How to automate my business with AI in San Diego is a queue: the work that already repeats, with fields you can name, a write path into the system you already open, and a human gate on money and promises. You do not automate judgment. You automate retype, chase, and reminders.
Nidrosoft is the San Diego studio of Cyriac Ndo Zeh, a design engineer who implements AI for small and mid-size businesses: audits, custom builds, phone and receptionist agents, training, and systems the team can own.
Written by Cyriac Ndo Zeh, Founder of Nidrosoft — 12+ years building software and AI products for San Diego small businesses. Also on LinkedIn.
Key takeaways
- Automation is a queue with
NextAttemptAt, not a robot that "runs the company." - Five steps: list repeating jobs, pick one, write fields and gates, choose leftover type, measure a number you already track.
- A typical 10-person San Diego shop can move a follow-up pile from 10 hours a week to about 2 if the queue writes the row and a person still sends.
- Tasks that should never go to a bot: price exceptions, permits, clinical notes, "we are on the way," firing people.
- Seats draft. Zapier or Make moves clean fields. A custom agent handles messy speech into a trusted row.
- Soft next step: automate one queue this month. Leave the rest of the shop analog.
Contents
- How to automate my business with AI without boiling the shop
- A five-step automation plan
- Before and after in a typical San Diego shop
- What belongs in a queue
- Phrases owners use when they mean automate
- Which tasks should never go to a bot
- San Diego queues that already exist
- Glue, seats, and a custom leftover
- Frequently asked questions
- Queue the repeat, keep the gate
How to automate my business with AI without boiling the shop
How to automate my business with AI is the question after the third Saturday of the same follow-up. Owners want the repeat work off their chest. They do not want a science project that touches payroll, permits, and the published number in the same week.
The county map is AI implementation for small business. A route-shaped leftover sits on AI agent for cleaning company. City texture sits on AI implementation Santee.
Automation in a small shop is a list that moves without a hero. LastAttemptAt. NextAttemptAt. OwnerName. HumanGate. The model fills what it can. A person clicks the gate. If you cannot draw that list on a whiteboard, you are not automating. You are hoping.
Nidrosoft's method is Embed, Diagnose, Scope, Build, Own. Live public leftovers include Omnira Dental, Octana, Eviction Wizard, Guidera, TokenTra (10-30% wasted spend cut when a team can see the bill), Protectron, and Nerlude. Nexuvo dealers acquire 3x more inventory with 80% less manual outreach. NanoBrief turns a long brief into something under 2 minutes versus 3+ hours. Those are queues with gates. Steal that shape.
Formerly Anthropic, Microsoft (Bing and Edge), Intuit, and Gap Inc. is background. 12+ years, 125 products shipped, 16 public. Your queue still needs a Tuesday inheritor.
A five-step automation plan
-
List repeating jobs that already happened 10 or more times last week. Missed-call retype, estimate chase, recall texts, invoice reminders, review asks, weekly route confirms. If a job happened twice, it is not a queue yet.
-
Circle one job. Write
WorkflowNamein one sentence. "Estimates older than two days get a drafted text the coordinator sends." If the sentence includes "and also the website," you circled two jobs. Put one back. -
Write fields and gates. Fields:
RecordId,LastAttemptAt,NextAttemptAt,Status,CallbackOrEmail,DoNotContact. Gates: money, legal or clinical content, customer-visible sends if you want a person on every send, dispatch promises. -
Choose leftover type. Clean fields already in Jobber: Zapier, Make, or the automation toggle you already pay for. Messy speech on a phone: pickup SaaS or a custom agent. Drafting only: ChatGPT, Claude, Copilot, or Gemini. Write
AlreadyTried. -
Run 30 days against a metric you already track and a queue count:
items_aged,items_closed,items_skipped,hours_on_retype. Turn the leftover off if the queue grows and nobody owns exceptions.
Five steps. A sixth job in the same sprint is how a 10-person Santee shop loses the first leftover.
Before and after in a typical San Diego shop
This is a composite of a typical 10-person cleaning company that runs homes from Santee into La Mesa, not a named client.
Before: weekly confirms lived in a group text. No-shows showed up as a hole on the route. New-lead follow-up lived in the owner’s head. ChatGPT drafted a hiring post twice. Timing: about 10 hours a week on "are we still coming Thursday" and "did you get my quote" threads. They wanted automation. They did not want a new brand of vacuum.
