Where AI belongs in a revenue system
Give AI the narrow jobs: drafting a reply from rules the business wrote down, sorting messages and summarising them. Give plain rules the screening, routing, reminders and sending each message once. Keep price, scope, delivery dates and exceptions with a person, and have someone at the business approve every drafted reply. The AI never invents a price.
When I build a system that turns leads into revenue, the AI does a few narrow jobs in the middle of it. Plain rules do the steps that should happen the same way every time, and a person keeps every decision that commits the business. I've built two systems this way: the inquiry desk, which sorts the email that reaches SignalNine, and lead to reply, which answers website form leads. Both run the screening, sorting, drafting and approval steps today. The routing, summary and reminder rows below are where I'd put those steps when I add them.
Each step, and who does it
| Step | Who does it | Why |
|---|---|---|
| Screening out junk, newsletters and automated mail | Rules | A fixed check catches these the same way every time. It runs before any AI step, so in the 2026-09-07 run on SignalNine's own inbox, junk cost $0.00 to handle. |
| Telling a new lead from a sales pitch | AI, with a rule acting on its answer | People describe their needs in too many ways for a fixed check. The AI picks from a short list of answers, and a rule decides what happens next. |
| Routing each message to the right person | Rules | Who handles an existing client or a supplier is a decision the business makes once. Written as a rule, it holds when someone is away. |
| Summarising a long email for whoever acts on it | AI | It saves reading time, and the person reads the summary before acting, so a miss gets caught there. |
| Drafting the reply | AI | It drafts from the facts and rules the business wrote down, in the way the business writes. It never puts in a price nobody gave it. |
| Approving each drafted reply | A person at the business | The owner, or someone they name, reads it on their phone and approves it, edits it or turns it down. |
| Price, scope, start and delivery dates, and exceptions | A person | Each one commits the business in writing, and a wrong one is hard to walk back. |
| Reminders and follow-ups on dates already agreed | Rules | Once a person has set the date, sending the reminder on time is the same step every time. |
| Sending once and logging the run | Rules | A reply has to go out exactly once, even if approve is pressed twice. That is a job for a fixed lock, which works the same way on every run. |
- Rules screen and route each message
- AI drafts from your written rules
- A person approves anything that names a price, scope or dateHUMAN
- The system sends once and logs it
The jobs I give to AI
AI gets three kinds of job: drafting from rules the business wrote down, sorting a message into one of a few kinds, and summarising, which neither system does yet. Each one has a short output that a rule or a person checks right after, so a mistake stops at the next step instead of reaching a lead.
A narrow job can still go wrong, so I test each one alone before it runs on real mail. On the inquiry desk I tested the sorting step on sample emails on 2026-09-12, with 60 runs per setting. An inquiry from someone describing how their own team works was thrown out as a sales pitch in 8 of 60 runs with the old wording on a smaller AI model. After the wording change on a larger AI model, it was thrown out in 0 of 60. The new wording helped on its own, and the larger model took it the rest of the way.
That kind of test is only possible because the job is narrow. The same sample email goes in many times, and the misses come out as a count I can compare after each change.
What plain rules do better
Rules take the steps where the right answer is already known: screening, routing, reminders on agreed dates, and sending each message once. A rule does the same thing on every run, costs nothing to repeat, and is easy to read back when someone asks why a message went where it did.
The inquiry desk is a system I built, running on SignalNine's own inbox since 2026-09-07. Its screen runs before any AI model is called, so in the 2026-09-07 run on SignalNine's own inbox, junk and out-of-scope mail cost $0.00 to handle. The trade-off today is that an email the desk sets aside as junk or a pitch raises no alert, so a real lead set aside by mistake shows up only in the run record.
On one test lead, the AI step was quick and cheap. On lead to reply, a system I built for website form leads, one test lead took 5.35 s from the form to a score and a drafted reply. On 2026-09-16, the single AI call that scored one test lead and drafted its reply cost $0.0001179, about a hundredth of a cent. So the person stays on the reply because of what a reply can commit the business to.
