Apollo's 2026 AI in Sales survey found that 84% of revenue teams use AI for prospecting and outbound, making top-of-funnel the dominant AI deployment in GTM by a wide margin. This is not an experiment anymore. It is the default motion.
And it is churning hard.
The easy explanation is that the AI writes bad emails. That explanation is wrong, and it is costing people a lot of money. The models improved every quarter for two years and the copy improved with them. A first-year rep can now send a thousand messages that each name the prospect's last funding round.
Something else broke. I put it this way to an investor on a call this week:
"An agent is only as good as the context. So if a prospecting agent does not have context of what happened prior in the GTM motion, you just 10x the inefficiency. And that's what you're finding in the AI SDR space. Massive churn, because the promise of the outcome, which was meetings, they're not getting. Because meetings is not a consequence of just doing an outbound. Meetings is a consequence of multiple things together. And if your AI SDR is just focused on the context in the prospecting motion, and not what happened earlier in the marketing motion, they're not effective."
Everyone is selling 10x productivity. Nobody mentions that the multiplier runs both ways. Point automation at a broken input and you do not multiply output. You multiply error, and it happens faster than anyone can catch it.
An agent is only as good as what it can see
Start with what the agent can actually reach. MIT Technology Review surveyed 300 data and technology executives this month. At the average company, AI can reach 45% of that company's data. For the laggards, 30% or less.
Read that with a sales lens. You hire a rep, hand them the CRM login, and walk away. They can't see the campaign history, or which of your emails this person already ignored, or that the account evaluated you a year ago and picked someone else. Then you set a quota and grade them on reply rate.
Nobody would do that to a person. It is roughly what the market did when it handed outbound to AI agents.
A meeting was never an outbound outcome
The AI SDR category sold one thing - MEETINGS. Not emails sent or personalisation at scale. That is what the pricing assumed and what the board deck promised.
But a meeting is a consequence of multiple things together, and almost none of them are the outbound email. Someone read something last month, or a peer brought you up, or they hit your pricing page twice in a week and then went quiet. Sometimes the person who downloaded your report last quarter has already moved to the company you are now emailing. By the time the message arrives, most of what decides the reply has happened somewhere the sales stack cannot see.
Point an agent at prospecting alone and you have handed it the last ten percent of a story and asked it to write the ending.
Watch a good rep before they type. First they check whether marketing already touched the account, and what the person read versus what they ignored. Then they ask a colleague whether we lost this logo before, and why. Only then do they write four sentences.
The four sentences were never the hard part. But we simply automated the four sentences.
The errors multiply, and you removed the thing that caught them
An agent failure is not like a software failure. Software fails loudly. It throws an error, the pipeline stops, someone gets paged at 2am. An agent missing context does not stop. It reasons cleanly from a wrong premise and hands the result to the next step as fact.
Fiddler's production data shows enterprise agents succeeding about 60% of the time on a single run. Measured across eight runs in a row, that falls to 25%. Chain three agents that are each right 70% of the time and the whole thing works 34% of the time. That is not a bug.
Humans currently hide a lot of this failure from us. Nearly half of marketers say they encounter AI errors several times a week. They notice that the company is wrong, the account already churned, the customer said no six months ago, or the recommendation makes no sense.
An autonomous workflow removes that person by design.
That can still be the right tradeoff. But you need to understand what disappeared along with the labor.
And it is about to get more expensive, not less
The market has not priced in the next part yet. This week, Gartner gave it a name: the Inference Paradox. Even as the price of a token keeps falling, the cost of running one agentic workflow will rise more than fivefold through 2028.
The reason is simple. A chatbot reads a question and answers it. An agent, in Gartner's words, must "constantly reason, negotiate, and question itself," and every one of those loops burns tokens. Route a task through an agentic reasoning model and you pay at least five times what a simple call costs, often far more as the task gets complex.
