How to choose an AI agent vendor: eight questions to ask before you sign

You choose an AI agent vendor by what the agreement says about ownership, pricing, responsibility and measurement, not by the demo at the first meeting. Eight questions in writing separate a proposal you can sign from one that will leave you stuck: ownership of the agent and the conversation data, the fate of the number on the day you leave, what the billing is based on, who maintains the knowledge base, who detects an outage at 10 p.m., who is responsible for what the agent promises, and what you will measure in the third month.

How do you choose an AI agent vendor when all three proposals sound the same?

You compare them by what happens when the agent gets something wrong, when you want to leave and when the business changes, not by how well it talks in the demo. The software side of building an AI agent for a business has become widely available, and almost any vendor will stand up an agent that holds a reasonable conversation in Hebrew. What is not widely available is the operational side, so every AI sales agent project should open with an ownership and responsibility appendix, before a single line of code is written.

These are the questions worth sending in writing to everyone who gave you a proposal:

• Who owns the agent and the conversation data, and in which account?
• What happens to the number and the WhatsApp Business account if we part ways?
• Do you bill per message, per minute or per month, and what happens in a busy month?
• Who writes the knowledge base, and who updates it when something changes?
• Who detects an outage at 10 p.m., and how does a customer reach a representative?
• Who is responsible if the agent quotes a price we did not approve?
• Will the demo run on a generic scenario or on our data?
• What will we measure in the third month, and what is the criterion for stopping?

Questions 1 and 2: who owns the agent, the data and the number?

Ownership comes down to three places: the accounts with the infrastructure providers, the database where the conversations are stored, and the configuration files that make up the agent. If any of them is registered in the vendor's name, you are renting an asset, not owning it, and that is fine as long as you knew it before signing and not on the day you wanted to switch.

The three points where ownership is tested

First, the accounts: the messaging platform, the connection to the AI engine, the telephony, the hosting server and the database. Ask to see on a shared screen that you are listed as the owner of each one. Second, the data: a conversation with a customer is a sales record, what they asked for and which price made them hesitate. Ask whether you can export the messages and transcripts at any moment in a readable format, because a vendor that offers reports instead of a raw data export keeps you on a short leash. Third, the configuration: the behavior instructions, the knowledge base and the test suite. Finally, ask what you would receive if you ended things tomorrow, in what format and within how many business days.

The number and the WhatsApp account: the trap to know about in advance

The number and the business account should be registered to your business, with the vendor set up as a partner with working permissions, not the other way around. There is a technical detail here that is worth money: a number connected to the business platform stops working in the regular app, and going back is a deregistration process. Ask under which entity the account will be opened, who holds the two-step verification code, and what the process is for moving the number to another vendor. On the voice channel, find out in whose name the virtual number is registered.

Question 3: is the pricing per message, per minute or per month?

In a proposal you can compare, the payment is broken into three components: a one-time setup, a monthly payment, and variable costs based on usage. The third is the one that surprises people, so that is where you demand a breakdown of who counts, what is counted and what stops an overage.

It is worth knowing how the mechanism works on WhatsApp, because many proposals still describe it incorrectly. The messaging platform no longer charges per 24-hour conversation window, but per template message sent, at a different rate for each category (marketing, utility and authentication) and according to a per-country rate card. Messages within a window opened by the customer are not charged, and an inquiry from a click-to-WhatsApp ad opens a window at no charge. In other words, an agent that answers incoming inquiries generates a low platform cost, and it is proactive outreach that drives the cost up.

How to create a yardstick for comparison without guessing

Ask every vendor for three line items, in numbers: the rate for a marketing template message in Israel according to the official rate card, the cost of a call minute on the voice channel, and the cost of processing an average conversation in the AI engine. Then ask them to run those line items on two scenarios: the inquiry volume you have today, and three times that. The gap between the two results is the risk you are buying.

Compare the cost for a full year including setup, not a monthly price. We don't publish a public price list for agents, because the scope varies with the channels and integrations, but every proposal goes out with a written breakdown of the three components. You can see what a transparent price structure looks like on our SEO services price list, and that is the level of detail it is reasonable to demand here too.

Question 4: who writes the knowledge base, and who maintains it?

Whoever writes the knowledge base effectively decides what the agent knows. There are three approaches: a vendor that organizes the materials you provide, a vendor that interviews your salespeople, and a vendor that pulls content from your website and declares the job done. The third is the cheapest and the weakest, because your website contains what you tell a new customer, not what you answer a hesitant customer in the third conversation.

