Human Approval in AI Outreach: Where a Person Must Sign Off

Picture an AI SDR agent with a list of 2,000 contacts and one wrong assumption baked into its prompt. Maybe it believes every company in the segment uses Salesforce. Maybe it thinks the persona owns a budget they do not own.

A rep with that wrong idea sends it to 30 people and notices on day 2 when the replies come back confused. The agent sends it to all 2,000 before anyone reads a reply.

That is the case for human approval. This article covers the 5 points where a person must sign off, what a reviewer checks at each one, how to keep review fast, the numbers that tell you autonomy can widen after month 3, and what goes wrong when approval is skipped.

What Human Approval Means

Human approval is a sign-off step where a person reviews an agent's output before it reaches a buyer or changes who the system targets.

It sits on top of an AI sales team, which we define in What Is an AI Sales Team?. Agents do the watching, research, list building, drafting and reply sorting. People decide who to target, approve what goes out and answer the buyers who are interested.

The principle behind where to put approval is multiplication. A decision that applies to one message needs light review. A decision that applies to every message in a segment needs a person, because one error becomes hundreds of identical errors.

The 5 Approval Points

Every outbound system we build has the same 5 approval points. Between them, agents run without waiting for anyone.

Fig. 1The 5 approval points in AI outreach

1. 01Targeting

   Who we reach and why: segments, personas, signals, exclusions.

2. 02Lead list

   Every account and contact checked before a message is drafted.

3. 03First touch

   Each first message read against its research brief.

4. 04Replies

   Interested buyers and objections get a human-checked answer.

5. 05Anything new

   A new angle, segment, channel or offer returns to full review.

1. Targeting

Before any agent runs, a person approves who the system will reach: segments, personas, which signals count, and who is excluded. This is the decision with the widest blast radius. A wrong segment wastes every message sent to it.

We build targeting as an ICP matrix: 1 ICP split into 3 segments by 2 personas, giving 6 lists and 3 angles each, 18 tests in total. The client signs off on the matrix in week 1.

2. The lead list

The agent builds the list, then a person checks it before any message is drafted. The checks are about fit and conflict: does each company match the segment, is the contact the right persona, and is anyone on the list already a customer, an open deal or someone your team spoke to recently.

3. The first touch

Every first message is read against the research brief it came from. The reviewer is checking that the facts are true, the trigger is recent, the tone fits and the ask is clear. This is where most edits happen in the first month.

4. Replies

Agents sort replies into interested, not now, wrong person, objection and unsubscribe. Clear cases like unsubscribes and out-of-office messages can be handled automatically. Interested buyers and objections get a drafted answer that a person checks before it sends, because that answer decides whether a meeting gets booked.

5. Anything new

A new angle, a new segment, a new channel or a change to the offer goes back to full review, even if the system has earned more autonomy elsewhere. New work has no track record, so it has not earned trust yet.

What a Reviewer Checks

Review gets fast when the reviewer has a short checklist and the agent shows its sources. Without a checklist, every draft becomes a fresh debate about style. With one, the reviewer scans for the few things that matter and approves the rest.

The list check matters more than it looks. Sending a cold email to a current customer, or to a prospect in the middle of a deal with your sales team, damages a real relationship. The CRM hygiene play covers keeping that data clean enough to check against, and Fix Your CRM Before You Scale Outbound covers the order of work.

Keeping Review Fast

Approval fails when it becomes a bottleneck. A founder who has to rewrite 40 drafts every morning will soon stop reviewing. These habits keep it to minutes.

• Sources next to claims. Every fact in a draft links to where it came from: the job post, the press release, the public post. The reviewer checks the link instead of searching.
• Batch review. Drafts arrive once a day as a batch, in one place, with approve, edit and reject buttons. No chasing messages across tools.
• Fix the system behind the draft. If the reviewer makes the same edit 3 times, change the prompt or the research step so the agent stops making that mistake.
• Short drafts. A first touch under 110 words takes seconds to read. Long drafts are slow to review and slow to reply to.
• Clear reject reasons. A tagged reason ("wrong fact", "wrong persona", "tone") feeds back into the system. A silent rejection teaches it nothing.

One system our founding team built shows how fast review can get. A thought-leadership engine for a global consulting firm's TMT practice monitors 500+ sources and filters out 95% of signals automatically. The 50+ senior partners who publish with it spend under 10 minutes of review per partner. The agents do the reading; the person makes the final call quickly.

Who Should Review

The reviewer matters as much as the checklist. Three traits make a good one.

• They know the buyer. They have been on calls with this persona and can tell when a draft assumes something a real buyer would never say.
• They know the offer. They can spot a promise the service does not deliver, or a price hint that is out of date.
• They did not write the prompts. The person who built the system reads drafts as they meant them. A second person reads them as the buyer will.