Put the two numbers next to each other. You 10x the inefficiency, because the agent is working from half the context. And you pay five times more to do it, because reasoning is the expensive part. Both multipliers are pointed the wrong way, and the pitch deck only showed you one of them.
The bill is moving to your buyer
Worse than the wasted budget is what this does downstream, as per Forrester B2B companies will lose more than 10 billion dollars this year to ungoverned use of generative AI, through incidents that hit enterprise value, legal settlements, and fines. That is the tail risk everyone is quietly carrying.
The number closer to home is smaller. Nineteen percent of buyers using AI tools say they came out of the evaluation less confident than when they went in, because the information the AI gave them was wrong or unreliable.
One in five buyers is leaving an AI-assisted evaluation less sure. We spent three years promising AI would shorten the buying cycle. For a fifth of the market it is doing the opposite, through the same mechanism breaking your outbound. A confident system, reasoning from partial context, producing something fluent and wrong.
Forrester expects the correction to show up as demand for humans. In 2025, only 17% of buyers found conversations with product experts meaningful in the final stage, against 30% who valued genAI tools. Forrester reads this as a reversal, as buyers go looking for someone who can check what the agent told them.
Where they work. Where they don't
AI SDRs work in narrow places. High-volume, low-consideration, short-cycle motions where the buyer's journey really does start with your email - the tools do about what they claim. The failures cluster in enterprise and upper mid-market, where the buying group is nine people, the cycle runs five months, and almost everything that matters happened before anyone picked up a phone.
I build a context layer for a living, so discount me accordingly. Test the argument against your own data. Pull your AI SDR's reply rate for the last two quarters, then pull the same rate for accounts where marketing had already reached the buying group. If those numbers come out close, ignore everything above. In the accounts I have seen, they have not, though my sample skews toward companies who called me because something was already broken.
What to check before you renew
The 2025 AI SDR contracts are coming up for renewal about now. Four things worth doing before you sign.
1. Measure what your AI can reach, not what you store.
The benchmark is 45%. List every system that holds a customer interaction, then mark the ones your agent can actually query at the moment it acts. Product usage, support tickets, call transcripts, paid engagement. Most teams find their number is worse than the benchmark, and that the unreachable half is exactly the half that predicts intent.
2. Count the links in the chain, then price them.
Write down how many agent steps sit between a trigger and an action that touches a customer. Do the compounding math: four steps at 70% each is a 24% success rate reaching real buyers with your logo on it. Then multiply that by cost, because each of those reasoning steps is now the expensive part of the bill, not the cheap one.
3. Audit one failure by asking what the agent knew.
Take a sequence that went badly. Ignore the copy and the tone. Reconstruct what the agent could actually see when it acted. Almost every time, the message was not wrong. It was written to a version of that person who no longer existed, or who had already said no somewhere the agent could not read.
4. Renew the layer, not the actor
Models are converging and getting cheaper. Agents are becoming a weekend build. The identity-resolved picture of how your company actually goes to market is the one asset in the stack that gains value the longer you hold it, and the only one a competitor cannot buy on Tuesday.
The layer underneath
One small moment from that investor call keeps coming back to me. His first comment, before we really started talking, was that there were more AIs in the meeting than people.
Three notetakers and two humans. That barely feels unusual anymore.
Every one of those bots is quietly building a well-organized memory of the conversation for somebody. Meanwhile, most GTM teams still cannot answer a basic question in one query:
"What has happened with this account over the last two years?"
Your buyer's AI may have better recall of your last four conversations than your own stack does. That gap is the part I would worry about.
We have spent the last two years putting increasingly capable actors on top of fragmented customer data. Adding another agent does not fix the fragmentation underneath it.
The agent was never the hard part. Building the context it can trust is.
P.S. If your AI SDR renewal is on the desk this quarter, run check number one before you sign and tell me the number you get. I am collecting them, and I suspect 45% is generous for most stacks. Reply here or reach me at [email protected]