How to work more efficiently with AI is a habit change on tools you already pay for, plus one owned playbook. Time three Tuesdays. Save prompts next to the files. Draft in the model, send as a person. Measure hours on the job you already track. Custom software waits until the playbook still leaves a Friday Sheet.
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
- Efficiency is hours back on a named job, not a new tab you open when you feel behind.
- Four habits: time the week, save prompts beside the files, draft then send as a person, kill prompts nobody asked for.
- Seats win for drafts. A $50/mo receptionist SaaS wins for pickup. A custom queue wins when two people still double-text.
- Token waste is a real leak. TokenTra teams have cut 10-30% of wasted spend when they can see the bill. NanoBrief-shaped speed is under 2 minutes versus 3+ hours for a brief someone requested.
- Book a free 15-minute AI readiness call if the playbook is in use and Friday still dies in a Sheet.
Contents
- What efficient AI work looks like on a Tuesday
- How to work more efficiently with AI in four habits
- A typical accounting desk before and after
- A week you can print
- Where people lose the hours they thought they saved
- Seats, playbooks, and when a queue should take over
- Token spend and briefs that nobody asked for
- How Nidrosoft trains a crew to keep the hours
- Frequently asked questions
- Keep the playbook next to the files
What efficient AI work looks like on a Tuesday
Efficient work with a model is boring on purpose. The operator opens the same Drive folder, the same Jobber or QuickBooks or HubSpot view, and a short prompt card that names the fields. They paste the source, get a draft, edit the three lines that matter, and send. They do not open five chat threads to rediscover the voice. They do not paste a full client folder into a personal account. They do not ask the model to invent a number that lives in the ledger.
People ask how can I use AI to get through my inbox faster. The honest answer is a triage card: archive, draft, escalate. The model drafts the middle pile. A person still sends anything with money, a filing, or a promise of a date. People ask what does an efficient AI workday look like. It looks like the same calendar with shorter draft blocks and a written stop time, not a longer day stuffed with more generations.
This page has no city in the query. The studio still sits in San Diego County. If you want the implementation pillar after the habits stick, read AI implementation for small business. If the painful job is client follow-up in a tax shop, read AI agent for accounting firm. If you want a city map later, AI implementation Encinitas is one coastal cut.
Workshops have been run in La Mesa. A room can install the four habits. It will not empty a queue that needs a lock.
How to work more efficiently with AI in four habits
-
Time three Tuesdays before you change tools. Write
JobName,HoursThisWeek, andAlreadyTried(ChatGPT, Copilot, a prior intern). If you cannot name the hours, you will buy a seat and still feel behind. A grounded range for a solo operator who drafts estimates or client emails is often 6-10 hours a week on first drafts. A coordinator who replays voicemail can sit in the 8-12 hour band. Time your shop. Do not import someone else's number. -
Save one prompt card per job, next to the files, in a tenant you can revoke. Custom instructions that live in ten personal accounts are not a playbook. Name the voice, the forbidden claims, the field list, and the human gate. People ask how do I stop wasting time prompting from scratch. You stop by refusing to start from scratch. The card is the product.
-
Draft in the model. Send as a person. Money, clinical notes, customer promises, and filings stay gated. Identify as AI when a caller is on a phone agent. If you record, California is widely reported as two-party consent. This is not legal advice. Spanish-first threads need a signed script, not a hopeful translation.
-
Kill work the model invented. A brief nobody asked for is waste. A second summary of a meeting that already has minutes is waste. A rewrite of a clean email is waste. NanoBrief is the public pattern for speed on a requested brief (under 2 minutes vs 3+ hours). Copy the gate, not the product name.
Do these in order. Buying a fifth seat before you time the week is how a 20-person shop adds a line item and no hours.
A typical accounting desk before and after
This is a composite, labeled as a typical shop, not a named customer.
Before: a 22-person accounting firm. Two staff accountants and one office manager share client email. Each person has a personal ChatGPT login. Partner review still happens on every outbound note because the voice drifts and someone once pasted a Social Security number into a thread. April looks like a 60-hour week with 12-15 hours of first-draft emails and organizer chases. The unofficial stack is a desktop folder of last year's letters and a Slack channel where people paste "make this nicer."
After the four habits, not a custom agent yet: one firm ChatGPT or Copilot workspace. Three prompt cards: organizer chase, extension notice draft, and "questions for the partner" (the model lists missing items; it does not file). A one-page data rule: no SSNs, no full returns, no bank logins. The office manager owns the cards. Staff still send. Partners still sign anything that changes a filing date. Draft time on those three jobs drops from the 12-15 hour band toward 4-6 hours of edit and send. The remaining leak is a follow-up queue with no NextAttemptAt. That leak is a later agent, described on the accounting vertical page, not a fifth prompt.
