Sainso Technology AI Systems Integration
← Home All solutions 繁中 Book a Call →
White Paper · 2026 / 9

AI Digital Employee
Implementation White Paper

Sainso Technology designs AI as a job role, with a defined scope, permissions, direct supervisor, handoff rules, and performance metrics. This white paper covers six digital employee roles already in production: what they do, how they connect to existing systems, what went wrong after launch, how we fixed it, and how to calculate implementation benefits. Digital employees can work on the platforms customers already use, including LINE, WhatsApp, Telegram, and WeChat, with language and data deployment regions configured for each market.

Data through 2026-09-30 6 roles 10 field incidents Multichannel · Multilingual Built and operated by our Taiwan team
Book a 30-minute diagnostic → Calculate benefits now
01 · Summary

Four key points

If you do not have time to read the full paper, these four points summarize how we implement AI digital employees.

i

One digital employee is one job role

Write the job description first: what the role owns, what it must not touch, and when it hands off to a person. Only then choose the model and connected systems.

ii

Facts come from your systems; AI handles understanding and expression

Prices, inventory, and time slots always come from your database. AI only determines what the customer wants. If the answer cannot be found, the case goes to a person; AI does not guess.

iii

We run it on ourselves first

Digital employees handle Sainso's website content, social channels, and operations monitoring every day. Every failure listed in this white paper has been fixed and documented as standard practice.

iv

Measure benefits with the same yardstick

Measure a baseline before implementation, then use the same metrics for pilot acceptance. We make the case with your numbers, not illustrative figures.

02 · Definition

What is an AI digital employee?

A chatbot answers questions; a digital employee completes a job. The role has permissions and handoff rules, and every step is logged. The flow below shows the full path for handling one inquiry.

CHANNEL

Work where customers are

LINE, WhatsApp, Telegram, WeChat, website, email, or an internal admin console.

UNDERSTAND

Determine intent

The AI model distinguishes requests for pricing, bookings, rescheduling, complaints, and other needs.

VERIFY

Read facts from your systems

ERP, appointment calendar, product catalog, and price list. AI does not invent answers.

GUARD

Check every item before sending

Numbers, links, and personal data must match the source. When uncertain, stop.

DELIVER

Complete the work

Reply to the customer, generate a draft, open a work order, or notify the store.

HAND OFF TO A PERSONComplaints, refunds, personal data, and out-of-scope questions go to the owner with the conversation history and a suggested reply.
WORK LEDGEREvery decision, usage cost, and error code is logged for auditing, traceability, and cost accounting.
ComparisonTraditional chatbotAdditional or outsourced staffAI digital employee
Basis for answersPrewritten scriptsPersonal experience and memoryYour system data plus explicit rules
Unknown questionsReturns a stock message or loopsHandled by individual judgmentHands off by rule with a suggested reply
Service hours24 hoursScheduled shifts24 hours
System integrationUsually nonePeople switch screens to look things upConnects to ERP, booking, and CRM; every write-back item requires approval
Quality managementHard to measureManager spot checksEvery interaction is logged; the standard test set runs before a model change
Cost structureLicense feeSalary, labor and health insurance, training, and turnover costsBuild fee, monthly operations fee, and per-interaction usage

Work on the platforms customers already use

The job description, guardrails, and work ledger do not change by channel. Only the channel integration and language test set change. The same digital employee can connect to different messaging platforms and serve customers in different countries.

ChannelCommon marketsCurrent status
LINETaiwan, Japan, ThailandIn production: tire distributor store inquiries and demo-store smart customer service
WebsiteGlobalIn production: website content and online contracting
TelegramGlobalInternal production: operations alert notifications and internal commands
WhatsAppSoutheast Asia, South Asia, Europe, the Middle East, and Latin AmericaIntegration available, but no production case yet; WhatsApp Business account review required
WeChatMainland China and Chinese-speaking communities worldwideIntegration available, but no production case yet; an Official Account or Service Account must be obtained under local regulations
Messenger／InstagramGlobalIntegration available, but no production case yet
EmailGlobalPartially complete: notification email is in use; AI email replies are not yet live

Multilingual support: the content role has published 123 bilingual Chinese and English articles; customer-service roles currently operate in production in Traditional Chinese. Before adding any language, we build a standard test set and guardrails for it—quantifiers, dates, and telephone formats vary by country—and enable it only after it passes. The data deployment region can be chosen based on the client location and regulations.

