.
• A process in motion. If you change the way you sell or the way you charge every two or three weeks because the business is still taking shape, automating means freezing a version you will not be using a month from now. Wait until the process repeats identically three months in a row.
• An accountability problem disguised as a tool problem. If nobody answers the WhatsApp because it is not clear whose job it is, a bot covers it up for about 60 days. Then it hands off to a human, that human still does not exist, and the customer gets twice as angry because now they actually were expecting an answer.

How we answer them

StriqTech's answers to the three that weigh the most, so this does not stay a list of demands nobody meets. The code and the workflows belong to the client and are delivered exported, with the list of credentials and who holds them. The WhatsApp number and the Business Manager always stay in the client's name, and we come in as partners. The exit plan is written into the proposal before you sign: 30 days' notice and delivery of everything in an open format. In the subscription model hosting is on us, and we say so explicitly so you know which of the two models from question 2 you are buying.

If you want the breakdown of what gets paid and to whom, it is in How much an AI WhatsApp chatbot costs in LATAM; to lay the project out over time, the 90-day roadmap. The 15-minute audit is free and it starts the opposite way round from what you expect: with what is not worth automating for you. Write to us at info@striqtech.com.