The 4-6 hour remainder is a grounded desk range for shops that already had a repeating letter. Your April will differ.
A week you can print
| Day | Job | Tool | Gate | Stop rule |
|---|---|---|---|---|
| Mon | Inbox triage | Seat + triage card | Person sends | 45 minutes, then the live book |
| Tue | Recurring letters or estimates | Prompt card beside files | Partner or estimator | No second rewrite |
| Wed | Meeting notes to action list | Seat | Owner confirms owners | No extra summary |
| Thu | Follow-up list | Sheet or CRM | Person clicks send | If the list has no dates, do not generate more copy |
| Fri | Token or seat review | Invoice or TokenTra-shaped view | Admin | Cut unused seats; cut prompts nobody ran |
Print it. If Thursday has no dates, you do not have an efficiency problem on Friday. You have a queue problem. AI implementation for small business is the path for that leftover.
Fit is roughly 30-250 people, plus smaller shops if the workflow is clear. A 8-person shop can run this week if one person owns the cards. A 80-person shop that lets every department invent a voice will spend the seats and keep the 12-hour draft pile.
Where people lose the hours they thought they saved
They regenerate. The first draft was good enough. The third draft is later and not better. Put a stop rule on the card.
They paste secrets. Then counsel pauses the whole program. Efficiency becomes a shutdown. Write the data rule in week one.
They prompt in the car on a personal phone and cannot find the thread on Tuesday. The playbook has to live where the files live.
They ask the model for a number that is in QuickBooks. The model will sound sure. The ledger is still the ledger.
They buy a receptionist SaaS and still replay the vendor log into the CRM by hand. Pickup got better. The morning retype did not. That is a write-path job, not a prompting job.
They run a workshop and treat the burst of energy as the new baseline. Workshops have been run in La Mesa. The hours hold only if the cards are still open in week six.
They ignore commute and coverage. An Encinitas operator on Coast Highway walk-in days cannot "just prompt more." Coverage is a staffing fact. The model does not sit the desk.
Seats, playbooks, and when a queue should take over
| Leftover | Hours it can return | What it cannot do |
|---|---|---|
| ChatGPT or Copilot seats | Draft time on repeating letters | Lock a follow-up so two people do not text |
| Shared playbook in Drive | Consistency and revocation | Answer the trunk at 8 p.m. |
| Receptionist SaaS (~$50/mo class) | Pickup on a short script | Trusted Jobber or Open Dental writeback |
| Custom queue | State: last attempt, next attempt, do-not-contact | Replace partner judgment on a filing |
Buy seats when how to work more efficiently with AI is the whole question. Buy a queue when the playbook is in use and the Sheet still has no lock. Buy an audit when you cannot rank the job. Typical bands: readiness call is free; a scoped audit often lands in the low thousands; a first custom agent often lands in a mid four- to five-figure project band depending on systems.
If the ranked job is client chase inside a tax practice, the vertical write-up is AI agent for accounting firm. If you later want a North County room or a remote embed, AI implementation Encinitas is a city page, not a requirement for this habit work.
Token spend and briefs that nobody asked for
Software teams and any shop that wires models into a product will leak money on retries, forgotten evals, and features that call the model twice. TokenTra is the public meter. Teams have cut 10-30% of wasted spend when they can see the bill. An owner who only buys ChatGPT seats still has a smaller version of the same leak: five people generating unused summaries.
Write FeatureName or PromptCardName and a monthly cap you will actually look at. Nerlude exists because keys still land in Slack. Put keys in a vault. Protectron is the public EU Artificial Intelligence Act artifact pattern. It is not legal advice and not a reason to generate more internal memos.
NanoBrief is the other public number: under 2 minutes versus 3+ hours when someone asked for the brief. The efficiency move is to refuse the brief when nobody asked. That sentence saves more hours than a better model.
How Nidrosoft trains a crew to keep the hours
The method is Embed, Diagnose, Scope, Build, Own. For efficiency work, Embed is watching Tuesday, not a vision offsite. Diagnose is the hour table and the leftover list. Scope is often "playbook plus seats" and a written note that a queue is out of scope. Build, if it happens, is one write path. Own is the inheritor and a 30-day metric you already track.
Cyriac has 12+ years, 125 products shipped, 16 public, formerly Anthropic, Microsoft (Bing and Edge), Intuit, and Gap Inc. Live systems you can open: Omnira Dental, Octana, Eviction Wizard, Guidera, TokenTra, Protectron, Nerlude. Those are ownership patterns. They are not a Tuesday prompt card.