The follow-up question that reveals everything: what happens when a price changes on Sunday

Ask who makes the change, whether you can do it yourself in the interface or need to open a ticket, how long it takes to go into effect, and whether a log is kept of who changed what and when. Also ask whether a test suite runs after every update, because a change in one section can break an answer in another, and that comes to light through a surprised customer, not through the system.

Questions 5 and 6: who answers at 10 p.m., and who is responsible for what the agent promised?

An outage at 10 p.m. comes down to one question: do you find out about it from a system alert, or from a customer who writes in the morning that nobody answered them yesterday? Responsibility comes down to a similar question: does the agreement say who bears the damage when the agent states a price or a date?

What the monitoring layer should actually check

Here is what we run on our own agent, and it is the level of answer that is reasonable to ask for. A scheduled job runs every 15 minutes and sends a test message through the entire path, from the webhook to the reply. We learned the key point the hard way: a 2xx response from the messaging API means the message was accepted for sending, not that it was delivered, so the monitoring tracks the delivery status that comes back through the webhook. When something fails, the alert goes out to the business owner and to us at the same time, on WhatsApp and by email, and the response time outside office hours is written into the agreement and measured from the moment of the alert, not from the moment of the complaint. A layer like this is part of every AI agent setup.

Handoff to a representative: who sees what, and who goes quiet

Handing off to a person is a procedure, not a button. Define the triggers: an explicit request, a complaint, a threat to cancel, a request for an unusual discount, and a question that has no answer in the knowledge base. After the escalation, the representative needs to enter the same conversation and see the full history, the agent needs to go silent the moment a person has replied and not speak again until it is explicitly released, and every escalation needs an owner, not just a tag.

Hours with no representative need a separate rule. An escalation at 10 p.m. on a Friday should not disappear: the agent collects details, gives a time frame you can meet, and the escalation goes into a queue that opens on Saturday night, when Shabbat ends, while an urgent inquiry goes out as an alert to a phone. Don't guess the escalation rate: measure it in the first two weeks and set a target against your own number.

Responsibility for a commitment made in a conversation

This is the question that is almost always forgotten, and the only one that can hit your wallet. Ask for an operating rule and a liability clause. The rule: the agent does not state a binding price and does not confirm a delivery date, but gives a range or refers to a signed quote. For us, this is an iron rule on the voice channel, because a price said on a call sounds to the customer like a commitment. The clause: who bears responsibility if the rule was broken because of a configuration error on the vendor's side, and what the liability cap is. Also ask for the transcript to be kept, because without a record there is no basis for a discussion of who said what.

Question 7: what must a demo on your data prove?

A demo is worth something only if it runs on your real information and your hard conversations, live and not as an edited recording. It is not meant to prove that the agent speaks nicely, but that it knows how to stop, pass the conversation on and write a correct record into the system.

Three messages to send the agent yourself

The first: a question whose answer is not in the materials you provided, where the agent should say it is checking and not produce an answer that sounds plausible. The second: an attempt to squeeze out a commitment, for example an unusual discount or a timeline you never promised. The third: a real, messy message, with no punctuation and with spelling mistakes, and if possible also a voice message recorded while driving, because it is transcribed before the agent understands it, and background noise is where this breaks.

Also ask to see an action written into your system, for example a meeting booked in a test calendar, and then open the calendar and see the record with your own eyes. At the end, ask for the full transcript, because a vendor that hands it over without hesitation shows you what the day-to-day will look like after signing. You can arrange a test like this with us through the contact form on our website.

Question 8: what do you measure in the third month, and how do you structure the deal?

You measure four numbers against where you stood before setup: time to first response, the share of inquiries that got an answer, the share of conversations that reached the outcome you defined, and the share of escalations to a representative. You want to see the first three improve and the fourth go down over time, because it measures how fast the knowledge base learns.

A measurement baseline before going live

You can't compare without a starting point, so for the two weeks before launch, record how many inquiries come in each week and from which channel, how long it takes until a first answer, what share arrives outside business hours, and how many inquiries were never answered at all.

