Bottom line up front: don't pick by feature-list length — pick by the answers to six questions. Every vendor's feature grid looks about the same, but the answers to the six questions below differ from vendor to vendor — and every one of them can be settled in a single demo.
First, tell the four supply types apart
The "AI for global trade" label covers a lot. Sort the market into four directions first — pick the wrong direction and everything after it is wasted:
- Outbound AI SDRs — they hunt new customers and blast cold outreach. They solve "where do customers come from".
- Reception agents — they catch the inquiries you already receive. They solve "the customer arrived — now what".
- Platform built-ins — like the assistant bundled with your B2B marketplace membership; check whether it's already included first.
- Generic support chatbots — scripted Q&A machines, too shallow for pre-sales reception.
If your problem is "inquiries already come in, but we can't keep up — replies are slow and the languages fall behind", the second type is what you're looking for.
Dimension 1: After launch, who operates it?
Be clear about whether you're buying a piece of software or an outcome. Tool vendors hand you a dashboard and a login: from then on, writing prompts, maintaining the knowledge base and fixing wrong answers are all your team's job. Fully managed vendors take those jobs over. Most stalled projects don't stall on the model — they stall on "nobody owns the ongoing maintenance".
How to verify: ask who configures everything before launch, and who fixes wrong answers three months in. Get both answers written into the contract.
Dimension 2: Where do its answers come from? Will it make things up?
Whether replies can be traced back to your own documents is the single hard metric in this category. One invented MOQ, certification or lead time can cost you a buyer outright. A properly built agent answers only what can be verified in your catalog, pricing rules and documents — and says plainly when it can't find the answer.
How to verify: run a blind test with five of your real historical inquiries and require a cited source on every reply. Then deliberately ask something your documents don't cover, and watch whether it bluffs or falls back to a human handoff.
Dimension 3: Whose accounts does it run on? Who owns the data?
Accounts and conversation history should stay on your side — otherwise you're renting your own customer relationships. Some vendors keep customer records on their platform: fine while it lasts, but the day you want to switch, you find the leads, chat history and tuned knowledge don't come with you.
How to verify: one question is enough — when the cooperation ends, can you export all leads and historical conversations in full, at no extra charge?
Dimension 4: Does it hold up in the languages you actually sell in?
"Supports 100+ languages" carries no information. General translation stopped being the problem years ago — industry terminology is the problem: tolerances, material grades, certification names, part numbers. That's exactly where general models drift, and exactly what buyers use to judge whether you're professional. What matters is bilingual semantic understanding grounded in your own product materials, not a language count on a slide.
How to verify: ask a technical question about your own product in the language of your biggest non-English market, then have someone fluent in that language grade the reply.
Dimension 5: When does it escalate to a human — and how?
A reception agent's most important skill is knowing when to stay quiet. Price floors, exclusive distribution, payment terms, contract wording — none of these are places for on-the-spot improvisation. A well-designed agent recognizes those topics, records the context, and hands the conversation — full history attached — to a specific person.
How to verify: in the demo, ask it directly about your price floor and exclusive distribution rights. If it answers instead of escalating, you have your answer.
Dimension 6: How is it priced? What's the real yearly cost?
Per seat, per conversation, per project — three pricing models put the risk in three different places. Per seat gets expensive the moment your team grows; per conversation gets most expensive exactly when your campaigns start working; per project or annual fees are predictable, but you must pin down in writing what "ongoing operations" actually covers.
How to verify: take your real monthly inquiry volume, project it across 12 months under each model, then ask: what does the bill look like in the month inquiries double?
Six signals it's time to walk away
- They promise a number of inquiries but can't explain where those inquiries come from
- The demo runs only on their sample data — they won't test live with your materials
- They can't say who's responsible when the agent answers wrong, or how it gets fixed
- Case studies are a logo wall — no specific industry, action or timeline
- Leads and chat history live on their platform and can't be exported in full
- The quote never spells out what "ongoing operations" actually includes
Two delivery models — which fits you better?
| Fully managed service | Productized self-serve (like TradovaAI) | |
|---|---|---|
| Getting started | The vendor builds and runs it for you, scheduled as a project | Sign up, connect your channels, go — try core features free first |
| Reply transparency | You can't see how answers are produced; you rely on the vendor watching | Review workbench plus cited sources — every draft is checkable and correctable by humans |
| Cost | Negotiated project fee plus a monthly retainer | Flexible model integration with transparent running costs |
| Data | Mostly lives inside the vendor's systems | Full self-hosted deployment available — data never leaves your network |
| Best for | Teams with nobody to run it who want to hand it over entirely | Teams that want to verify before committing and keep control of data and process |
Neither model is absolutely better — but if you want to see results before you decide and keep your team able to check what the AI is answering at any moment, productized self-serve is the steadier starting point.
One honest closing line
If your monthly inquiries are only a handful, your people can keep up — don't deploy any agent yet; straightening out your process is worth more than any tool. Come back to this page when time zones and languages have buried your inbox.