Natthan.AI Lab · A playbook for owners
What AI actually does for a business like yours
What you'll be able to do
You can hand an email audience to an owner agent, the agent responsible for one business, and check it against consent records before approving a send.
Your owner agent can prepare an email audience or draft a lending page from your records. Use Natthan.AI Lab's company brain, the operating layer that runs a business on AI agents, as the pattern for checking that work. Start with the ecommerce account below: a separate check found 250 reachable people after the first pass reported four.
Give your owner agent a job with a check
Pick the email audience as your job and write your consent rule, the condition a contact must meet before receiving marketing email. Keep “what stays with the owner”, the decisions the agent cannot take, beside that rule.
You know it worked when your instruction requires current consent and leaves the send waiting for you, as the ecommerce consent and owner checks require.
Give the agent the records behind the job, using the record groups on the company brain map:
- Pull customers and orders, including consent from both platforms.
- Attach money and metrics from the sales or subscription records.
- Load rules and knowledge, including your consent rule and do-not-contact list.
You know it worked when the count the check reports is the count you can reproduce from the records, the way 250 was in A green checkmark said 4 people were reachable. The real number was 250.
Ask your owner agent for the first audience list with the source of consent beside each contact. Put the handoff on the bus, the shared message log agents use to pass work to each other.
You know it worked when a separate reviewer can see whether both consent sources were read, following the Getting customers seat's check on the assumption that left only four people reachable.
Put a gate, a checkpoint with the power to stop work, between the audience draft and the send. Require a receipt, evidence from the actual result that shows what the check found, with the merged count and any unresolved opt-outs.
You know it worked when each address clears the merged consent and do-not-contact check described on the ecommerce page, and a failed check holds the send.
Decide what waits for you: require approval for money or external sends, and keep deletions behind your approval too. Ask for a pinpoint, a signal worth your attention, when a consent conflict needs your decision.
You know it worked when a money-page change still waits for your sign-off after its checks pass, as the lending owner's rule requires.
Read the receipt before repeating the job on a heartbeat, a scheduled run that checks for work and stops when nothing is due. Keep the work-due gate, the check that stops an idle run, ahead of any paid model call.
You know it worked when the next run either picks up due work or exits before a paid model call, as the Running the business seat specifies, with the last result still there to read.
Try the lending check on a draft
Give a lending draft the same stop point as the lab's marketplace: run the compliance and citation check before the page ships. Use a stale legal citation or a “guaranteed approval” phrase to test that it blocks the draft. Keep the money-page change waiting for your sign-off even after the corrected copy passes.
Request an out-of-family reviewer, an AI from a different model family than the builder, for a high-stakes change. Follow the lending practice of calling that review when needed; do not assume it runs automatically on every change.
Worked example
Follow the ecommerce brand's email audience in A green checkmark said 4 people were reachable. The real number was 250.
Owner's instruction. I set the consent rule: build the campaign audience only from people whose marketing consent is current.
What the agent produced. The first pass reported four reachable people and marked the job done. It cited the consent rule and spot-checked samples, but the tool read consent from only the current ecommerce platform.
What the check caught. A separate agent questioned whether that platform held all the consent. The legacy email platform held consent too. Reading both sources produced 250 reachable people. The tool had buried an extra check against only the ecommerce platform, shrinking the audience back to four each time.
What the owner decided. I decided merged consent was the correct rule. The agents repaired the audience tool and resolved the real opt-out conflicts between systems. The send waited for my approval, as the company brain map's audience handoff shows.
Copy that division of work: you decide which consent rule to accept; the agent repairs the tool and brings back a count you can reproduce.
Mistakes to avoid
- Check both consent sources before trusting a green count: A green checkmark said 4 people were reachable. The real number was 250.
- Trace a real report through the checker before relying on it: We built the lie-detector. It just wasn't plugged in.
- Recheck the checker against the live app; it reported false failures until its checks tested actual outcomes: A green checkmark that isn't the AI's own word: building a quality check with teeth.
- Demand evidence of the reported work; an agent had called a blank page complete: The rule wasn't 'don't lie to me.' It was 'don't lie to any agent, including yourself.'
- Open the page a customer sees before accepting the test count: 463 passing tests. A blank website. Both true at once.
Ten minutes, today
Write one rule using the ecommerce send boundary: “The agent may draft X; it may not send/spend/delete without me.” Replace X with your actual email job.
Hand the agent one real customer consent record and one real job: draft that customer's audience entry with the source of consent attached.
Check the entry's consent source against the customer record yourself. Ask whether consent also lives somewhere else, the assumption that turned 250 reachable people into four. Hold the send until both consent sources and any opt-out conflicts are checked.
Save the checked entry as your receipt beside the owner rule. Keep the lab's line: nothing goes out without the check you'd have made yourself.
Frequently asked questions
What can AI actually do for a small business?
Give it the audience job from the ecommerce account: build the list from consent records and bring it back for review. Or give it a lending page to draft under the compliance and citation check. You keep the consent rule and approval of anything touching money. The lab calls this operating layer a company brain: shared records support agents that carry the work, with checks before shipping and one place for the owner to decide.
Do I need to be technical to use AI for my business?
Start with the owner's part of the ecommerce case: choose the consent rule and check the audience against the records. I spent most of my time deciding that merged consent was the right rule; the agents repaired the tool. You can try the consent check in Ten minutes, today before setting up the company brain's technical gates.
Where should I start?
Use the ten-minute exercise on this page: one real consent record, one audience entry to check. Write that the send stays with you before handing the record over. Follow the Getting customers seat's rule: give the review to something other than the worker and question whether the source contains all the information.
Isn't it risky? What if the AI gets something wrong?
The ecommerce agent reported four reachable people where the merged records supported 250. Keep that failure in your check: question whether all consent sources were read. Test the checker on the real path too; the lab once built a checker that no live report passed through. Some lab checks are wired today and others remain written rules, so ask for evidence that yours actually stops the send.
Read what a company brain is to place your audience job on the same map as the lab's owner decisions.
Use the Getting customers seat to carry this consent check into your next audience job.