How Can AI Improve Sales Prospecting ?
AI is good at two things here: finding accounts that resemble the customers you already win, and doing the research that makes outreach worth reading. It is dangerously good at a third — generating volume — which is how most teams destroy their domain reputation. Used for targeting and research with a person sending, it works. Used to send more, it stops working quickly. In the EU there is a harder ceiling than deliverability: Article 14 of the GDPR turns every enriched contact into a notice you owe that person.
Manual vs. AI-Assisted Prospecting
| Step | Manual Process | AI-Assisted Process |
|---|---|---|
| Building a list | Filters on industry and headcount | Accounts resembling the ones you actually close |
| Qualifying | Firmographic guesswork before any contact | Signals from hiring, product changes and public activity |
| Research | Fifteen minutes per account, or skipped | A brief per account, produced in seconds |
| Personalisation | A merge field, and everyone can tell | A genuine reason for the message, drawn from research |
| Timing | Whenever the rep gets to the list | Prompted by a trigger event worth referencing |
Model Your Wins, Not Your Ideal Customer Profile
Most ICPs are written in a workshop and describe who a company wishes it sold to. Your closed-won data describes who actually buys, and the two are often noticeably different.
A model trained on won and lost deals finds the attributes that genuinely predict a close — which are frequently unglamorous, like a specific tech stack or a particular growth stage. It also finds who you consistently lose to, which is just as useful and rarely discussed.
Run it against your existing pipeline first. If it ranks your current opportunities in an order your sales leader recognises, it has learned something real.
Volume Is the Trap
Every capability here can be pointed at sending more. It is the one application that reliably makes things worse — deliverability degrades, your domain gets flagged, and buyers who now recognise generated outreach at a glance discount everything from you.
The version that works uses the same technology to send fewer, better messages. Better targeting means a smaller list; better research means each message earns a reply. That is a harder sell internally than a tenfold volume increase, and it is the one that still works next year.
The Rules This Has to Satisfy
Outbound tooling routinely breaks two structural requirements of the CAN-SPAM Rule, 16 CFR Part 316. Every commercial message needs a clear and conspicuous valid physical postal address of the sender — a street address, a registered PO box, or a registered private mailbox. And the opt-out must not require the recipient to pay a fee, to provide anything beyond their email address and opt-out preferences, or to take any step beyond a reply or visiting a single web page, which rules out login-gated preference centres and multi-step unsubscribe flows. Because the Rule reaches whoever the message is sent on behalf of, per-sender domains and rotating inboxes do not partition the obligation: suppression has to be global across every sending identity.
In the EU and EEA, a list built from an enrichment vendor or from scraping triggers Article 14 of the GDPR, Regulation (EU) 2016/679 — a notice duty to each person, naming the source the data came from. That is the practical brake on volume this page keeps describing. Article 14(5) narrows it in four cases — the person already has the information, notice is impossible or a disproportionate effort, disclosure is laid down by Union or member-state law, or the data is covered by professional secrecy — but none of them describes routine outbound to a bought list. So in practice a hundred thousand enriched contacts is a hundred thousand Article 14 notices. The design implication is to keep provenance per field at ingestion, and to treat the notice as an automated pipeline stage with its own delivery record, rather than something an SDR is expected to remember.
Which Model We'd Shortlist for This
Rates below are per million tokens, input then output, taken from each provider's own pricing documentation and checked on 7 August 2026. Each name links to that model's page, where the source and the capture time are shown in full.
Gemini 2.5 Flash-Lite — $0.10/$0.40, and $0.05/$0.20 batched. Prospecting is a volume game where the per-call rate directly sets how wide a net you can afford.
Ministral 3 8B — Apache 2.0, with output priced the same as input at $0.15 per million tokens. That is the right shape for generating many short outreach drafts from thin inputs.
GPT-5.4-mini — $0.75/$4.50 across a 400,000-token window with no long-context surcharge published, so enriched account research can be pasted in whole.
Claude Haiku 4.5 — $1/$5 with a 200,000-token window for the drafts that go out under the company's name, and a 50% batch rate for the scheduled runs.
Where This Fits
This is one part of our work in AI for Sales & Marketing. See the full set of AI use cases for the equivalent in other industries and functions.
Frequently Asked Questions
Can AI write outreach that does not sound like AI?
Yes, when it has something real to say. Generated messages read as generated because they are built from nothing — no research, just a template and a name. Given a genuine reason for the contact, models write perfectly natural outreach. The tell is not the writing, it is the emptiness underneath it.
Should we automate sending?
We advise against it. The time between an approved draft and a sent email is seconds, so full automation saves almost nothing, and it removes the last check before something reaches a prospect. One badly judged message to a target account costs far more than the seconds saved across a whole quarter.
How do we avoid damaging our domain reputation?
Send less, and keep the standard high. Reputation damage comes from volume with poor engagement — messages ignored, deleted or marked as spam. A smaller, well-targeted programme with real reply rates protects the asset. Once a domain is flagged, recovering it is slow and it affects everything, including your invoices. Rotating domains does not solve it and does not help legally either: the CAN-SPAM Rule attaches to whoever the message is sent on behalf of, so your suppression list has to be global across every sending identity.
What signals actually predict buying intent?
The ones specific to your product, which is why generic intent data disappoints so often. Hiring for a role that implies your problem, a leadership change in the relevant function, a public product launch that creates the need — these are observable and meaningful. Work out which correlate with your own closed-won deals rather than buying a general score.
How does this fit with our SDR team?
It removes the research and list-building that consumes most of their day, and leaves the conversations. Expect the role to shift toward fewer, better-prepared touches. Teams that instead keep the same activity targets and just add AI end up with the volume trap, and their reply rates fall within a quarter.
Rank your current pipeline first — if the order looks right, the model learned something.
Get a Free Proof of Concept within weeks.
Let’s talk
Tell us what you’re
trying to build
The first 45-min alignment session — and a small PoC — are free.
Or just say hello or write us an email.