← The reading table Using AI in sales without sounding robotic
AI in sales goes wrong when it is handed the conversation. Give it the work that happens before and after one, and the outreach stops sounding like it was addressed to a record in a database.
Give AI the work that happens before and after a sales conversation, and keep the conversation itself human. Research, enrichment, scheduling, logging, pipeline hygiene and first-touch sequencing are volume work with predictable logic. Reading a reply and deciding whether to push or wait is judgment, and it is the part that is being bought.
"Hi [First Name], I came across your profile and thought our solution could really help [Company Name] achieve its goals in the current market landscape."
The tokens are showing. Everyone has received this email, and most people have received it twice this week from the same company.
Why does AI sales outreach sound robotic?
Because personalisation gets treated as a data problem when it is a relevance problem. Feed a system a job title, a company size and a recent funding round and it will produce something factually about the reader that has no reason to exist. Accuracy is not the same as relevance, and only one of them makes somebody reply.
Relevance requires a judgment about what matters to this person, this week, given something you actually know. That judgment is what the sender is supposed to supply. When it is delegated, automation scales the volume and scales the absence of judgment along with it, which is why the problem gets worse as the tooling gets better.
Which sales tasks should AI own?
The ones that repeat, have a checkable output, and take a person's attention without needing their opinion.
Logging activity. Updating records. Pulling research together before a call. Scheduling. Flagging deals that have gone quiet. Sequencing first-touch outreach that opens a door. Ranking inbound leads by behaviour rather than by arrival time, which is where most of the available improvement actually sits.
Pipeline hygiene deserves its own mention, because it is the least discussed and the most reliable. A pipeline is an optimistic document. Deals sit in stages they left weeks ago, contacts who have not replied since July are still marked warm, and notes get updated inconsistently, so the manager's view is assembled from wishes. A system tracking time in stage and actual engagement flags the stuck ones without anybody having to admit anything. That is clarity, and it arrives without a conversation about rep discipline.
Which sales tasks should stay human?
Everything inside the conversation, and the decision about what to do next after one.
Reading tone in a reply and knowing whether to push or pull back. Recognising that a deal is stalled on timing rather than interest. Hearing that a difficult-sounding client is a cautious one, and adjusting. Responding to an objection nobody scripted.
The boundary is worth writing down as a list, because it drifts. Every tool update adds a capability that sounds like it belongs on the other side of the line, and the drift is gradual enough that nobody decides on it. OWASP files the general version of this as Excessive Agency, sixth in its 2025 Top 10 for LLM Applications: a system granted more autonomy or more permissions than its task requires. In a sales context the exposure is not a data breach. It is the relationship, and it fails quietly, which makes it harder to notice and more expensive to repair.
How should you actually use AI to write outreach?
Write the thing you mean first, then hand it over to be tightened.
The template problem starts with asking a model to produce the whole message from nothing. Supply the specific reason you are reaching out, what you know about this person, and what you want them to do, even as three rough lines. Then use the model to cut the length, fix the structure and check the tone. You keep the point of view. It removes the throat-clearing.
The test is reading it aloud. If it sounds like something you would say to someone, it goes. If it sounds like a brochure, it does not, and how well the merge fields were populated makes no difference to that.
What does sending more of it cost you?
Deliverability, and the threshold is published.
Since 1 February 2024, anyone sending more than 5,000 messages a day to Gmail addresses has had to meet Google's sender guidelines: SPF, DKIM and DMARC authentication, one-click unsubscribe, and a spam rate in Postmaster Tools kept below 0.30%, with Google advising you stay under 0.10% and never reach 0.30% at all.
Automation makes volume the easiest lever in the building, and that rate is what it is levered against. A sequence that annoys people does not cost you those people. It costs you the inbox placement of every campaign that follows, including the good ones.
The same applies underneath, where the data lives. A CRM with duplicate contacts, three definitions of the same pipeline stage and leads from 2023 nobody closed does not improve when a model is pointed at it. It produces outputs that faithfully reproduce the mess, faster. An afternoon of standardising fields does more than any amount of prompt work, and knowing what that data is and where it goes is the same job viewed from the other end.
Before the call: research, scheduling, enrichment, sequencing. After the call: notes, follow-ups, pipeline updates. During the call: you.
MorningLabs AI builds the agents, automations and internal systems a business runs on, and audits AI systems that are already live.