03 · Roster

Digital employee roster

All six roles below are deployed in production on Sainso-owned or client systems. LINE smart customer service currently operates through a demo store, and the social editor has posted on Sainso’s production Facebook Page since 2026-09-29 onward. Each ID badge states the role’s permissions and handoff owner; the production record gives the measurement date and scope.

Sainso Digital Staff
DS-01 · CUSTOMER SERVICE

LINE Smart Customer Service Specialist

SHIFT
24 hours
CHANNELS
LINE; can connect to WhatsApp, WeChat, Messenger
ACCESS
Read store services and pricing; reply to customers
HANDOFF
Complaints, refunds, personal data, out-of-scope requests
MANAGER
Store manager

RESPONSIBILITIES

  • Answer questions about prices, service duration, and available times
  • Provide the correct booking link
  • Draft a reply for store staff when handing off
PRODUCTION RECORD

In a live-device test at the demo store in production, it returned the correct price and booking link within 2 seconds at a cost of US$0.00066 per reply (2026-09-14). Any model change must pass the standard test set; the latest score is 100 out of 100 points.

Sainso Digital Staff
DS-02 · STORE DESK

Store Support Clerk (ERP × LINE)

SHIFT
24 hours
CHANNEL
LINE
ACCESS
Read ERP products and inventory; show admin-maintained prices by membership tier
HANDOFF
New-contact approval and order confirmation
MANAGER
Store supervisor

RESPONSIBILITIES

  • Check inventory by warehouse and product specifications
  • Quote by membership tier with automatically consistent discounts
  • Place orders through LINE and notify the store
PRODUCTION RECORD

The tire distributor client launched in 12 phases beginning on the recorded date (2026-06-16), and 159 existing member accounts have been approved. Price and inventory lookups allow no error, so this role deliberately uses a rules engine, not generative AI.

Sainso Digital Staff
DS-03 · FRONT DESK

Appointment Desk Specialist

SHIFT
24 hours
CHANNELS
LINE chat and website; can connect to WhatsApp and WeChat
ACCESS
Check time slots; create and modify bookings
HANDOFF
Open a work order for store staff
MANAGER
Store manager

RESPONSIBILITIES

  • Determine intent and open a booking card directly
  • Answer common questions
  • Let customers check and reschedule bookings themselves
PRODUCTION RECORD

During development acceptance, the conversation required to complete a booking fell from 12–16 turns to 2 turns. It passed 739 acceptance tests before launch. When 10 people competed for 2 openings simultaneously, exactly 2 bookings succeeded. Booking writes never pass through AI.

Sainso Digital Staff
DS-04 · CONTENT

Content and Social Editor

SHIFT
Posts at fixed times every day
CHANNELS
Website and Facebook Page
ACCESS
Write drafts; publish only to authorized accounts
HANDOFF
Drafts below the score threshold
MANAGER
Marketing lead

RESPONSIBILITIES

  • Select and score topics from public sources
  • Write bilingual Chinese and English articles and social posts
  • Generate covers and publish only after quality checks pass
PRODUCTION RECORD

The Sainso website has automatically published 123 bilingual Chinese and English articles since 2026-06-28 onward. The social pipeline has posted automatically to Sainso’s production Facebook Page since 2026-09-29: at most 2 posts per day, 6 hours apart, and only with a score above 0.7 to publish.