In a B2B company above about $100k/month, that is often the founder for the first weeks, then a sales leader or senior rep once the checklist is settled. The founder's time is the scarce resource, so the goal of month 1 is to turn the founder's judgment into checklist items and prompt changes that someone else can apply.

Set a fixed time for review, such as 20 minutes each morning, and a rule for what happens if review is missed. The safe default is that nothing sends until it is approved. A day of delay costs little. A day of unreviewed sends can cost a segment.

When Autonomy Can Widen

Approval is heavy in month 1 by design. It should get lighter as the system proves itself, one approval point at a time.

Fig. 2How approval changes over the first 6 months

1. Month 1
   Full review

   Every list, every first touch, every reply to an interested buyer.

2. Month 2
   Faster review

   Checklist settled, sources attached, batch approvals in one sitting.

3. Month 3
   First widening

   Proven angles move to sampled review if edit rates stay low.

4. Months 4 to 6
   Steady state

   Spot checks on proven work, full review on anything new.

The question is what counts as proof. Track these numbers for each angle and segment, and widen autonomy only when they are low and stable for several weeks.

• Edit rate. The share of first touches the reviewer changes before approving. For example, a team might decide that fewer than 1 in 20 drafts edited across 4 straight weeks qualifies an angle for sampled review.
• Reject rate. Drafts thrown out entirely. A rising reject rate means the research or the targeting has drifted.
• Fact errors. Any draft with a wrong fact resets the clock for that angle, however low the edit rate was.
• Negative replies. Replies that point out an error or complain about the message. These are the clearest sign that review was needed.
• Spam complaints. Gmail and Yahoo expect bulk senders to stay below a 0.3% complaint rate, and Google advises under 0.1%. Any upward drift means review should tighten.

When the numbers qualify, move from reviewing every draft to reviewing a sample, for example 1 in 5. Keep reviewing everything new. Use the matrix below to decide which tasks can widen and which never should.

Fig. 3Where autonomy can widen and where it should not

How often the task repeats, low to high

Automate with spot checks

Frequent and low risk: reply sorting, CRM updates, out-of-office handling.

Approve every time, keep it fast

Frequent and high risk: first touches, answers to interested buyers.

Automate fully

Rare and low risk: formatting, scheduling, data cleanup.

Person decides

Rare and high risk: targeting, new segments, new offers.

Cost of one mistake, low to high

Two approval points should never move to full autonomy: targeting and anything new. In Oriel's own builds, lead lists and first messages stay on full approval as well. Both are rare decisions with a high cost of error, and both are where a person's judgment about the market matters most. For a broader view of where agents add value and where they cause harm, see AI SDR Agents: Where They Help and Where They Hurt.

Failure Modes When Approval Is Skipped

Skipping approval looks efficient in week 1. These are the ways it tends to fail, and each one is harder to repair than to prevent.

Wrong facts at volume

The agent fills a gap in its research with a plausible guess. A buyer gets congratulated on a funding round that belongs to a different company with a similar name. Multiply by the list.

Burned segments

A weak first message goes to an entire segment. Those buyers now know your name and associate it with a bad email. The second, better message gets ignored.

Deliverability damage

Messages that feel automated get marked as spam. Complaint rates climb, inbox placement drops, and good messages stop arriving. Warmed inboxes take about 2 to 4 weeks to build, and much longer to repair.

Conflicts with live deals

A cold email reaches a buyer your sales team is already negotiating with. Now your best rep spends a call apologizing for your own outreach.

Mishandled replies

An agent answers an interested buyer's pricing question with a calendar link, or replies to an objection with the same pitch. The meeting that was almost booked does not happen.

No one learns

Without review, nobody reads the drafts closely enough to see patterns. The prompts never improve, and the system stays as good as it was on day 1.

How We Run Approval at Oriel

Our own operation keeps a person on every reply. Ashwanth's daily LinkedIn Sales Navigator routine sends connection requests to qualified owners each morning and a researched first message to everyone who accepts, with the person's profile and recent posts read before any message so it opens on something real. Replies are flagged for a human to answer.

When we run outreach for a company, that company approves every lead list and every first message before anything sends. Everything runs in accounts the company owns, so the approval history, the lists and the drafts stay with them. The capital projects advisory system our founding team built followed the same ownership principle: synced to the CRM, then handed over. The client now runs it unassisted.

FAQs

Run Approved Outreach With Oriel

Oriel is the engine room behind the go-to-market team at B2B companies above about $100k/month in revenue. Agents do the research, list building, drafting and reply sorting, and you approve the targeting, every lead list and every first message, with sources shown next to every claim so review takes minutes.

On a 30-minute call, we will walk through where your approval points would sit, who on your side should review, and what the first month of drafts would look like. If you are not ready for this yet, we will tell you that too.