To evaluate a sales AI agent before you sign, first define what type of tool you need, review five axes with one concrete question for the vendor, add up the total cost and not just the license, ask for a demo with your own conversations and decide with a pilot measured against your baseline.
A clarification before going on: at Vantegrate we build agents of this kind, so take this guide as what we would ask any vendor, including ourselves. You won't find a list of vendors or third-party pricing here, just criteria to compare using your own data.
The purchase often starts off crooked: demos are run with sample conversations, proposals arrive in different units and you end up choosing the neatest presentation. The full set of questions is in the checklist for choosing a sales AI agent, and the neutral description of each tool category is in the comparison of AI agents for sales on WhatsApp. This guide is the walkthrough for using them in order.
What type of tool do you need?
Three different things are sold under the name AI for sales, and comparing them with each other is like comparing apples to oranges:
Agent: talks to the customer, queries your systems, quotes, takes the order and hands off to a person when needed. What you buy is finished work.
Copilot: summarizes conversations, suggests replies and drafts messages, but the sales rep takes the action. What you buy is each person's productivity.
Analytics: reads what already happened and shows patterns, such as the most repeated objections or the reasons for lost deals. What you buy is information.
The expensive mistake is buying one when you need another. A copilot doesn't answer anyone at nine at night, and an agent won't explain why your quotes get lost. To get oriented quickly: if the problem is that nobody responds on time, you want an agent; if the sales rep responds but takes too long to do the work, a copilot; if you don't know what is going on, analytics.
If your doubt is between a menu bot and an agent that understands natural language, we cover it in chatbot or AI agent for selling on WhatsApp.
Five axes and the question to ask the vendor on each
Once the type of tool is clear, the evaluation is organized into five axes, each with a question that forces the vendor to show instead of promise:
Result: which indicator in your business does the agent move, and how will it be measured with your data? If the answer is another company's percentage of improvement, ask them to explain how they would measure it for you.
Integration: does it read and write in your CRM and your ERP, on your objects and with your permissions, or does it export spreadsheets? What happens when a write fails?
Answers: what does it do when it doesn't have a piece of data? Can you define what it may promise, such as discounts or lead times, and what it must hand off? We explain it in how to prevent an AI agent from making up prices or stock.
Cost: how does the bill change if conversations double in peak season or drop by half?
Governance: who adjusts the agent when the catalog changes, who can see what it did and why, and where is your customers' data stored?
For the first meeting, these five are enough. With two or three vendors in the running, the checklist breaks each axis into finer questions and builds the request for proposal with your priorities, so that all the answers arrive in the same format.
What decides the result on WhatsApp
Evaluation guides tend to think about email and phone, but in Latin America the customer writes on WhatsApp. There, seven points separate an agent that sells from one that puts your number at risk:
Official API: the agent must operate on the official WhatsApp Business API, with the number and the Meta account in your company's name, so you can take them with you if you change vendors. If you sell from the app today, see how to migrate to the API without changing your number.
Templates: to start a conversation or resume a follow-up, WhatsApp requires templates approved by Meta. Ask who writes them and under which category they are submitted.
Quality: Meta rates each number based on how recipients react, and blocks and reports lower it (AWS, on Meta's documentation). Ask how the agent protects that rating in follow-ups.
Meta cost: Meta charges for each delivered message according to its category and the recipient's country (Meta, 2026). Ask whether you pay Meta directly or the vendor with a markup.
Handoff: according to Gartner (2024), 64% of customers would prefer that companies not use AI in customer service, and the most cited concern, for 60%, is that it becomes harder to reach a person. Ask when the agent hands off, to whom and with what context; we detail it in handoff to a sales rep.
Regional Spanish: your customers write the way they talk, with voice notes and abbreviations. In Bogotá they ask you to "gift" them the price, in Monterrey to "set aside" the product, and in Buenos Aires they use a different word for a quote than in the other two cities. Require the demo to include messages like these.
Writing to your systems: the lead has to end up in Salesforce or Oracle with its source campaign, the quote linked to the opportunity, and the order loaded in the ERP, not in a spreadsheet. How each system connects is covered in selling on WhatsApp with the stock and prices from your ERP.
None of these points show up in a demo with sample data, and all of them show up in the first week of operation. Ask about them before you sign and see them work in the trial.
