How AI-assisted messaging is written
AI prepares the research and first drafts, human writers write around the real trigger, and a person approves every sequence before it sends.
The short answer
Deligatr writes outreach grounded in the specific signal that triggered contact, not a generic template. AI does the research and first drafts, human writers refine them within defined constraints (no guarantees, no performance claims, no false urgency), a person approves every sequence, and weekly reviews decide what to keep.
This picks up where the research engine stops. The channels these messages travel through are covered on the Outbound overview, and the wider process is on how managed outbound works.
In practice
What a signal-led message looks like
The signal on the left decides what the message says on the right.
The signal
A regional developer files a new land acquisition permit in a market that a construction firm serves.
The message
“Hi [First name], I saw the filing on the [Neighbourhood] parcel. We have worked on a nearby site with a similar density constraint, which may be a useful reference point. Worth a short call before you lock in your design team?”
Illustrative wording with placeholders. A reference like this is used only when it is true and the signal is confirmed.
From signal to sent message, in four steps
The order never changes: context first, a constrained draft second, human approval third, and performance feeding what comes next.
The workflow
Four steps, always in this order
Signals enter at step one and replies feed back at step four.
- 1
Step 1
Signal becomes context
The trigger found in research is passed in as grounding context, not a generic prompt.
- 2
Step 2
AI drafts, writers refine
First drafts follow set constraints and the tone of your sector. Human writers refine them.
- 3
Step 3
A person approves
Every sequence is approved before it sends. You review the direction and can change it any time.
- 4
Step 4
Performance shapes what follows
Replies feed back. Angles that work are scaled, the rest are revised or retired.
What every message must and must not do
A short standard keeps messages relevant, honest and easy to answer.
Always
Never
Reference a specific, real, current trigger
Open with a generic capability statement
Lead with one clear, low-friction question
Push a hard meeting ask up front
Keep it short and easy to answer
Write a long feature-by-feature pitch
Match the tone the industry expects
Reuse identical phrasing across every industry
Say plainly what we do if asked
Use vague, buzzword-heavy language
Stay honest about results
Make guarantees, performance claims or false urgency
What the data says about AI personalisation
More AI does not automatically mean better outreach. Specific, real references do the work.
LinkedIn connection requests
Specific beats generic, and AI hyperpersonalisation did not help
Expandi’s analysis separated hyperpersonalised AI openers from specific, real references.
13.2M
Connection requests analysed
28.5%
Average acceptance rate
12%
Lower acceptance with AI hyperpersonalisation, reply rate flat
1.8 to 2.7x
Acceptance for notes that referenced a real shared tie, against the 30.2% baseline
Expandi, analysis of 13.2 million LinkedIn connection requests. The 28.5% is the average across all requests. The 1.8 to 2.7x figure comes from 240 campaigns with at least 50 contacts each, measured against that comparison’s own 30.2% baseline acceptance rate.
That is why the engine starts from a confirmed signal and why writers and a person stay in the loop. For the mechanics of triggers, see the guide to signal-based outreach. Our cold email agency and LinkedIn lead generation pages cover the channels.
Guardrails and approval
Constraints are set before drafting, and a person checks the result before it reaches a prospect.
Every engagement excludes certain language by default. Messages are grounded in verified enrichment data and confirmed signals, so a detail that cannot be confirmed is left out. For regulated industries your compliance team can review the direction before launch.
Tone follows the sector, set in the ICP workshop. See industries for how different sectors are approached, and data privacy for opt-outs and compliance.
Approval gate
Nothing sends until it passes
The checks every sequence goes through.
Sequence for Northbridge Advisory
Three touches, drafted around one trigger
- Grounded in a confirmed signal
- No guarantees or performance claims
- Tone matched to the sector
- Approved by a person before it sends
Questions about messaging
Straight answers on who writes, who approves and what the AI is allowed to say.
Does a human ever see the messages before they send?
Yes. AI researches each account and prepares first drafts, human writers write and refine the sequence, and a person approves every sequence before it sends. You review the messaging direction during onboarding and can ask for changes at any point.
How is this different from a mail-merge template with a first name inserted?
A mail-merge changes one variable in an otherwise identical message. Here the message references the actual signal that triggered contact, so the substance changes, not just the greeting. The evidence points the same way: Expandi’s analysis of 13.2 million LinkedIn connection requests found AI hyperpersonalisation lowered acceptance by 12% and left reply rate flat, while connection notes that referenced a real shared tie (a follow, a group or a shared network) were accepted at 1.8 to 2.7 times the 30.2% baseline acceptance rate.
Can we set rules about language the AI should never use?
Yes. Every engagement excludes certain language by default: no guarantees, no performance claims and no false urgency. Further restrictions can be added for your compliance needs or brand voice.
Does the AI ever make things up about a prospect?
Messages are grounded in verified enrichment data and confirmed signals, not invented details. If a signal cannot be confirmed, it is not referenced. A person approving every sequence is the second check.
How does messaging differ across industries?
Tone and framing follow the ICP workshop and the sector. For example, accounting and financial services usually call for a consultative tone, while architecture and construction can lead with project context. There is no single universal template.
Can we review and change the messaging approach after launch?
Yes, at any point. Performance is reviewed weekly, which is a natural moment to request changes, but you can raise a messaging concern whenever it comes up.
Keep reading: Outbound overview, deliverability, research engine and documentation home. Packages and prices are on pricing page.