AI SDR: What the Evidence Actually Supports, and What It Cannot

Somebody will pitch you an agent this quarter that costs less than a graduate and never takes a holiday. Before you compare it to a salary, find out what has actually been measured. The honest answer is that almost nothing has, and the few numbers in circulation are older and weaker than the slides suggest.

This teardown is for a leader deciding whether to spend on autonomous prospecting software. It separates what the available research says from what the category claims, and ends with a test you can run on your own list in four weeks.

The Headline Being Sold

The pitch is straightforward: an agent performs top-of-funnel development work, so a team needs fewer people doing it. Kaleigh Moore's review for Close, published in November 2025 and updated in May 2026, defines the category as "software powered by LLMs and automation to perform sales tasks" covering research, outreach and booking. Her verdict on whether they work is yes, with an immediate condition attached: "With the proper guidance and human involvement." Close sells a CRM with its own AI features, so read the conclusion with that in mind. The figures below are not Close's own.

Where the Numbers Actually Come From

Chasing the citations back changes the picture. The adoption figures that circulate in this debate, usually attributed to a 2025 business-school article, trace to McKinsey's ninth annual B2B Pulse survey of 3,942 decision makers across thirteen countries. Digital Commerce 360 reported in September 2024 that 19% were already implementing generative AI for B2B buying and selling, with 23% building or experimenting. That is two years old, it describes implementation rather than results, and it is not about sales development specifically.

A second set of figures, about how many companies see AI driving innovation against how many can measure a profit effect, appears in the same discussion with no date and no sample attached anywhere we could reach. We have taken those out. A widely repeated single case about a wealth management firm, with a revenue lift and a preparation saving, is one company described second-hand, and we have taken that out too.

What remains verifiable is buyer research, and it is more useful than any of the above.

What Survives a Second Look

Adoption is not an outcome, and the adoption data is old

Most organisations have tried something. Far fewer can describe a result, and the most-quoted measurement of that gap is from 2024. If your team experimented, found it unremarkable and moved on, you are in the majority rather than behind. Treat any vendor who presents adoption statistics as evidence of effectiveness as having told you which one they could not produce.

The documented gains sit in preparation, not conversation

Every credible account of where this technology helps lands in the same place: account research, drafting, record-keeping and meeting preparation. None of that is a replacement for a development rep. It is time handed back to one, and whether that time converts into pipeline depends on what the rep does with it. Our template for the account brief a rep will actually use is the concrete version of that handover.

Buyers still rate a person higher exactly where deals move

Gartner surveyed 645 B2B buyers between August and September 2025 and compared their experience of sellers against generative AI. Buyers were 32 percentage points more likely to credit a rep with making them confident in the decision, and 21 points more likely to say a rep helped quantify the benefit for their organisation, with similar gaps on understanding needs and advancing the purchase. The release is on Business Wire because Gartner's own site refuses automated access, so treat the figures as quoted from the syndicated copy.

One line from that release deserves more attention than the headline it was published under. Buying groups with low dysfunction were 13x more likely to report a high-quality deal, and buyers who spent more time with supplier reps reported the least dysfunction. If that relationship holds, cutting human contact at the top of the funnel is not a cost saving further down. Our piece on why buyers take AI answers back to a person works through the same pattern from the buyer's side.

Four Things the Evidence Cannot Carry

• No independent trial exists. We looked for a controlled comparison of agent-generated and human-generated meetings with a stated sample and sponsor, and did not find one we could verify. The conversion and cost figures circulating in this category come from vendors and name neither sample nor timeframe.
• A forecast is not a measurement. Gartner predicts that AI will begin 95% of seller research workflows by 2027, against under a fifth in 2024. That is a projection about research tooling, and it says nothing about whether anyone is employed to do the rest of the job.
• Adoption surveys do not segment development work. A company implementing generative AI somewhere in buying and selling may be summarising support tickets.
• The interested parties fund most of the research. Check who paid before you quote the number in your board pack.

A Four-Week Split Test on Your Own List

1. Take two hundred accounts of comparable quality and split them evenly. Same segment, same offer, same period.
2. Give one half to a development rep and the other to the tool, with a named person reviewing and sending every message, which is what Close itself recommends during a rollout.
3. Run four weeks without changing the offer on either side.
4. Count qualified meetings per hundred accounts, then count how many became opportunities an account executive accepted. The second number is the one that matters.
5. Log every human hour on both sides, review time included, and nothing else.

Measure acceptance, not bookings. A system that fills calendars with weak meetings transfers the cost to your account executives and shows up as a pipeline problem two months later, which is why the handoff standard in our five questions that make a handoff easy is the right acceptance test. For the structured version of the pilot, see our two-week pilot design for prospecting agents.

A Reviewer Is Headcount. Count It.

Here is the position. Every serious recommendation in this category, including from the vendors themselves, puts a human on the send button and asks them to review each message. That person is headcount. If a tool needs an hour a day of review to stay safe and on-brand, the saving is not a role removed, it is a role changed, and the business case has to be written that way or it is simply wrong.

This matters because the two framings produce opposite decisions. Headcount removal justifies an annual contract signed centrally. Time returned justifies a pilot measured against what the reps did with the hours. Run the second and you will usually find the gain is real but smaller and differently shaped than the proposal said, which is a much better position than discovering it after renewal.

Proof would look like a published trial, with a stated sample and a disclosed sponsor, showing accepted opportunities per hundred accounts from an unsupervised system matching a supervised one. Until somebody runs that and publishes it, the reviewer is part of the cost, and anyone selling you otherwise is selling the part of the product that does not exist yet. The role itself is also moving, which is the subject of our explainer on what a development rep actually does.

What We Would Say in a Pipeline Review

The evidence supports a changed job, not a removed one, and the strongest verified data points the other way entirely: buyers still rate human contact higher on the moments that advance a decision. Use the tools where they are documented to help, keep a person accountable for what gets sent, and judge the pilot on accepted opportunities rather than meetings booked. Results vary by market, list and offer, and nobody reviewing this evidence can promise you a number.