What is an AI agent is software that takes a job from a phone, inbox, or list, uses tools you named, writes a row in a system you already pay for, and stops at a human gate. A person still owns money. The leftover is yours on Tuesday.
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
- An agent has a job, tools, a write path, and a stop. A chatbot has a conversation.
- ChatGPT is a model you prompt. An agent is a system that can call Jobber, place a hold, and wait for a person.
- Typical shops feel 8-12 hours a week back when the agent removes Monday retype, not when it "sounds smart."
- Human gates belong on paper: money, filings, clinical notes, dispatch promises.
- A $50/mo receptionist can be a thin agent. A custom leftover starts when your no-book list and zones have to win.
- Soft next step: write the job in one sentence, then decide seats versus a write path.
Contents
- What is an AI agent in plain shop language
- A four-step way to tell if you need one
- Before and after in a typical San Diego shop
- Agent, chatbot, and a person at the desk
- Questions people ask when they mean this
- What the agent is allowed to touch
- Where San Diego shops already use the shape
- When a custom agent is the leftover
- Frequently asked questions
- Write the job, then name the stop
What is an AI agent in plain shop language
What is an AI agent in a shop that sells time: a program with a written job, a list of tools, and a rule for when a human has to click. The job might be "after-hours calls become a Jobber request." The tools might be the phone system, Jobber, and SMS. The stop might be price and same-day promises.
A model is the engine that predicts the next sentence or the next field. ChatGPT, Claude, and Gemini are public models you can buy as seats. An agent wraps a model in your job and your tools. If there are no tools and no write path, you have a chat.
The county map is AI implementation for small business. A legal-desk shape sits on AI agent for law firm. If you want a city page after the definition, AI consultant San Diego is the local leftover.
Nidrosoft's method is Embed, Diagnose, Scope, Build, Own. Embed watches the real job. Diagnose writes fields and gates. Scope is one paragraph. Build happens in your tenant. Own is a runbook. 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. Steal the gate.
Formerly Anthropic, Microsoft (Bing and Edge), Intuit, and Gap Inc. is background. 12+ years and 125 products shipped, 16 public, is studio history. The definition does not change because of a resume.
A four-step way to tell if you need one
-
Write the job in one sentence that a coordinator would recognize. "Missed calls become a text and a row." "Recall list gets a drafted text." "Inbox leads get a hold request." If the sentence is "answer anything a customer might ask," you do not have a job. You have a wish.
-
List the tools the job needs. Phone, SMS, Jobber, Open Dental, HoneyBook, Gmail, a Sheet. If the only tool is a browser chat, buy a seat. If the tool is a trusted row, you are in agent territory.
-
Write the stop. Money, filings, clinical notes, "we are on the way," contract language. If you cannot name a stop, do not turn the leftover on. An agent without a stop will invent a truck.
-
Time the hours the human already spends on that job. If it is under about 4 hours a week, seats usually win. If it is 8-12 hours or more and the work is fields, an agent can pay. Use your timing. This page will not invent a study.
Four steps. If step 2 is empty, you wanted a chatbot. If step 3 is empty, you wanted a risk. If step 4 is empty, time three Fridays.
Before and after in a typical San Diego shop
This is a composite of a typical 9-person law office near downtown, not a named client and not legal advice.
Before: the intake line rolls to a paralegal cell. After-hours callers leave stories. Monday morning is 50 minutes of notes into the matter list. Consult calendars slip because the story never became matter_type, opposing_party_if_known, and callback_number. ChatGPT already drafted a blog paragraph nobody published. Timing: about 9 hours a week on intake retype and "did you get my voicemail" threads.
After, two months later: an after-hours agent collects matter type from a written list, a short fact pattern, a callback, and a conflict-check name if the office already asks that on paper. It writes an intake row. It does not give legal advice. It does not quote a fee. A person books the consult. Monday is exceptions: emergencies on the written list, Spanish callbacks, a caller who refused to give a name. Timed leftover: about 2 hours a week on the same pile.