Sainso Digital Staff
DS-05 · DOCUMENTS

Document and Contracting Specialist

SHIFT
Activated per project
CHANNELS
Website signing page and email notifications
ACCESS
Generate quotes, contracts, and signing links
HANDOFF
Send signing links and affix seals
MANAGER
Sales lead

RESPONSIBILITIES

  • Produce project quotes and contracts (PDF and Word)
  • Create online signing and track signing status
  • Attach the PDF for archiving after signing completes
PRODUCTION RECORD

Sainso’s online contracting platform is live with 9 built-in, customizable contract templates. It was used for an actual client contract on the recorded date (2026-09-17), with both parties signing online. A person always sends the signing link.

Sainso Digital Staff
DS-06 · OPERATIONS

Operations Duty Officer

SHIFT
24 hours, scheduled inspections
CHANNELS
Telegram notifications and internal admin console
ACCESS
Read monitoring data; generate reports and drafts
HANDOFF
Deployments, sending contracts, payments, and external email
MANAGER
Operations lead

RESPONSIBILITIES

  • Daily operations brief
  • Infrastructure and security inspections
  • Compile tender and grant intelligence
  • Draft business development messages and proposals without sending them automatically
PRODUCTION RECORD

A dedicated host runs 24 hours a day and has accumulated 4,178 operations reports since 2026-06-06 (counted 2026-09-30). Whenever this role needs to deploy, send a contract, or move money, the action always enters a human approval queue.

04 · Case Files

Real production records

Client cases are anonymized under confidentiality agreements. Each record states the starting point, approach, current status, and problems encountered after launch.

Client projectAutomotive · Tire retailDS-02

Tire distributor: ERP × LINE store inquiries

STARTING POINT
List prices, specifications, and inventory by warehouse were all in the ERP. Store and customer quotes required manual lookups, and staff had to remember VIP discounts.
APPROACH
We integrated the existing ERP: products and inventory are read-only; inventory counts entered through the admin console write to ERP count sheets without changing the product master. Seven membership prices are maintained in the admin console we built. We created the admin console and LINE inquiry clerk, then launched in 12 phases, opening each phase only after validating the previous one.
CURRENT STATUS
Phases 11 and 12—member approval and the daily downtime window—were completed on 2026-09-26 as recorded. 159 existing member accounts are approved; new contacts can check prices and place orders only after approval.
POST-LAUNCH FIXES
Multi-warehouse inventory originally returned only the final warehouse; it now sums all warehouses. When the admin console is temporarily unavailable, the system explicitly says “the system is busy” instead of returning potentially stale data.
Sainso-owned platformServices · Single locationDS-01 · DS-03

Service-business platform: LINE smart customer service and AI copywriting

STARTING POINT
Small stores often have one person on duty. They cannot reply while serving customers, and after-hours inquiries wait until the next day.
APPROACH
Each store uses its own LINE Official Account, with credentials stored encrypted. The model only determines intent; the system composes customer-facing sentences from store data. Across the platform’s 19 AI services, 6 make no large-language-model calls at all.
CURRENT STATUS
A live-device test at the demo store in production returned a reply in 2 seconds on 2026-09-14 as recorded. The platform moved to the Google Cloud Taiwan region beginning 2026-09-15; migration used backup restoration with row counts compared table by table. It currently operates as a demo store while we recruit the first implementation customers.
POST-LAUNCH FIXES
The standard test set passed in full, yet correct answers in real conversations were handed to people. Guardrails now remove only the problematic fragment. After an AI provider changed a parameter specification and replies failed, we added automatic fallback and logging of the original error code.
Internal Sainso useContent marketingDS-04

Sainso website: 123 articles from the AI content editor in three months

Articles published automatically each month: month 6 had 6; month 7 had 28; month 8 had 40; and month 9 had 49 0 25 50 6 28 40 49 Month 6* Month 7 Month 8 Month 9
Articles published automatically each month, 123 total, each in Chinese and English. *Month 6 counted from the 28th. Source: website publication records.
QUALITY GATES
At least 3 section headings, a required list, deduplication against existing articles, and blocking of suspected fabricated statistics.
POST-LAUNCH FIXES
A configuration-format change caused 36 consecutive failures and an 18-day outage. Now an alert is sent after 3 consecutive failures. The program still reported success after the topic pool ran out; now it adds 40 topics automatically when fewer than 20 remain and reports failure when exhausted.
Client projectBrand · Social advertisingDS-05

