Can an AI Agent Replace Your Sales Team? What Tools Like Harvey Actually Do
Not for most businesses today. AI agents can research accounts, build lists, draft personalised emails, time follow-ups, sort replies, and schedule meetings. People still decide who to target, judge what is safe to send, run real conversations, negotiate, and stay accountable. Even Harvey, the open-source agent behind the headline, holds every email for approval by default.
What the tool in the video actually is
The video, from the YouTube channel Woke Engineer, features Harvey, which its description calls an open-source AI sales agent on GitHub that uses Claude Code. The repository, ethanplusai/harvey, is MIT licensed and describes itself as an autonomous sales agent powered by Claude Code or Codex that runs on a Claude or ChatGPT subscription rather than per-token API billing. It is not related to Harvey, the AI company that builds tools for lawyers.
According to the README, Harvey finds businesses (by default from OpenStreetMap data), reads their websites without model calls, proposes signals that the user must confirm before it collects them, writes cold email sequences, classifies replies, and keeps a local dashboard. Three details from the same README matter for the headline: every outgoing email waits for approval by default, a deterministic pre-send gate rejects things like unrendered merge tags and overlong emails, and calendar auto-booking is listed under the roadmap rather than as done. The README also states that a fully autonomous AI cold dialer will not be built, citing US rules on artificial voices in calls.
In other words, the tool's own documentation describes a supervised assistant, not a replacement for a team.
The sales job, split into tasks
Asking whether AI can replace a sales team is too coarse. It helps to split the work and ask the question task by task.
Strong fit for agents today
Finding companies that match written criteria, reading their websites, enriching records, drafting first-touch emails from verified facts, scheduling follow-ups, stopping sequences on reply, classifying replies as interested, not now, or opt-out, and offering meeting times. These are high-volume tasks with checkable outputs.
Works with review
Personalised openers, replies to simple questions, and summaries before a call. Output is often good and occasionally wrong in ways only a person who knows the product and market will spot.
Still needs people
Choosing the market and the offer, deciding when not to send, discovery calls, demos, pricing and negotiation, handling a sceptical or upset buyer, multi-stakeholder deals, and owning the relationship after the sale. These depend on judgement, context that is not in the CRM, and trust.
Why replacing the team tends to fail
The appeal of full automation is volume, and volume is also the risk. When a person writes a poor email, one prospect sees it. When an agent follows a poor instruction, every enrolled lead does. Sending reputation is shared across a domain, so a spike in complaints or bounces from one unsupervised campaign can push everyone's mail toward spam. Buyers also recognise templated personalisation quickly, and a reply handled badly by an agent can lose a deal a person would have saved.
There is a legal point too. Harvey's README tells users plainly that they are the sender. Consent and opt-out rules apply to the business that sends the message, whether a person or software wrote it, and rules differ by country and by type of recipient. An agent cannot carry that accountability for you.
What a realistic small team looks like
For a founder-led or small sales team, a workable shape is one or two people who own targeting, conversations, and closing, plus agents that handle research, drafting, and follow-up timing under supervision. Some teams add independent reps paid per held meeting for extra conversation capacity. The daily rhythm is short: review and release agent drafts, handle replies that need judgement, run calls, and check the metrics. The work shifts from typing outreach to supervising it, which is a real change, but it is not the disappearance of the team.
How to test the claim yourself
Run a small pilot rather than a migration. Choose one segment and one offer. Have the agent research and draft for that segment while a person approves every send for a few weeks. Compare against a person working a similar segment on the measures that matter: positive replies, meetings held, opt-outs, complaints, bounces, and the time spent reviewing. If the agent's drafts pass review consistently and its held meetings keep pace, widen its scope and raise its daily cap one step at a time. If not, narrow the task.
Questions to ask any tool that promises replacement
Ask what the agent can do without approval, how approval works in practice, and whether there is a kill switch. Ask how suppression, opt-outs, and regional consent are enforced, and whether they are enforced by code on every send or only by the agent's prompt. Ask how sending volume is capped per mailbox, what happens when bounce or complaint rates rise, and whether every action is logged against the agent that took it. Vague answers to these questions matter more than a strong demo.
Where Koryo fits
Koryo is built on the assumption that people and agents share the work rather than one replacing the other. Agents connect over MCP as operators with their own identity, permissions, and daily caps, and they do not take a paid seat. Autonomy has three levels: suggest only, send with approval (the default, with new agents capped at 100 messages a day), and autonomous within limits. Agents cannot approve drafts, convert, close, or change settings unless granted, and every message passes the same compliance gate as a person's.
One practical caution before running it: third-party scanners flag that Harvey's setup instructions use Claude Code's permission-skipping flag, which lets the agent act without per-action prompts. If you try it, run it in a container or a throwaway account rather than against your main mailbox and machine.
Frequently asked questions
- Can AI replace SDRs?
- AI can take over a large share of SDR tasks: research, list building, first drafts, follow-up timing, reply sorting, and scheduling. It does not reliably replace the judgement an SDR applies about who to contact, what to say to a sceptical buyer, and when to stop. Most teams that succeed use agents to raise each person's capacity, with people reviewing what goes out.
- What is Harvey in AI sales?
- Harvey is an open-source project on GitHub, published under the MIT licence, that describes itself as an autonomous sales agent running on a Claude Code or Codex subscription. Its README says it finds businesses, reads their websites, writes cold email, and handles replies, with every outgoing email held for approval by default. It is unrelated to Harvey, the legal AI company.
- Is it safe to let an AI agent send cold email on autopilot?
- Only after it has earned it. Start with every message held for approval, review drafts for factual errors and tone, and keep a low daily cap. Suppression lists, consent rules by region, opt-out handling, and bounce and complaint limits should be enforced by the system on every send, not by the agent's instructions, so a bad prompt cannot bypass them.
- What should a small team hand to an AI agent first?
- Pick one narrow, repetitive task with an easy quality check, such as researching accounts and drafting first-touch emails for one segment. Keep a person approving every send for the first few weeks. Expand the agent's scope only when its drafts pass review consistently and the held meetings it contributes match what a person achieves on the same segment.