What stays with a person
Price, scope, delivery dates and exceptions stay with a person. Each of them is a promise the business has to keep, and a lead reads a number in an email as an offer. So the hard rule on every system I build is that the AI never invents a price. On the inquiry desk, a rules check runs on every draft and flags a made-up price, a delivery promise or a risky ask before anyone reads it, and a price goes into a reply only when a person gives it one.
Dates sit on both sides. A person commits to a start or delivery date, and once it's agreed, rules send the reminders on time. Exceptions go to a person too, with the context to decide: an odd request, an upset client, or a lead asking for something the business doesn't do.
Someone at the business approves every drafted lead reply from their phone, and the business picks who: the owner or someone on the team. On SignalNine's own inbox that person is me. Everything else acts on its own, like logging the lead, creating a follow-up task or sending a standard welcome email the business signed off on at setup.
Placing AI in a system you already run
- Write the rules down first: what the business sells, what it doesn't, how it prices, and examples of how it writes. The AI drafts from that, and you can change it any time.
- List each step from the first message to a signed deal, and mark each one rules, AI or a person, using the table above.
- Give AI only the steps with a short output that something checks right after, and put a rule or a person straight after each one.
- Test each AI step alone on the same sample emails many times before it touches a real lead, and test again after every change.
- Name who approves replies, and a backup for when they are away.
If you'd like to mark up your own steps this way, we can go through them on a call.
Questions
Can AI write the whole reply to a lead?
It can draft the whole reply, and on the systems I build it does. Someone at the business approves every drafted reply before it goes out, and a price goes in only when a person gives it one.
Should AI answer pricing questions?
Only with a price a person has already set, either in the written rules or while approving the draft. When the price depends on scope, the draft says someone will confirm it, and the person approving the reply decides what to say.
Where should a small business start with AI in sales?
With drafting replies to new leads. The draft saves the most typing, and a person still reads it before it goes. Sorting and summarising come next, while routing, reminders and sending stay on rules from the start.
What if the AI sorts a real lead as a sales pitch?
It can happen, which is why I test the sorting step alone on sample emails and count the misses after each change. On the inquiry desk today, an email set aside as a pitch raises no alert, so a real lead set aside by mistake shows up only in the run record.
Does AI make a revenue system expensive to run?
In the runs I've measured, the AI step is the cheap part. On lead to reply, the single AI call that scored one test lead and drafted its reply on 2026-09-16 cost about a hundredth of a cent. On the inquiry desk, the screen runs before any AI step, so in the 2026-09-07 run on SignalNine's own inbox, junk cost $0.00.
Who should approve the replies?
Someone at the business who knows its prices and its clients, usually the owner or a person they name. They approve from their phone, and a named backup covers when they are away.
Related
- A drafted reply to each lead, approved from my phone before it sendsAI drafts replies to lead emails that pass a junk screen, from my business facts. Risky asks get flagged, and nothing sends until I approve it on my phone.
- Website form leads, scored and answered with a reply I approveA system I built: each website form lead is checked, added to HubSpot once and scored. Gmail sends the drafted reply when I approve it in Telegram.
- Human-in-the-loop approvalWhat human-in-the-loop approval means in an automation, which steps should wait for a person, and how an approval card makes sure an email sends once.
- The lead-response bottleneck: how fast to answer a new leadHow fast to answer a new lead, why a slow reply hands the decision to someone else, and a system that drafts from your rules for a person to approve.
- What a reliable automation looks likeFive properties any automation should have, checkable by an owner who doesn't build: logged runs, named alerts, one send per approval, safe reruns, an owner.
- Build it, hire for it, or install a system: a one-page decisionDecide whether to build it yourself, hire a person or install a system by the kind of work it is, with a one-page memo to fill in for your own business.