Quick answer
On 25 June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. An AI SDR pilot can die of all three. To stay out of that 40%, write four things down before the first email goes out: the scope (one job, one segment, capped volume), the success measure (qualified opportunities and cost per opportunity against a control group, not meetings booked), a time box (12 weeks of sending with checkpoints at weeks 4 and 8) and the kill criteria, plus the name of the person who makes the call.
What Gartner said, and when
Gartner published the prediction in a press release on 25 June 2025: "Over 40% of agentic AI projects will be canceled by the end of 2027." It names three reasons: escalating costs, unclear business value and inadequate risk controls.
Anushree Verma, the Gartner analyst behind it, described most agentic AI projects as early experiments driven by hype and often used where they don't fit. Gartner also said many use cases sold as agentic don't need an agent at all: use agents where a decision has to be made, automation for routine workflows and assistants for simple lookups.
Two details get lost when the number is repeated:
- It isn't about sales. The 40% covers agentic AI in general. An AI SDR, which picks prospects, writes and sends emails and handles replies, is just one of the clearest examples in a revenue team.
- It doesn't come from a poll of 3,400 companies. The release mentions a January 2025 poll of 3,412 webinar attendees, but that poll asked how much their organisations had invested in agentic AI (19% significantly, 42% conservatively, 31% waiting or unsure). The release doesn't say the 40% comes from it.
The same release named "agent washing": vendors rebranding chatbots, RPA and assistants as agents. Gartner estimated that only about 130 of the thousands of agentic AI vendors are real.
I set up cold email and LinkedIn outbound inside B2B teams, with 2000+ meetings booked for B2B clients, so I look at an AI SDR the way I'd look at any new sender on a client's domains: what can it break, and how fast would we notice?
What Gartner added in 2026
Three newer Gartner releases are more specific about why agent projects get dropped.
- 26 May 2026, governance. By 2027, 40% of enterprises will demote or decommission autonomous AI agents over governance gaps found only after something goes wrong in production. Gartner's fix: govern each agent by its autonomy level (observe, advise, act with approval, act autonomously) instead of one rule for all. Source.
- 17 August 2026, cost. Inference costs per agentic workflow will rise more than fivefold through 2028, because an agent uses far more tokens than a chatbot. Your vendor pays that bill, so ask how your price is tied to usage and what's fixed at renewal. Source.
- 29 September 2026, vendor-built agents. By 2028, 70% of enterprises will abandon agentic AI built for them by a vendor's forward-deployed engineers, stuck with rising costs and unable to change it themselves. Gartner's advice: an executive sponsor accountable for business outcomes, written terms on IP and knowledge transfer, plus an exit plan from day one that you actually carry out. Source.
For a sales team the message is plain: before you start, decide what the agent may do on its own, what it will cost at full volume and how you get out.
How an AI SDR pilot ends up in the 40%
| Gartner's reason | How it shows up in an AI SDR pilot | What to write down before day one |
|---|---|---|
| Escalating costs | The pilot price covers a small volume. Credits, data, extra mailboxes and someone's review time come on top at full volume. | The full monthly cost at the volume you plan to run and a ceiling per qualified opportunity |
| Unclear business value | The dashboard shows emails sent and meetings booked. Nobody checks whether those meetings became pipeline. | One primary measure, a control group and your definition of a qualified opportunity |
| Inadequate risk controls | The agent sends from your domains while nobody watches spam complaints, made-up facts or emails to existing customers. | Approval rules, an exclusion list, health thresholds and who checks them every week |
| No exit plan (from Gartner's 2026 advice) | The vendor set it up, nobody inside knows how it works and the renewal arrives before a decision. | An end date, the person who decides and what you keep if you stop |
All four are decisions someone skipped before launch, and you can make them in an afternoon.
Step 1: Scope the pilot to one job
Start with Gartner's own split. If every prospect in a segment gets the same three emails, you need a sequencer, not an agent. An agent earns its price where it has to decide something: which accounts to contact after a signal, which angle to use, how to answer a reply.
Then run the agent washing test. Ask what the agent does differently for two accounts that look the same on paper, when one just raised a round and the other just hired a new sales leader. If both get the same sequence, you're paying agent prices for automation.
A pilot scope I'd sign off fits in six lines:
- One job. For example, first emails and follow-ups to one segment. Interested replies go to a person.
- One segment, split in two. Half goes to the agent, half to your current process. That control group is what makes the result readable.
- A volume cap. This is how I set it up for clients: 20 to 30 emails per mailbox per day, after 3 to 4 weeks of warmup. An AI sender gets the same limits as a person.
- An exclusion list. Customers, open deals, partners and anyone who asked not to be contacted.
- Separate sending domains. Not the domain your invoices and support emails go out from.
- Approval first. The agent drafts and a person approves every send, which is Gartner's "act with approval" level, until the drafts stop needing edits.
Automate the work, not the thinking. AI owns the speed, you own the judgment: the agent can research, write and send, but who you target and what you offer stay with you. If the agent you're testing is the one built into your CRM or sales engagement tool, I checked what those agents actually do in Gartner's 40% task-specific agents forecast against 7 sales tools. For the brand guardrails themselves, see how to deploy an AI SDR without wrecking your brand.
Step 2: Write the success measures before day one
Meetings booked is the easiest number for an agent to hit and the least useful one to manage to. Pick one primary measure and a few health checks, and write their definitions down.
