The arms race that worked, then stopped working
Since GPT-4 made personalization at scale cheap, the playbook was clear: pull better data, write a sharper hook, add an AI layer, and reply rates would climb. The math worked because the gap between "personalized" and "generic" was real and visible to a buyer.
That gap closed in 2025. The modern enrichment stack pulls the same data from the same sources, while every team plugs the same LLM into the same prompt patterns, so the output converges. Today, three reps from three different vendors will send the buyer three emails that read like the same email — because, structurally, they are.
Even after the AI personalization wave rolled out across the GTM stack, the trend hasn't reversed: 95% of cold emails generate no reply, and the ones that do see an average reply rate of just 3.43% — down from 5.1% the year before, per Martal Group's 2026 benchmark. AI didn't fix the funnel; it commoditized the part of the funnel reps thought was a moat.
What buyers actually pattern-match now
Open any director's inbox in 2026 and the pattern is unmistakable: the same opener structures, the same proof-point formats, the same three-bullet bodies repeating across vendors. The "{first_name}, I noticed {recent_event}…" opener and the "We help {persona} solve {pain}…" frame have become so common that buyers don't have to read the message to recognize it — the structure does the work.
This is what the next generation of AI personalization is up against. Better data and tighter prompts produce a slightly more polished version of a template buyers already recognize, which means the arms race is happening at a level of detail the buyer never gets to.
The signal AI can't fake
There's exactly one signal in the buyer's inbox that doesn't compress into a recognizable template: a real person the buyer trusts saying "talk to this team." AI can do an enormous amount of useful work around that signal — writing the message, scoring accounts, mapping who in your team's extended network knows whom (which is what Via does) — but it can't manufacture a relationship that doesn't already exist. Your CEO's investor either knows the buyer or doesn't, and no model release will change that. AI can surface the warm paths your team already has, but it can't generate trust that isn't there.
This is why warm outbound response rates have held — and pulled away — while cold has collapsed. Warm outreach response rates reach 10–34%, compared to 1–5% for cold email — a 5–10x improvement, and 82% of B2B buyers trust coworkers and internal management as information sources, making them the most-trusted source of all. AI can imitate the words around that trust, but it can't generate the trust itself.
Signals tell you when. Warm paths tell you how.
Some shops are still winning with cold + AI, and it's worth being honest about who they are. The most common cases are teams selling into a niche the buyer hasn't been targeted to death in yet — early-stage category buyers, narrow vertical roles where the AI-generated noise hasn't fully arrived. Another set is teams whose personalization involves real human judgment in the loop rather than pattern-fill. A third group has product positioning sharp enough, timed to a pain the buyer is actively feeling, that the cold message cuts through anyway. That one is the hardest of the three to engineer — landing positioning cleanly across an entire target list is genuinely hard, and most teams don't get there even with strong PMF. If that's you, keep doing what's working. For everyone else — selling to ICPs every competitor is also targeting — the math has shifted, and the answer isn't to stop using AI; it's to use it for the part it's actually good at.
AI and signals are still the best tools for figuring out which accounts deserve attention right now: job changes, fundraising rounds, hiring spikes, product launches, and intent data all surface the right target at the right moment. The mistake is treating those signals as the whole job, when they're really only half of it. The other half is access. Once a signal tells you an account is in-market, the next question is how you actually reach the buyer, and that's where signals and warm paths combine: signals tell you when to act and who to act on, and warm paths tell you how to land the meeting once you've decided to act. Teams that pair the two outperform teams running either play in isolation.
Why the gap widens as AI gets better
Here's the part that surprises people: better AI doesn't fix the AI outbound problem — it deepens it. Every model improvement makes generic outreach incrementally more "personalized" and incrementally more uniform across vendors, so every team gets the same lift at the same time, and the buyer experience is more polished AI-written emails arriving in higher volume with the same trust profile as before, which is to say none.
Meanwhile, the value of a real warm path goes up rather than down, because as inbox noise rises and cold reply rates fall, the relative gap between "an email from a stranger" and "an email vouched for by someone the buyer trusts" widens with every model release. The teams that figured this out early aren't optimizing prompts harder — they're shifting from prospecting to pathfinding, starting with the buyer they want to reach and working backward through the relationships their team already has.
What to actually do
The shift isn't subtle, but it is specific. It comes down to three changes:
- Use AI for triage and prep, not for the message itself. AI personalization at scale has commoditized — even teams that haven't adopted it well are facing buyers numb to the output — so treat it as table stakes for research, prioritization, and prep, not as the wedge that wins the meeting.
- Treat your team's network as inventory. Advisors, investors, customers, board members, and former colleagues are the only outreach input your competitors don't also have, so map that graph the way you'd map any other pipeline asset.
- Pair signals with paths. Use AI and signals to decide which accounts are worth pursuing this week, then check for a warm path before you write a single email — and if a path exists, use it. If it doesn't, you can still go cold, but go in knowing your starting reply rate is 3.43% and your trust starting point is zero.
The AI outbound arms race ends the same way every arms race ends: a new asymmetry emerges that the old weapons can't reach. The new asymmetry is your team's relationships.