The short version.
An AI sales agent researches companies, writes to people and sends in your name. Before you buy one, ask eight questions. Each answer can be tested, not only trusted.
We sell such an agent, so read the list with that in mind. At the end we answer every question for SYNC, including where it falls short. One gap first: on 29 September 2026, 230 of the 570 dossiers SYNC had marked ready rested on a cited signal without a date.[7]
Questions about the research and the writing.
1. Where does the data come from, and how old is it?
Ask for the sources, and separately where email addresses and phone numbers come from. Then open the sources behind three accounts the vendor prepared. A link does not prove that the page says what the claim says. In a study of search engines that answer with citations, only 74.5% of the citations supported the sentence they were attached to.[1] Check two dates as well: when the page was read, and when the event on it happened.
2. What does the agent do when it finds no reason to write?
Ask how many researched accounts get no message, what reason is recorded for each, and to see some of them. An agent that writes to every account it finds has no rule here.
3. Who approves what, by default?
“A person approves” can mean the first message only. Ask what a new account starts on, who can change it, and whether follow-ups need approval too. Research on automation finds that people rely too much on decision aids that are sometimes wrong, experts included, and that training or instructions do not prevent it.[2] So ask what the reviewer sees next to the draft.
Questions about the sending.
4. How many messages can one person receive, and who decides?
Ask for the limit per person and per company, over what period, and whether a user can raise it. When someone says no, is the address closed for you only, or for every customer of the vendor?
5. Whose mailbox and domain carry the risk?
If the agent sends from your mailbox, complaints count against your domain. For mail to personal Gmail accounts, Google asks senders to keep the spam rate reported in its Postmaster Tools below 0.10%.[3] Ask who holds the mailbox connection, and whether the vendor ever sees your password. More in What mailbox providers actually require.
Questions about the vendor.
6. What does the vendor measure, and what does it refuse to publish?
Ask for every figure with its count, its period and the number of customers behind it. An open is counted when an image in the message is downloaded, and Apple says of its mail privacy feature:
“It also prevents senders from seeing if you’ve opened the email message they sent you.”
Two vendors of sending software publish the same average reply rate, to two decimals, which suggests one source and not two.[5][6]
7. Which costs grow with use?
Ask what uses the allowance, whether a team shares it, what happens at the limit, and whether data, verified addresses and AI are billed on top.
8. What happens to your data when you leave?
Ask how you get an export, how long deletion takes, where the data is stored, and which companies process it for the vendor.
How SYNC answers today.
The case for and against the approval step is in What a human approval is worth.
What speaks against this.
A vendor wrote this list, around questions SYNC was built for. A buyer who needs LinkedIn messages sent, a language beyond Dutch, English, German and French, or strong data outside the Netherlands should ask those questions first. SYNC falls short on all three.
Being strict has a price we cannot measure. Nobody knows what the 720 dossiers left alone would have done with a message. 6sense, a vendor, reports that buyers choose one of the four suppliers on their first shortlist 95% of the time.[8] A seller who waits for a reason may arrive after that list is made.
An approval step is not proof of a careful review. A study of 41 policies that require people to oversee government algorithms warns that such oversight can give “a false sense of security”.[9] It is about government, and we read only the abstract.
Our answers were checked against our code by us, and by nobody outside. Some will change. None says whether you may write to a given person at all. That is a question of law. No lawyer has read this for us yet, so we make no statement on it.
The questions come from the concerns buyers raise in public discussions of AI outreach, and from claims our own audit had to correct on this site.
Our figures come from one read-only count that day, across all workspaces together: 1,303 dossiers in five workspaces, by verdict and recorded reason, and 203 messages with a counted open in four workspaces, by the time to the first open. The workspaces include our own and test workspaces. The figures hold no history, and we publish no split by workspace. Each outside source is labelled with how far we checked it.
- [1]Liu, N. F., Zhang, T. & Liang, P. (2023). Evaluating Verifiability in Generative Search Engines.abstract onlythe abstract, read on arXiv; about search engines, not sales email
- [2]Parasuraman, R. & Manzey, D. H. (2010). Complacency and bias in human use of automation: an attentional integration. Human Factors 52(3), 381–410.abstract only
- [3]Google. Email sender guidelines. Gmail Help.read in the originalapplies to mail sent to personal Gmail accounts
- [4]Apple. Use Mail Privacy Protection on iPhone. Apple Support.read in the originalread on 30 September 2026
- [5]Instantly. Cold Email Benchmark Report 2026.vendor researchInstantly sells email sending software
- [6]Woodpecker. Cold Email Statistics.vendor researchrepeats the Instantly figures; not an independent source
- [7]iSyncSO. Dossiers by verdict and recorded reason, and recorded opens by time since sending. Measured 29 September 2026.our own measurementsee the method note
- [8]6sense (2025). B2B Buyer Experience Report.vendor research6sense sells sales and marketing software
- [9]Green, B. (2022). The flaws of policies requiring human oversight of government algorithms. Computer Law & Security Review 45.abstract onlyabout government algorithms
