Zeliq's AI Sales Agent: The Buying Tests That Matter

Zeliq raised €7 million to launch an AI sales agent, and the round tells you the company can afford to build one. It tells you nothing about whether the agent should be allowed to write to your CRM or send from your domains. Those are separate questions, and they get answered by an acceptance test you design before the trial starts, not by a demo the vendor drives.

What Was Announced

TechFundingNews reported on 22 September 2026 that Zeliq, a Paris-based company, raised a €7 million seed extension. The Moon Venture led the round. New investors Julien Jean and Sylvestre Blavet joined, alongside existing investors Ora Global and Cleo Ventures. The company said the funding will support the launch of its AI sales agent.

That is the verified set of facts. Anything else you read about the product's capabilities this week is either the vendor's own claim or someone repeating it.

Our view: the round is worth weighing for one narrow reason. A vendor you hand a workflow to needs to be solvent long enough for the switching cost to pay back, and a fresh €7 million buys that runway. Weight it in the vendor-risk column of your evaluation, then score the product separately on the tests below.

Write the Approval Boundary Before You See a Demo

The single decision that determines how an AI sales agent behaves in your business is where the human click sits. Decide it in writing, before the sales process starts, because it is very hard to argue for a tighter boundary after you have watched an impressive demo.

Four positions are common, and they carry different risk:

Most teams evaluating their first agent should start at draft and queue, then move one position at a time after reviewing real output. An agent's failures are invisible at the point they happen and visible three weeks later in a reply rate you cannot explain, so each step up should follow a batch of output someone has actually read.

Put the chosen boundary in the trial agreement. If the product cannot be configured to sit at that position, that is the evaluation result.

Inspect Every Write Before You Grant It

An AI sales agent earns its keep by acting on your systems, which means it needs write access to at least some of them. Ask for a written list of every field it creates or changes, in your CRM, your sequencer, and your enrichment layer.

Specific things to establish:

• Which CRM objects and fields the agent can create, update, or overwrite, and whether it can overwrite a value a human entered.
• Whether agent-generated records carry a source marker so your reporting can separate them from human work later.
• What happens to a record the agent judged wrongly: is the change reversible, and is there an audit trail of the prior value.
• Whether the agent can enroll a contact into a sequence without a human enrolling them.

Run a real reconciliation during the trial. Take fifty records the agent touched, pull the before and after state, and check the changes by hand. Vendors rarely object to this. The exercise usually finds one or two field behaviors nobody expected, and it is far cheaper to find them across fifty records than across your whole database.

Prove the Sending Controls Against Your Own Setup

Sending is where an agent can do damage that outlasts the contract, because a reputation problem attaches to your domain and stays there after you have switched tools.

The obligations sit with the sending domain, not the software. Google's sender guidelines require every sender to authenticate with SPF or DKIM and to "keep spam rates reported in Postmaster Tools below 0.3%," with SPF, DKIM, DMARC and one-click unsubscribe required above 5,000 messages a day. If the agent sends from your domain, that spam rate is yours.

Establish which mailboxes and domains the agent sends from, and confirm they are yours. Confirm the volume caps, the ramp behavior when a new mailbox joins, and what the agent does when bounces rise. Ask what happens when a prospect replies with a negative response: does the agent stop the sequence, and does it write a suppression entry that every other tool in your stack can read.

Test the suppression path deliberately. Add a seed address to the do-not-contact list in your system of record, then ask the agent to build a campaign that would include it. If the address gets a message, the agent is not reading your suppression source of truth, and no amount of good copy fixes that.

Confirm You Can Leave With the Work

This test decides whether you are buying infrastructure you own or renting a workflow, and it is easiest to answer during the trial rather than during a cancellation.

Ask, concretely: if you cancel in month four, what do you keep? The prospect records, the message history, the reply classifications, the sequence logic, and the enrichment you paid for should all come out in a format another system can read. Ask for a sample export during the trial rather than a promise in the contract. A vendor confident in its retention will show you the file.

The same question applies to the mailboxes and domains. If the agent runs on infrastructure registered to the vendor, you are renting the warm-up history that makes your email land. Register domains and mailboxes in your own company name, and the agent becomes a layer you can swap rather than a dependency you cannot leave.

Score the Trial on Decisions, Not Volume

Set the success criteria before the trial and make them decisions rather than outputs. Useful criteria include the share of drafts a human approved without editing, the number of records the agent changed that a human had to correct, the reply quality against a baseline you already run, and the hours the team actually stopped spending.

Message volume is the metric a vendor will offer, and it is the least useful one. An agent that produced four hundred messages your team rewrote has cost you time.

Pick one campaign, one segment, and one boundary, and run the agent against a comparable segment your team runs the usual way. Four weeks of that comparison will tell you more than any category report about whether this product belongs in your stack.

If you want a second opinion on your ICP, your current stack, and whether an AI layer would help or just add another tool to manage, book a free ICP and campaign-fit discovery call →.