01. The agent hype cycle vs reality
In 2025, every SaaS vendor has an “AI agent”. Most are wrappers around GPT-4 with a prompt. Real agents — those that reliably take actions, use tools, and complete multi-step workflows — are still narrow but incredibly valuable in specific slots. Sales is one of the highest-ROI slots.
02. Where AI sales agents actually win
The three highest-ROI deployments we've shipped: (1) Inbound qualification agents like our own Aria which pre-qualifies leads in 60 seconds. (2) Meeting-summary + follow-up agents that turn call transcripts into next-step emails in 30 seconds. (3) Outbound-research agents that build ICP-targeted lists with personalisation angles.
03. Where agents still fail
Complex negotiation. Emotional deal-saving conversations. Anything requiring nuanced political judgement inside a buyer's organisation. Also: any workflow where the cost of one bad AI action is high. Design your guardrails accordingly.
04. The right architecture: agents + humans
Best-in-class deployments give agents scoped autonomy: they can send certain emails, book meetings, tag CRM records — but escalate to humans for anything above a threshold. We call this “assisted autonomy”. It's the pattern that scales.
05. Cost, guardrails, and audit trails
Every AI agent we deploy logs every action to a queryable audit trail. Every action above a set risk score requires human approval before execution. This is non-negotiable for regulated industries.
Tags:AI AgentsSalesGPT-5Automation
Frequently Asked Questions
Can AI agents replace SDRs?+
They can replace the bottom 30–50% of SDR work: research, first-touch email, meeting scheduling. Not the whole SDR.
Which model should power a sales agent?+
GPT-5 for reasoning + JSON tool use. Claude Sonnet 4.5 for long-context CRM analysis. Use both if budget allows.
What is the payback period for an AI sales stack?+
For mid-market B2B teams: 60–90 days. For enterprise: 120–180 days. For SMEs: often instant.