Workshops have been run in La Mesa. A private team session can install the four habits on your files. Next work after the room is either a quiet month of the printed week, or a build if Thursday still has no dates.
Book a free 15-minute AI readiness call if you already know the playbook is not enough. Bring three Tuesdays of hours and the name of the card owner.
Prompt cards, meetings, and a 30-day drill
A card is a short spec, not a personality. Print this shape and fill the brackets with your nouns.
JobName: organizer chase. Audience: client who has not returned documents. SourceFields: ClientName, MissingItems, DueDate from the organizer, PartnerName. Forbidden: invent a due date, cite tax law, promise a result. Voice: plain, short, no jokes about the IRS. HumanGate: staff sends after a glance; partner sends if the date moves. MaxPasses: 1. StopRule: if MissingItems is empty, do not generate.
JobName: estimate narrative. SourceFields: JobType, ZipCode, SiteNotes from the camera roll, Exclusions the estimator already wrote on paper. Forbidden: a dollar amount, a schedule, a material the truck does not carry. HumanGate: estimator prices after the draft. MaxPasses: 1.
JobName: inbox triage labels. SourceFields: subject and first six lines only. Output: archive, draft, or escalate. Forbidden: sending. HumanGate: the person who owns the inbox.
Meetings waste hours when the model writes a second memoir. The efficient version is a 12-line action list: Owner, Due, SourceQuote (a short phrase from the call). If the meeting already has minutes in HubSpot or a notebook, skip generation. If three people need a brief the partner asked for, use a NanoBrief-shaped pass (under 2 minutes vs 3+ hours) and stop.
A 30-day drill you can run without a studio:
Days 1-3. Time the unofficial jobs. Write HoursThisWeek. Do not change tools.
Days 4-5. Move repeating work into a company workspace. Revoke one personal thread you find. Put keys in a vault. Nerlude exists because this step still fails.
Days 6-10. Write three cards. Run them on live work. Log regenerations. If a card needs a third rewrite, the card is wrong. Fix the fields. Do not buy a better model yet.
Days 11-17. Add the printed week table. Thursday must have dates or you write "queue leak" on the calendar. Token or seat invoice glance on Friday. Cut unused seats. TokenTra's 10-30% wasted-spend cut is the product-team version of this glance.
Days 18-24. Spanish-first or two-location exceptions, if you have them. Write the handoff sentence. Identify as AI if a phone leftover is in play. California recording consent is widely reported as two-party. This is not legal advice.
Days 25-30. Read the hour log against day 1. If draft hours dropped and Thursday still has no dates, you learned the boundary: habits worked, state did not. Open AI implementation for small business. If hours did not drop, the cards are closed or the job was never repeating. Do not add a custom agent to hide that.
Encinitas and other coastal shops should schedule the drill around walk-in hours. An efficiency program that needs a quiet 11:40 a.m. will lose to the waiting room. AI implementation Encinitas is optional geography, not a requirement for the cards.
Workshops have been run in La Mesa. If the crew will not write cards alone, book a room, then restart the 30-day drill. A room without the drill is a morning you enjoyed.
Cyriac has 12+ years on this kind of install, 125 products shipped, 16 public. The drill still belongs to your office manager. The studio cannot inherit Tuesday.
Frequently asked questions
How to work more efficiently with AI if I only have a personal ChatGPT account?
Move the repeating jobs into a workspace the company can revoke. Save the cards in Drive. Keep personal threads for personal mail. Efficiency that depends on one person's history dies when that person is in traffic.
Will a better model make me faster than a playbook?
A better model can shorten a draft. It will not find last week's card, honor your data rule, or stop a second rewrite. Buy the card first. Change models when the card is stable and the remaining minutes are still expensive.
How many hours should I expect back?
Grounded ranges, not a study: 6-10 hours of draft time for a solo operator who already repeats letters; 8-12 hours of replay time for a coordinator on a noisy trunk, if you later add pickup or a queue. Time your shop. If the number is 2 hours, buy seats and stop. If the number is 15 and the Sheet has no dates, plan a build.
Is a custom agent required to work more efficiently with AI?
No. Custom is for state. Habits are for drafts. Mixing them in one invoice without a boundary produces software nobody asked to operate.
What should I bring to a readiness call?
JobName, hours from three Tuesdays, tools you already pay for, and whether an inheritor exists. The call is free. Email cyriac@nidrosoft.com or use https://calendar.app.google/GCtfuTx3Ms6K7d579.
Keep the playbook next to the files
Time the week. Save the card. Send as a person. Cut the extra brief. If Thursday still has no dates, you need a queue, and AI implementation for small business is the next page. If you want a human to rank the leftover, book a free 15-minute AI readiness call.