A pilot with an exit door and milestone payments

The protection against a bad answer is how the deal is built, not just the question. A paid pilot on one channel, with an end date and an acceptance test written before you start, lowers the risk for both sides. Split the payment into milestones: part at the start, part on delivery of the knowledge base and the configuration, and a final part only after the acceptance test has passed, where the test is a list of scenarios you run yourself. Add a checkpoint on day 30 and on day 90 and a stop criterion, for example that if the response time has not improved against the baseline you measured, you can end it. Ownership passes to you at acceptance.

What is different in the Israeli market, and what does Amendment 13 to the Privacy Protection Law require?

Two things are different here: Hebrew and a calendar that most platforms were not built for, and privacy regulation that has tightened since August 2025. You test the first in a 10-minute demo; the second goes into the agreement.

These are checks you can run yourself: write the channel's name in three forms, וואטסאפ, ווטסאפ and also ווצאפ (three common Hebrew spellings of WhatsApp), and see whether the agent recognizes all three; test the same word in full and in short spelling (Hebrew is written both with and without vowel letters); send an English product name inside a Hebrew sentence; write a phone number with and without hyphens; and ask for a meeting both as 3.4 (day first, the Israeli way) and as שלישי באפריל ("the third of April"). Then check the calendar: a meeting on Friday afternoon, the eve of a holiday, and Chol HaMoed (the intermediate days of Passover and Sukkot). An agent that was translated into Hebrew without being built for Israel will give a polite but wrong answer there.

Amendment 13 to the Privacy Protection Law, 1981, came into force in August 2025 and bears directly on the first question here. The vendor that stores your customers' conversations is in effect a database holder and you are the database owners, so it needs to be written down who counts as what, what the purposes of use are, which types of information are kept, what the security obligations are and who is allowed to audit. Add a security incident clause: who reports, to whom and within what time, since the reporting duties and the enforcement powers of the Privacy Protection Authority have been expanded.

Three more points are worth a line in the agreement: where the data physically sits, because the AI engine is usually outside Israel and transferring information abroad is subject to conditions; whether the content of the conversations is used to train models, and ask for the agreement to say that it is not; and how long recordings and transcripts are kept. On the voice channel, add an opening message saying that the call is recorded and handled by a digital system, and for proactive outreach, Section 30A of the Communications Law (Israel's anti-spam provision) also applies. This is not legal advice, and it is worth having your lawyer go over the appendix.

Finally, ask for an Israeli reference you can call, and ask them what went wrong the last time and how long it took to get a response, who found the problem, how many hours a month the agent demands of them, and what they asked to change and didn't get. We have gathered the practical side on the AI agent development for business page.

Frequently asked questions

What documents should I ask for before signing an AI agent development contract?

In most cases, three are enough: a specification document that sets out what the agent does and does not do, an ownership appendix that defines in whose name the accounts and the data are registered and who holds the database, and an operations appendix with a response window and an escalation procedure. If one of them is missing, you can add it as a short appendix. The real gap between proposals shows up at this stage, not in the price line.

How do I check an AI agent vendor's references without wasting a week?

Ask for two references whose agent has already been running for a few months, not a customer who went live last week. Ask one question: what happened the last time something broke, how long it took to get a response, and who found the problem. A detailed answer tells you more than any presentation.

What counts as a good pilot, and what happens to the deliverables when it ends?

A good pilot is defined in advance by its scope, its end date and an acceptance test written before you start, for example one channel with a list of scenarios the business runs itself. It is important to have in writing what happens at the end: whether the configuration and the knowledge base pass to the business, and whether the payment is credited against the setup. A pilot without an acceptance test drags on for months without a decision.

What should happen to the conversation data when the contract ends?

The agreement should state that the business receives a full export of the messages, transcripts and records in a readable format, within a defined number of business days and at no charge. It should also state what is deleted on the vendor's side, when, and what deletion confirmation is delivered in writing. This applies to recordings too, which are worth a shorter retention period.

How can I tell whether a proposal includes maintenance or only setup?

Check whether it spells out what the monthly payment includes, what counts as additional work, and how many change requests a month the payment covers. A proposal that talks only about setup leaves open the question of who handles the updates, and that is the work that starts after the agent meets real customers. A follow-up question: how long a small update takes and who does it.

The bottom line

The difference between proposals for an AI agent is almost always in the ownership, the pricing breakdown, the responsibility during the hours when nobody is in front of the screen, and the metrics agreed in advance. The eight questions here take less than an hour, and asking for written answers filters out some of the proposals before the second meeting. If you add a pilot with an acceptance test and milestone payments, even a bad answer will cost you no more than a month.