The office did not replace a paralegal. The paralegal stopped being a voicemail printer.
If your shop is a trade, swap matter type for job_type and ZIP. The shape of the agent stays.
Agent, chatbot, and a person at the desk
| Kind | What it does | What it writes | Stop |
|---|---|---|---|
| Chatbot on a site | Talks | Maybe an email | Often none |
| ChatGPT or Claude seat | Drafts when a person prompts | Whatever you paste | The person |
| Thin receptionist SaaS | Takes a name on the phone | A transcript or a simple form | Vendor template |
| Custom agent | Runs a written job with tools | A row in your system | Your named gates |
| Person at the desk | Judgment, money, exceptions | Anything | Their license and your rules |
A chatbot can be useful for hours and a published FAQ. It becomes a problem when it invents a Saturday slot. A seat is useful for first drafts. A thin receptionist is useful this afternoon. A custom agent is useful when the row has to match your no-book list.
A person at the desk is still the product in a clinic, a firm, and a truck. The agent is the leftover around that person.
Questions people ask when they mean this
The first conversational phrase is is an AI agent just a chatbot. Sometimes the vendor sold it that way. If it cannot use a tool and cannot write a row, it is a chatbot with better lighting. If it can place a Jobber hold and wait, it is an agent even if the voice sounds like a chatbot.
The second is does an AI agent replace staff. In a shop that already runs thin, it replaces retype. It does not replace the coordinator who knows which regular gets the last table. Shops that fire the phone person and hope the model will smile are buying a review.
The third is how is an agent different from ChatGPT. ChatGPT waits for you. An agent waits for the phone, the list, or the clock, then acts inside the job you wrote. You can build a clumsy agent on top of ChatGPT with copy-paste. That is still you in the loop on every row.
Those three questions are the definition in street language. If you can answer them for your shop, you do not need a glossary.
What the agent is allowed to touch
Allowed: collect fields you listed, map speech to those fields, write a request, send a sentence you approved, page a human on a written list, decline out-of-area ZIPs you listed.
A clean allow list looks boring on purpose. job_type from a map. zip_or_neighborhood from a zone sheet. callback_number repeated out loud. priority_flag from a list of ten phrases, not from the model’s mood. existing_customer only when the CRM match is clean. language persisted when you staff it.
Write the allow list as verbs: collect, map, write, send approved sentence, page, decline. If a vendor demo adds "delight," "upsell," or "prioritize revenue," ask which verb that is. If they cannot point to a field, it is not on the list.
Not allowed unless you wrote a rare rule and a person still reviews: money, discounts, legal or clinical advice, permit dates, "someone is on the way," deleting a customer, changing a contract.
A clean deny list is shorter and louder. No invented trucks. No invented chairs. No invented Saturday. No invented chemical. No invented legal right. No silent cancel of a recurring stop. No delete. No discount. If the caller asks for a deny-list item, the agent says it will place a human callback and stops.
San Diego shops should also treat recording and texts as counsel questions. Identify the line as automated. California two-party consent is widely discussed if you record. TCPA can apply to texts. This page is not legal advice.
Logs belong in your tenant. Keep VendorCallLog7Days or the equivalent export. An agent you cannot audit is a leftover you do not own. Keys stay in your vault. Nerlude exists because keys still land in Slack. The same rule applies to transcripts. If the only copy lives in a vendor dashboard you lose when the card expires, you automated a rental.
If two people on the team still disagree about a deny-list item, write the disagreement down and pick an owner. An agent that inherits an argument will pick a side at 10 p.m. and you will hear about it from a customer, not from a dashboard.
Where San Diego shops already use the shape
Trades use after-hours intake agents. Rooms use reservation and waitlist agents. Clinics use recall-draft agents. Studios use inquiry agents. Property desks use maintenance-request agents. Accounting desks in tax season use document-chase agents. None of those names require a research paper. They require a job sentence.