Northern Taiwan brand: first week of managed social advertising

STARTING POINT
The client wanted us to manage social advertising, with everything from quoting to contracting completed remotely.
APPROACH
The quote, contract, and signing were completed entirely online. Both parties signed online on the recorded date (2026-09-17), with media spend and service fees itemized separately.
CURRENT STATUS
The first campaign is being prepared for launch.
WHAT WE LEARNED
Control of advertising accounts, Pages, and business assets must be inventoried clearly in the first week of implementation, or launch will be delayed. This is now mandatory in the first week of every project.
05 · Field Notes

What only happens after launch—we have seen it all

The real test begins after a digital employee launches. The ten incidents below happened in production or were caught in testing. Every issue has been fixed and incorporated into our standard practice. Each entry first states what happened, then what we do now.

A green status light does not mean the customer received it

RELIABILITY
  • Every auto-reply switch was on, yet not one message was sent. The trigger was “outside business hours,” but business hours were set to 24 hours, and the admin console showed no error.
  • A bulk re-engagement email reached only 1 of 3 eligible dormant customers, yet the system reported success. The deduplication identifier omitted the recipient. (Caught in testing.)
WHAT WE DO NOW

Before a new automated message launches, we receive it once through a real customer-side account and channel, confirming that it includes the store name and a working link. Batch jobs must verify that “N eligible” equals “N received.”

AI guardrails must block wrong answers and allow correct ones

ACCURACY
  • The standard test set passed 30/30, yet a real-device LINE conversation handed a correct answer to a person because “which time” was classified as a fabricated number.
  • An AI provider changed the specification for an optional parameter, and the AI could not answer the customer’s first message.
WHAT WE DO NOW

A fabricated number or unofficial link blocks the whole message and hands it to a person; a minor formatting issue is corrected only in that fragment, and the message is sent normally. Every handoff records its reason, and we reproduce the same real conversation more than 6 times before changing the rules. Unsupported external options trigger automatic fallback while preserving the original error code.

One store’s problem must not take down every store

MULTI-STORE ISOLATION
  • If one store saved a day-count field as text, birthday coupons, re-engagement, and loyalty automation stopped for every store. (Caught in testing.)
  • The nightly job reported success but processed no records. The relevant tables held 29 to 107 records each, while the job saw 0 records. Data-isolation rules were working, but the job lacked a store identity. (Caught in testing.)
WHAT WE DO NOW

Settings are validated when written. Scheduled jobs run independently for each store, so one store’s failure does not affect others. Tests deliberately include a “neighbor with a bad value” to confirm that other stores continue normally.

Silent failure costs the most

MONITORING & ALERTS
  • Automatic publishing failed 36 consecutive times after a configuration-format change, and no one noticed for 18 days.
  • A health check did not run for 77 days because its task name did not match, while the daily summary still showed normal status every day.
WHAT WE DO NOW

Every failure path notifies the owner; reporting does not happen only on success. Health checks confirm whether the most recent run completed, and the schedule list is regularly compared with actual execution records.

Privacy promises must cover every output

PRIVACY & DATA ACCURACY
  • The satisfaction survey screen showed only an average score, but the export included booking IDs and could be linked back to respondents.
  • The program calculated “today” in UTC while the database used Taipei time, so reports from 0 to 8 o’clock every day undercounted that day’s results.
WHAT WE DO NOW

For data promised to be anonymous, the data structure stores no field that can link back to a person. We inspect screens, APIs, exports, and system logs individually. The operating time zone is standardized on Taipei time and covered by automated tests.

06 · Benefits

How to calculate benefits

Every company has different inquiry volumes, labor costs, and workflows. We do not use someone else’s percentages to persuade you. Below are three structural benefits that can be calculated directly, followed by a calculator for your numbers. Formal figures are measured during the pilot.

4.2×

Service hours available

168 hours per week versus 40 hours from 9 to 17 on Monday through Friday. Night and holiday inquiries can also receive immediate replies.