After, six weeks later: a confirm queue texts the sentence they already used, 24 hours before the stop, and writes confirm_yes or no_reply into Jobber. A second queue drafts day-2 follow-ups on open quotes. The owner still prices extras and still handles a broken gate. Timed leftover: about 2 hours on the same piles. They automated two queues. They did not automate the cleaner.
What did not change: who gets the key, who handles a damage claim, and Saturday add-ons. What changed is the Thursday morning text storm.
If your shop is a clinic, the queue is recall. If your shop is a desk, the queue is documents. Same fields. Different nouns.
What belongs in a queue
Belongs: work with a record id, a next date, and a status that a person already understands. Confirms. Chase. Reminders. After-hours requests waiting for morning. Review asks after a completed job.
Does not belong: one-off emergencies, brand-new services you have never priced, a fight between two owners, anything you would not let a new hire send alone without a gate.
| Queue | Trigger | AI leftover | Human gate |
|---|---|---|---|
| After-hours intake | Missed ring | Agent or pickup SaaS | Price, same-day |
| Estimate chase | Estimate age | Draft plus send or approve | Discount |
| Appointment confirm | 24 hours before | SMS template, model only if speech is messy | Cancel and refill |
| Recall | Due date | Draft text | Clinical fit |
| Invoice reminder | Days past due | Template first | Collections tone |
| Review ask | Job complete | Template | Skip complaints |
If the row is already clean, use glue. If the row starts as a voicemail, use an agent. If the row is a paragraph you must write, use a seat. Mixing those three in one Zap is how queues rot.
Field names that keep a queue honest:
| Field | Why it exists | Fail if missing |
|---|---|---|
RecordId |
Ties the text to Jobber, Open Dental, or the book | You will double-text |
LastAttemptAt |
Stops the daily hero | The queue feels "busy" with no proof |
NextAttemptAt |
The only reason a queue exists | Items sit forever |
Status |
open, waiting, skip, closed | Nobody knows what to work |
AttemptCount |
Caps the leftover | Customers get a novel |
DoNotContact |
STOP, complaint, counsel | You will earn a bad morning |
OwnerName |
Exceptions need a person | The queue becomes a drawer |
HumanGate |
Who clicks send or money | The model freelances |
Print the table. If NextAttemptAt is blank, you do not have automation. You have a list of guilt.
Cap AttemptCount in writing. Three touches is enough for most estimate piles. Recall can sit on a monthly band. After-hours requests should not retry at 1 a.m. Write the cap before you turn the leftover on.
Santee and East County routes skip weeks when a customer is on vacation. Add SkipUntil if you already track that on paper. A confirm text that fires during a skip week trains customers to ignore you.
Phrases owners use when they mean automate
What can I automate with AI is the shopping version. You can automate the repeat list. You cannot automate taste. Start with confirms and chase. Leave the brand film alone.
How do I stop doing the same follow-up is the second phrase. You stop by putting the follow-up in a queue with dates, not by buying a chatbot that asks you what to do every morning. The bot that waits for you is still you.
Which tasks should never go to a bot is the third, and it belongs in the plan. Never: price exceptions, permits, clinical or legal advice, "we are on the way," deleting customers, payroll, hiring rejects, fake reviews. Write that list on the wall. The model will not honor a list you never wrote.
If you have asked all three, you are ready for step 1. List last week’s repeats.
Which tasks should never go to a bot
Money the owner would argue about. Anything a license sits behind. Anything a court or a board would read. Anything that pages a truck. Anything that names a chemical, a drug, or a legal right.
Recording and outbound texts still go to counsel. Identify automated lines. California two-party consent is widely discussed if you record. TCPA can apply to texts. This page is not legal advice.
Keys stay in your vault. Nerlude exists because keys still land in Slack. An automation that lives in a personal Gmail filter is a queue you will lose when the owner is on a plane.
San Diego queues that already exist
Trades already have a board. Automate after-hours into that board, then confirms. Rooms already have a book. Automate overflow holds, then no-show texts. Clinics already have a due list. Automate the draft, not the diagnosis. Cleaners and landscapers already have a weekly route. Automate the confirm. Property desks already have work orders. Automate status texts the resident can understand.