Total cost: what the license doesn't show
The monthly fee is the most visible line in a proposal and almost never the one that weighs the most. To compare proposals properly, break it down into these parts:
Implementation: mapping the process, connecting the CRM, the ERP and the payment methods, loading the catalog, and training the agent. Sometimes it is included and sometimes it is a separate project.
Subscription: per user, per conversation, per credit, or per platform. What matters is how it moves with your seasonal peaks.
Messages: they depend on your volume and on the countries your customers are in. The WhatsApp API cost calculator does the math.
Internal hours: reviewing responses, keeping prices and the catalog up to date, and handling what the agent hands off. They don't appear in any proposal.
Maintenance: who adjusts the agent when the commercial rules change, and whether that is billed separately.
Delay: every month the agent takes to go live, the sales you are losing today keep getting lost.
Once everything is added up, bring the proposals to a common unit, such as cost per resolved conversation or per order taken, with the same period for cost and benefit, and count the extra sales or the margin savings, not revenue. The pricing models are in how much an AI agent for WhatsApp costs, and the business case generator builds the case for leadership with the assumptions in plain view. Read the contract with the same attention: minimum term, exit conditions, and who owns the WhatsApp account.
Ask for evidence with your data
A demo with conversations prepared by the vendor shows that the product exists, not that it works with your customers. A trial that counts has two parts.
The first is a demo with your real conversations, without personal data and with the hard ones included: the long voice note, the "same as last time," the customer who haggles, the question about an out-of-stock product. Above all, watch what the agent does when it doesn't have the data: if it improvises a price in the demo, it will do the same with your customers.
The second is an integration test: the agent creates a lead, updates an opportunity, and leaves a quote in a test environment of your CRM, with the fields your team uses, and the order reaches the ERP. Ask what happens if the ERP doesn't respond at that moment.
That test also tells you whether your data is ready, which is less common than it seems: according to Gartner (2025), 63% of organizations don't have, or don't know whether they have, the data management practices AI needs. If your price lists live in scattered spreadsheets, it is better to find out in the trial than in production.
90-day pilot: baseline, control group, and gradual autonomy
The pilot is where most projects are lost. The MIT NANDA report on generative AI in business found that 95% of pilots do not reach production with measurable business impact (MIT NANDA, 2025). A pilot that isn't designed to reach a decision ends up like this: half working and without a number that justifies continuing or cutting it.
A three-month structure that tends to work:
Month 1, baseline: while the agent is being implemented, measure in your CRM how the process works today: first response time, qualified leads, quotes with follow-up, and sales. In our case implementation takes between 4 and 6 weeks, and the plan is in how a sales AI agent is implemented.
Month 2, one supervised use case: the agent handles a single use case, such as ad leads that arrive after hours, and someone on your team reviews a sample of conversations every day. If the vendor allows it, start with a person who approves the most sensitive responses before they go out.
Month 3, autonomy and comparison: where quality held up, the agent gains autonomy, and the results are compared against a control group: similar leads handled as usual, for example those from another region or another campaign, in the same period.
The final decision (expand, adjust, or cut) has to be written down before you start, with the metric and the threshold that trigger it. What to measure is in the KPIs for measuring a sales AI agent, and what usually goes wrong is in why an AI pilot dies in silence.
Build it in-house or get it turnkey
The last question is almost never asked out loud: who is going to build the agent and who is going to maintain it? There are three paths:
In-house development: your systems team builds the model, the connection to WhatsApp, the ERP and the CRM, the testing, and the maintenance. It gives you full control and requires people with the time to sustain it.
Platform: you pay for the software and your team designs the flows, loads the catalog and trains the agent. The fee is low; the internal hours do not appear in the proposal.
Managed service: the vendor configures, integrates, trains and maintains, and your team validates and approves.
One useful data point, with the caveat that it favors those who sell solutions, like us: in the MIT NANDA report, AI projects built with tools from external vendors reached implementation in 67% of cases, versus 33% for internal builds, according to the companies themselves (MIT NANDA, 2025). It does not measure sales agents in particular, but it matches something any IT department knows: the hard part is not building, it is maintaining.
To put numbers on your case, the self-service vs. managed comparator compares the first-year cost of a self-service tool and a managed service, including internal hours and the sales at stake while the agent is being fine-tuned. Sellium, our sales agent for WhatsApp, is delivered as a managed service: our team handles the configuration, catalog loading, integrations and training, and yours validates and approves.
If you want to ask Sellium these same questions, using your conversations and your systems, book a demo.
