Most of the questions I get about AI are some version of “can it do X?” The more useful question is “should it?” Almost everything in a sales business can be automated badly. The skill is knowing which parts to hand over and which parts to protect.
The rule we use
If a step repeats the same way every time and a mistake is cheap to catch, automate it. If a step commits the company to something, judges a person, or needs a relationship, keep a human on it — and make the system prepare everything that human needs.
That rule sounds obvious. In practice it means drawing lines in places people don’t expect.
Five things we never let AI own
1. Price. AI can apply your price book and draft the proposal. It never invents a number, grants a discount or agrees to an exception. In our proposal systems, prices come only from the owner’s price book, and a person approves every proposal before it goes out.
2. The close. Getting someone to a conversation faster is where AI earns its keep. The conversation that ends in a signature belongs to a salesperson who can read the room and be accountable for what they promised.
3. Hiring decisions. Our recruiting engine screens applicants against job-related requirements and shows the applicant’s own words as evidence. It cannot mark anyone hired. A person interviews, and the owner approves. Requirements are reviewed so they don’t act as proxies for protected traits.
4. Upset customers. A warranty problem or an angry homeowner never gets an automated reply. In our re-engagement system, those threads are excluded and become a task for the owner instead.
5. Unapproved messages. Anything that goes out in your name is sent from wording someone approved. In the operating system we’re building for ourselves, an email can only be sent if its content matches the version a person approved — change a word and it needs approval again.
What AI is actually good at
The list of things we do automate is long, and it’s where the time comes back:
- replying to a new lead in seconds instead of hours;
- asking only the qualifying questions a form didn’t cover;
- turning “yeah next week works” into a structured answer;
- drafting follow-ups that stop the moment someone replies;
- reading supplier invoices into the right fields;
- summarizing a call against your playbook so a manager knows where to listen.
None of those decide anything. They make sure a human decides with better information, sooner.
Why deterministic rules still matter
A lot of what people call “AI” in sales tools should be plain rules. In our prospecting engine, scoring, call eligibility and queue order are written rules, not model outputs — because when a regulator, a customer or your own team asks “why was this person called?”, you need an answer you can read.
We use AI where formats vary and judgment is cheap to check (reading a PDF, drafting a message), and rules where correctness matters (who can be contacted, which job gets the cost).
A test you can run on any AI tool
Ask the vendor three questions:
- What does it do when it isn’t sure? (Good answer: it says “unknown” and asks a person.)
- Who approves messages before they’re sent? (Good answer: a named person, with a record.)
- Can I see why it made a decision? (Good answer: yes, with the evidence.)
If the answers are vague, the tool is automating judgment it shouldn’t.