Santee, El Cajon, and East County routes feel heat and skip-week chaos in the same season. Coastal routes feel tourism and gate codes. Write gate_code if you already collect it. Never invent one.
Workshops have been run in La Mesa when a team needs to agree what a queue is. The room will not text your Thursday customers.
Fit is roughly 30-250 people, plus smaller shops if the workflow is clear. A 10-person shop can own two queues. It cannot own nine.
A 30-day automation calendar that stays boring on purpose:
Week 1: list last week’s repeats. Circle one. Write fields and gates. Get admin access. Name OwnerName and InheritorName.
Week 2: turn on the smallest leftover. Confirm template or after-hours request only. Meet 20 minutes on Friday. Count items_closed and items_skipped.
Week 3: fix two failures. Do not add a second queue. If skips are the story, the leftover is wrong or the sentence is wrong.
Week 4: keep the leftover or kill it. Write the runbook in Drive. If it lived, wait two more weeks before circling job two.
I-8 and Grossmont traffic still lie about same-afternoon windows. A queue that texts "we can add you today" without a zone check will create a route you cannot run. Keep windows behind the gate or behind a zone sheet you already trust.
If the shop texts customers, write the STOP handling before week 2. A queue that cannot honor STOP is not automation you want to own. Counsel owns the rest. This page is not legal advice.
A queue review that takes 20 minutes on Friday: open items_aged, close the ones a person finished by hand, skip the ones that should never have entered, and rewrite one sentence that failed. That is Own. A queue you never review is a drawer with a nicer name.
Glue, seats, and a custom leftover
| Leftover | Typical money | Automates | Fails when |
|---|---|---|---|
| ChatGPT, Claude, Copilot, Gemini | Seat cost | Drafts | Nobody pastes the row back |
| Zapier, Make, native toggles | Subscription you may already have | Clean field movement | The field is a voicemail |
| Rosie, Goodcall, Smith.ai | About $50/mo or per-minute | Name and callback | You still retype 40 minutes |
| Custom Nidrosoft agent | Audit in the low thousands; first leftover mid four- to five-figure band | Messy speech into your row | No inheritor, no admin |
Cost bands are typical, not quotes. The readiness call is free. Method is Embed, Diagnose, Scope, Build, Own.
Book a free 15-minute AI readiness call if the queue is named and the leftover is a write path. If the leftover is a Jobber toggle, turn the toggle on and skip the call. Email cyriac@nidrosoft.com or use https://calendar.app.google/GCtfuTx3Ms6K7d579.
Frequently asked questions
Is Zapier enough to automate the business?
It is enough to move clean fields. It is not enough to understand a 90-second voicemail. Most shops need glue plus either a seat or an agent, not glue as a personality. If the leftover starts as speech, buy pickup or a write path first and add Zapier after the field has a name.
Can we automate the whole phone?
You can automate after-hours intake first. Daytime main lines wait until the counter wants it and the greeting is done. Whole-phone cutovers in week one are how Friday dies. Keep a kill sentence in Drive so the owner can remove the forward before the weekend if the leftover lies.
What if the queue fills with junk?
Write DoNotContact, a skip reason, and an owner for exceptions. A queue without skips is a junk drawer. Review skips weekly. If junk is half the queue, the trigger is wrong. Fix the trigger before you add a second leftover.
How many queues should we start with?
One. Two if the second is a confirm template with no model. Nine is a program. Finish 30 days on the first queue, write the runbook in Drive, then circle job two only if someone besides the founder will own it.
Will staff lose their jobs?
Staff lose the retype. They keep exceptions, customers, and the work the license sits on. Shops that fire the coordinator and keep a silent queue get a louder Monday. Name the inheritor before you turn the leftover on so Tuesday still has a person.
Queue the repeat, keep the gate
How to automate my business with AI in San Diego is still a list with dates, a write path, and a named stop. Circle one repeating job. Write the fields. Pick glue, a seat, pickup, or a custom leftover. Measure a number you already trust.
Book a free 15-minute AI readiness call when the queue needs a write path. When the queue is a confirm template you already say out loud, put it in Jobber this week and keep the call for the messy door.