≈ 2cents

Cost of one AI reply

A LINE AI reply at the demo store cost US$0.00066 in testing (2026-09-14). At an estimated exchange rate of NT$32 per 1 US dollar, that is about NT$0.02 per reply. Ten thousand replies cost about US$6.6 in total.

Same questionSame answer

Answers do not change by person

Prices, inventory, and time slots are read from the system. They do not change across shifts, staff changes, or seasons. The system applies discount rules instead of relying on staff memory.

Benefit calculator

Defaults are examples, not client results. Replace them with your company’s figures; replace “share AI can handle automatically” with the measured pilot value.

inquiries/month
minutes
%
NT$/hour
NT$/inquiry
Converted from the measured US$0.00066 per inquiry; complex roles may cost more.
NT$/month
NT$

Calculated results (monthly)

Monthly net benefit—
Human hours freed—
Equivalent labor cost—
AI usage cost—
Build-fee payback period—
AI usage + operations feesNet benefit

Hours freed = inquiries × automation share × handling minutes ÷ 60
Net benefit = hours freed × labor cost − AI usage cost − operations fee
Payback period = build fee ÷ net benefit

The calculator includes only quantifiable hours. It excludes conversions from nighttime inquiries, return visits from faster replies, and errors avoided through consistent quoting; these are measured separately during the pilot.

Pilot acceptance metricPre-implementation baselinePilot acceptance method
First response timeSample 2 weeks of message records and calculate the medianThe system records the response time for every message
Automation shareAll handled by peopleShare not handed to a person and not followed by another customer question
Incorrect-answer rateNo recordStore manager scores a random sample of 50 each week
Handoff reasonsNo recordEvery human handoff records a reason for categorized analysis
Cost per inquiryLabor cost × handling timeActual usage in the work ledger
Human hoursShift schedule or time recordsMeasure again with the same method
07 · Process

Implementation process

A role takes about 6 to 13 weeks from discovery to production, depending on scope. Every step has a defined deliverable, and the next begins only after the previous one is accepted.

  1. 012–4 weeks

    Discovery

    Interview frontline staff, break down workflows, inventory data sources and account permissions, and measure the pre-implementation baseline. Deliver a job description and benefit baseline.

  2. 02about 1 week

    Role design

    Define scope, permissions, human-handoff conditions, and performance metrics, then build a standard test set specific to the role.

  3. 032–6 weeks

    Integration build

    Connect messaging platforms such as LINE, WhatsApp, Telegram, and WeChat, plus the website, ERP, booking system, or CRM. Read-only is the default; write-back items, such as inventory counts and orders, are listed individually and implemented only with your approval.

  4. 041–2 weeks

    Shadow deployment

    AI generates drafts first, and a person confirms them before sending. We compare answers and adjust guardrails until the role can operate independently.

  5. 05launch day

    Production launch

    Enable automatic replies. Every interaction is logged, and anomalies immediately notify the owner.

  6. 06monthly

    Monthly review

    Spot-check reply quality, analyze handoff reasons, and provide cost reports; rerun the standard test set before changing models.

08 · Security & Compliance

Security & compliance

Digital employees touch customer data and operational systems. These are the practices we use, with the scope currently applied stated explicitly.

Data stays in your chosen region

The service-business platform has been deployed in the Google Cloud Taiwan region since the recorded date (2026-09-15), with daily automatic backups and 7 days of point-in-time recovery. Overseas projects can choose a deployment region based on local regulations.

Recovery is actually rehearsed

A drill with demo data on 2026-09-12 took 87 seconds to export and 3 seconds to restore. Row counts matched table by table, and isolation rules remained effective across all 98 store tables. On the recorded date (2026-09-14), another backup was restored to the Taiwan-region database and matched table by table.

Mask data before sending it to AI

Telephone numbers, email addresses, and URLs can be masked before transmission, and prompts and outputs can be configured not to be stored. Scope is determined during discovery based on your data types.