A San Diego clock still matters after the definition. Inland heat changes night volume. Tourism weekends change coastal rooms. Military report dates change movers and some family sessions. The agent does not need to "understand the city." It needs your zone list and your closed dates.
Workshops have been run in La Mesa when a room needs the same definition. A workshop is a shared definition. It is how a team stops calling every chatbot an agent.
Walk the same definition through a few shop nouns so it stops being abstract.
A mover agent collects date, origin ZIP, destination ZIP, and access, then writes a quote request. The stop is a binding price and a booked crew.
A pest agent collects pest type, ZIP, interior or exterior, and pets, then writes a route request. The stop is product name and a same-day promise.
An electrician agent collects job type, ZIP, and has_power, then writes a dispatch request. The stop is "we will be there in an hour" and a permit date.
A photography agent collects session type, date, and location, then writes a HoneyBook lead. The stop is a custom package and "you are booked."
A law-office agent collects matter type and a callback, then writes an intake row. The stop is legal advice and a fee quote. Read AI agent for law firm for the desk shape. This page stays on the definition.
Each of those leftovers uses a model. Each is an agent only because tools and a write path exist. If you remove Jobber, HoneyBook, or the matter list, you are back to a chat.
People also ask whether an agent "has memory." Memory in a shop is the system of record. The last job, the last no-book date, the last Spanish preference. A chat history in a personal account is not memory you can own. A PreferredLanguage field on the customer row is.
People ask whether an agent "runs overnight." Overnight is a schedule. After-hours forward is a schedule. A weekly recall export is a schedule. If nobody wrote the schedule, the leftover waits for a hero.
People ask whether an agent is "autonomous." In this studio’s leftovers, autonomy stops at the gate. The useful autonomy is "collect fields at 10 p.m. without waking the owner." The useless autonomy is "decide a price because the caller sounded urgent."
When a custom agent is the leftover
Buy a seat when the job is drafting. Buy a $50/mo receptionist when the job is a name and a callback and you will retype for a while. Buy a custom leftover when the hold has to match zones, language, and a no-book list the vendor template will not honor.
Cost bands, not quotes: readiness call is free; a scoped audit often lands in the low thousands; a first custom agent or integration often lands in a mid four- to five-figure project band depending on systems. Fit is roughly 30-250 people, plus smaller shops if the workflow is clear.
If you want help naming the job and the stop, book a free 15-minute AI readiness call. Email cyriac@nidrosoft.com or use https://calendar.app.google/GCtfuTx3Ms6K7d579. You can also write the four steps and never call.
Frequently asked questions
Is a voice receptionist an AI agent?
It is an agent if it uses tools and writes a row, even a thin one. It is a voice chatbot if it only talks and emails you a novel. Many $50/mo products are thin agents. That can be enough.
Does the agent learn my business by itself?
No. It follows the job, the fields, and the stops you wrote. If you want it to "learn," you are asking for a weekly review of failures and a written change. That is Own. A chat history in a personal account is not learning you can hand to Tuesday’s coordinator.
Can one agent do the phone, the inbox, and the ads?
Not as a first leftover. One job. One write path. Shops that start with three doors get three messes. Pick the door that already produces paid work, finish 30 days, then write a second job sentence if the first leftover still lives.
Is this the same as robotic process automation (RPA)?
RPA clicks the same buttons every time. An agent can read messy speech and still write structured fields. A lot of shops need both: the agent for the call, a simple automation for the row that is already clean. If the row starts as a voicemail, start with the agent. If the row is already a field, start with glue.
Do I need an engineer on staff?
No. You need an operator, an inheritor, and an admin. An engineer leftover appears when the write path is custom and someone has to keep keys and logs. That someone can be this studio or a person you already trust.
Write the job, then name the stop
What is an AI agent, once the glossary is gone: a job, tools, a row, and a stop. Write those four on a page. If a vendor cannot point to each one, you are buying a chat.
Book a free 15-minute AI readiness call if you want a second set of eyes on the job sentence. If the sentence is already clear, decide seats versus a write path this week.