- Primary: qualified opportunities from the agent's half against the control half, using the definition your sales team already uses.
- Cost: everything the pilot costs (licence, credits, data, mailboxes, review time) divided by qualified opportunities, against the ceiling you set.
- Supporting: positive reply rate, no-shows and meetings held rather than booked.
- Health: bounce rate, unsubscribes, the share of drafts a person had to rewrite and the Gmail spam rate in Postmaster Tools, which Google says to keep below 0.10% and never let reach 0.30%.
The edit rate tells you more than the vendor's dashboard. If someone is still rewriting most drafts in week six, you've bought an assistant, and you should pay assistant prices.
Write down what result would make you scale it. "More qualified opportunities than the control, at a lower cost per opportunity" is a fine answer. "Let's see how it goes" is how pilots drift into the 40%.
Step 3: Time-box it to 12 weeks of sending
Cold outbound is slow to read: replies come in over weeks, and meetings take more weeks to become qualified opportunities. Too short and you kill an agent that works. Open-ended and a bad one lives on hope.
- Week 0: sign off the one-page plan: scope, measures, cost ceiling, kill criteria and the decision owner. New domains get 3 to 4 weeks of warmup first, outside the 12 weeks.
- Weeks 1 to 4: the agent drafts, a person approves every send. Track the edit rate and fix the instructions, not the individual emails.
- Week 4 checkpoint: health numbers within limits and edits falling? Let it send the steps it gets right without review and spot-check a sample every week.
- Weeks 5 to 8: run at the agreed volume and check spam rate, bounces and replies weekly.
- Week 8 checkpoint: first read on meetings held and qualified opportunities against the control.
- Week 12 decision: scale it or stop it. The only middle option is to fix one named thing and run 6 more weeks.
Hold the end date. Gartner's advice on vendor-built agents fits here too: carry out the exit plan you set at the start instead of extending because the team isn't ready.
Step 4: Agree the kill criteria and who decides
Kill criteria are what Gartner calls circuit breakers for autonomous agents: rules that stop the agent when a threshold is crossed, agreed before anyone is attached to the result.
Pause sending at once if
- The Gmail spam rate in Postmaster Tools climbs toward 0.30%.
- The agent emails anyone on the exclusion list: a customer, an open deal, someone who opted out.
- A message contains a made-up fact, a wrong name or a claim you can't stand behind.
- The bounce rate passes the limit you set in week 0.
Stop the pilot at a checkpoint if
- Cost per qualified opportunity is above your ceiling at week 8 and hasn't come down by week 12.
- The agent's half has no more qualified opportunities than the control by week 12.
- A person is still rewriting most drafts at week 8.
Write them in words nobody can reinterpret later:
"If the Gmail spam rate passes 0.2% on any day, sending pauses until [name] has reviewed it. If cost per qualified opportunity at week 12 is above [amount], or the agent's half has no more qualified opportunities than the control, we stop. We keep the domains, the lead data and the prompts."
Name the person who decides, and make it someone accountable for pipeline, not the person who chose the tool. Gartner says the same about vendor-built agents: the sponsor should answer for business outcomes, not for the implementation budget.
Finally, settle what you keep if you stop: domains, mailboxes, lead data, prompts and the playbook. Check that before you sign; my AI SDR contract checklist covers the clauses.
The one-page test. If you can't fit the scope, the success measure, the cost ceiling, the end date and the kill criteria on one page before the contract is signed, you're not ready to start the pilot.
Key takeaways
- Gartner predicted on 25 June 2025 that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls.
- The 40% doesn't come from the poll of 3,412 webinar attendees in the same release. That poll measured how much companies had invested.
- Gartner's 2026 releases add three more reasons to plan ahead: governance by autonomy level, rising inference costs and an exit plan for vendor-built agents.
- Before the first send, write down one job and one segment, a control group, a cost ceiling, a 12-week end date and the kill criteria.
- Judge the pilot on qualified opportunities and cost per opportunity against the control, and let someone accountable for pipeline make the call.
FAQ
What did Gartner predict about agentic AI projects being canceled?
In a press release dated 25 June 2025, Gartner predicted that over 40% of agentic AI projects will be canceled by the end of 2027, because of escalating costs, unclear business value or inadequate risk controls. The prediction covers agentic AI in general, not only sales.
What is agent washing, and how many agentic AI vendors are real?
Agent washing is Gartner's name for vendors rebranding chatbots, RPA or AI assistants as agents without real agentic capability. In June 2025 Gartner estimated that only about 130 of the thousands of agentic AI vendors are real.
Does Gartner predict companies will abandon vendor-built agentic AI?
Yes. On 29 September 2026 Gartner predicted that by 2028, 70% of enterprises will abandon agentic AI built by a vendor's forward-deployed engineers, stuck with rising costs and unable to change it themselves. Its advice is a sponsor who owns the business outcome and an exit plan from day one.
How long should an AI SDR pilot run?
About 12 weeks of sending, with checkpoints at weeks 4 and 8, plus 3 to 4 weeks of warmup before that if the domains are new. A shorter pilot rarely shows whether meetings turn into qualified opportunities.
How do you measure whether an AI SDR pilot worked?
Compare qualified opportunities and cost per qualified opportunity against a control group working the same segment. Track spam rate, bounces and how many drafts a person had to rewrite as health checks. Meetings booked on its own tells you very little.