Your data is not used for training

Every AI request from the service-business platform requires zero data retention. If a provider does not support it, the request automatically switches to prohibition on training use and leaves a record. Your project can use the same setting.

People approve high-risk actions

Whenever the operations duty officer needs to deploy, send a contract, move money, or send external email, the action always enters a human approval queue. A person sends every signing link.

Pass the test set before changing models

Before the service-business platform changes its AI model, the model must pass the standard test set before launch. Feature and role names are checked for trademarks before release.

09 · Pricing

Plans and pricing

Start with one role and one channel. All prices exclude tax and are reference ranges; final quotes depend on the number of roles and integration scope. See the Plans and Pricing page for full details.

Discover

Discovery and diagnostic

NT$30,000and up

2–4 weeks, producing a job description, benefit baseline, and implementation recommendations. The fee can be credited toward a later project.

Build

Digital employee build

NT$150,000and up

One-time build. Light implementations range from 15–40 in units of ten thousand NT$, and standard integrations from 40–120 in units of ten thousand NT$, tiered by the number of roles and connected systems.

Subscribe

Monthly platform-module subscription

NT$3,000and up/month

Subscribe directly when a ready-made module exists for the role, with tiers based on module and usage.

Operate

Operations and monthly review

NT$8,000and up/month

Monitoring and alerts, monthly review, test-set maintenance, and model updates; alternatively, 12–20% of the annual project value.

10 · FAQ

Frequently asked questions

What happens when AI gives a wrong answer?

Your systems provide prices, inventory, and time slots; AI only determines intent. Guardrails block content that does not match the data and hand it to a person. The standard test set runs before launch, every interaction is logged after launch, and replies are spot-checked in the monthly review.

Will our data be used to train models?

It can be configured not to be. The service-business platform requires zero data retention on every request. If a provider does not support it, the request automatically switches to prohibition on training use and leaves a record. Your project can use the same setting, and personal data can be masked before it reaches the model.

Do we need to replace our current ERP or booking system?

No. We integrate your existing systems. The tire distributor case retained its original ERP; products and inventory are read-only, prices are maintained in the admin console we built, and only inventory counts entered through the console write to ERP count sheets.

Is it limited to LINE? Can it serve overseas markets?

It is not limited to LINE. The same role definition can connect to LINE, WhatsApp, Telegram, WeChat, Messenger, a website, or email. LINE and websites currently operate in production, and Telegram is used for our own operations alerts. Other platforms can be integrated but have no production cases yet; their business-account and review requirements are confirmed during discovery. Reply language is configured by market, and every language must pass its own standard test set before launch. The data deployment region can also be chosen under local regulations.

Will employees be replaced?

Digital employees take over repetitive, rule-based work. Your team remains responsible for judgment, exception handling, and customer relationships. Handoff rules are defined with frontline staff.

Can we start with a small pilot?

Yes. Start with one role and one channel, proceed through shadow deployment and a production pilot, then decide whether to expand based on acceptance metrics.

Who maintains it after launch?

The same team handles planning, development, and operations. Anomaly notifications go directly to the responsible engineer, and monthly review reports go to your contact.

Next Step

Start with one role

Bring the 20 questions you receive most often to a 30-minute diagnostic. We will tell you on the spot which suit a digital employee, which should stay with a person, and which metrics the pilot should measure.

Book a 30-minute diagnostic → View plans and pricing

Contact Sainso Technology: info@sainso-tech.com | 04-2391-4106 | Online version and benefit calculator: sainso-tech.com/en/solutions/ai-digital-employee/

Sources and notes

  • Figures in this white paper come from Sainso Technology system records, including version-control release records, scheduled-job reports, acceptance tests, recovery drills, and deployment records. Measurement dates appear with each item.
  • Client cases are anonymized under confidentiality agreements. The service-business platform, website content editor, and operations duty officer are Sainso-owned systems.
  • The benefit calculator is an estimation model and its defaults are examples. Actual benefits are determined by pilot measurements.
  • Data through September 30 in 2026 (month 9).
Contents ↑