What AI Actually Changed About the SDR Role

For the past two years the standard prediction about sales development has been that AI would remove the role. The work looked like an obvious candidate. It was repetitive, high in volume, and easy to measure: build a list, find the addresses, write the email, send it, process the replies that say no.

Most of that has now been automated. The role has not disappeared. What changed is which part of it produces the result, and that is worth describing carefully, because it determines how sales and contact center teams should be staffed, trained and measured from here.

What became cheap

Three costs collapsed at roughly the same time.

List building came first. Data providers, enrichment services and scraping layers made it possible to assemble a few thousand contacts in an afternoon, with job titles, company size and technology stack attached.

Copy came second. A model will produce a first email, three follow-ups and five variations of each in the time it takes to describe the audience. The output is competent: grammatical, correctly structured, and close to what a person would send.

Sending capacity came third. Infrastructure that used to need a dedicated technical owner is now a paid configuration, with warming, rotation and deliverability monitoring handled by the platform.

The effect on our own work is visible in how often we launch. Through 2025 we were putting out two to four campaigns a month. Since February 2026 the same team has been launching between eight and sixteen, with no additional headcount. Preparation stopped being the constraint.

Why cheap stopped meaning better

Every one of those costs fell for everyone at the same time. The tools are sold on the open market at prices any team can pay, which means the advantage lasted exactly as long as adoption was uneven.

The consequence appears on the receiving end. The number of well-written, correctly personalised emails arriving in a director’s inbox has grown for the same reason your own output has grown. Attention did not grow with it, so each message competes with more messages of comparable quality and wins a smaller share of it, regardless of how well it is written.

We tested the obvious response and raised our sending volume substantially. The number of replies did not scale with it. More sending produced more sending, which is what moved our attention to the two things additional volume cannot fix: whether the offer is right for the segment, and whether the segment was the right one to begin with.

The offer did not get cheaper

A model will rewrite an offer in a dozen ways in a few seconds. What it cannot do is decide which problem is worth naming for a particular audience, because that depends on why your customers actually buy, and that information exists only in your own closed deals, lost deals and conversations.

This is the work that used to be squeezed between list building and sending, and it is now the main thing worth spending a week on. A precise offer for a narrow segment will outperform a general offer sent to a larger audience, and no amount of additional volume changes that ordering.

One thing that moved the other way

There is a counter-effect worth naming. When the default output of a language model became the median cold email, writing that is recognisably human became unusual again. A message with a specific observation, an ordinary sentence structure and no promotional vocabulary now stands out for the reason it used to be ignored.

This does not restore copy as a source of scale. Careful writing does not multiply the way volume does, and a good sentence sent to the wrong person still fails. What has changed is the floor: competent writing is now the baseline everyone has, and anything below it is immediately recognisable as automated.

Signals age faster than they used to

The same logic that ended the volume advantage applies to buying signals. Funding rounds, executive hires and open vacancies are pulled by everyone from the same sources, which is precisely why they have stopped providing an edge. By the time you write about a funding round, several similar emails are already in the inbox.

This is worth applying consistently, including to the signals currently considered clever. Any attribute that can be purchased or scraped will be purchased and scraped, and its value will decay on the same schedule. The signals that hold their value are the ones a company generates itself and nobody else can buy: what customers say in conversations, which problems keep coming up in support queues, how accounts actually use the product. Those are not available on any market.

Sorting replies is automated. Answering them is not.

Reply triage is no longer manual work, and it is worth being accurate about that. We classify incoming replies with a model: sentiment, intent, whether the person is the right contact, where the reply belongs. It does the job, and it does it before anyone opens the inbox.

What did not transfer is the answer itself. Over the past twelve months we categorised 3,704 replies to our campaigns, and direct meeting requests accounted for roughly one percent of them. Everything else arrived as a question, an objection, a request for detail, or a redirection to somebody else. Each of those has to be answered by a person who knows what is actually true: what the product does and does not do, what is flexible, which existing customer this situation resembles.

A standardised answer at that moment is worse than a slow one. Someone has just spent attention on you, which is the scarcest thing in this channel, and a template spends it for nothing. This is the one point in the entire process where a real buyer is paying attention, and it is the last place worth automating.

What this means for teams

Three consequences follow, and one of them is uncomfortable.

Team size and output are no longer proportional. A small team with accurate targeting will produce more pipeline than a larger team sending more messages to a weaker list. Said plainly, this means that for a given pipeline target, fewer people are needed than two years ago. The honest expectation is fewer roles in sales development, filled by people who are paid more.

The hiring criteria change accordingly. What matters is the ability to define a segment from evidence rather than from a database filter, to build an offer for that segment, and to read a reply correctly. Speed of execution is now the software’s job.

Training changes with it. Scripts matter less than review of real replies and real recorded calls, because that is where the information about the market actually is.

Looking ahead

We expect signals to keep moving toward behaviour and away from announcements, and to keep decaying once they become purchasable. We expect measurement to standardise around conversations held rather than activity completed. And we expect the gap between teams to come from the data they own rather than the tools they buy, since the tools are the same for everyone.

None of this makes the work easier. It makes it smaller and harder, concentrated in two decisions: who is worth contacting, and what to do with what comes back.