AI for SMBs: New Customer Service Standard & DeepSeek's Open-Source Multimodal Model

AI Customer Service National Standard (GB/T 47746‑2026)

Effective 2022‑09‑01, the 16‑page standard defines enforceable clauses: a visible hand‑off button, automatic escalation to a human agent at defined thresholds, and no requirement for the user to repeat the issue when switching. The China Consumers Association reports that 26.79 % of after‑sale complaints involve AI customer service, with transfer to a human agent being the hardest issue.

Self‑test steps: open the own chatbot, type “转人工”, and observe the behavior within 30 seconds; use the 35‑item checklist (clause numbers) to query the AI‑service provider; replace the KPI “interception rate” with resolution rate, escalation rate, and the proportion of cases neither resolved nor transferred.

Implementation difficulty ★★☆☆☆ (configuration changes and KPI adjustment, no model change). Estimated cost ¥0‑¥1000 / month.

Tencent Hy4 Preview Model

Hy4 preview is released with total 770 B parameters (49 B active) and a 1 M token context window. It targets software engineering, office work, game development, and research. Internal blind tests score higher than GLM‑5.3 and Kimi K3. Ant Financial announced the finance‑enhanced model Ling‑3.0‑Flash‑Fin.

Implications: the 1 M context enables ingestion of whole codebases or contracts for knowledge‑base use, content generation, and compliant customer service (combined with the new standard). High‑complexity agents (49 B active) require careful evaluation.

Implementation difficulty ★★★★☆ (requires sufficient compute or cloud API). Estimated cost ¥0‑¥1000 / month via cloud API; private deployment costs more.

DeepSeek V4 Multimodal Model (Open‑Source)

DeepSeek released the experimental multimodal model DeepSeek‑V4‑Flash‑Vision‑Exp on Hugging Face under the MIT License. It adds a vision module to the V4‑Flash backbone, enabling the model to read screenshots, charts, and UI elements while retaining text reasoning and agent capabilities. Weights (168 GB, 48 shards) and inference code are publicly available; an API was launched on 2022‑08‑21.

Benchmarks: Terminal Bench 2.1 score improved from 82.7 to 83.9; DeepSWE from 54.4 to 59.3, surpassing Opus‑4.8 on several metrics.

Use cases: developers can deploy locally or privately to build vision‑enabled agents for order filing, screenshot recognition, and chart analysis; teams without developers can use the API to add vision capability to existing agent toolchains. Local deployment requires multi‑GPU due to the 168 GB weight.

Implementation difficulty ★★★★☆ (multi‑GPU for local deployment; API option easier). Estimated cost ¥0‑¥1000 / month via API; private deployment incurs additional expense.

AI Coding Assistants

Cursor launched Origin, integrating repository, PR, CI, and agent functionality into the editor while keeping GitHub as the source of truth. OpenAI added a “persistent mode” to Codex, allowing continuous operation across sessions for up to 25 hours.

Workflow change: employee → give AI a task → AI edits code, creates PR, runs CI → employee reviews, replacing the traditional flow of IDE → write code → PR → CI → merge.

Suitable for software/SaaS companies with in‑house dev teams and enterprises with heavy internal tooling. Persistent mode consumes tokens continuously; usage limits should be set (refer to clause 6 of the new standard).

Implementation difficulty ★★☆☆☆. Estimated cost ¥100‑¥1000 / month (subscription).

AI Office Suite Competition

Three AI office products—豆包工作, WorkBuddy, and 千问办公—have moved to organization‑level solutions. Public testing shows comparable output for meeting minutes and PPT generation within free quotas; each has distinct failure points.

Measured efficiency gains (human‑verified): weekly report creation reduced from 3 hours to 10 minutes; meeting minutes from 1 hour to 5 minutes; PPT from 5 hours to 1 minute. Numbers require verification because AI may fabricate data.

Implementation difficulty ☆☆☆☆☆ (no technical integration needed). Estimated cost ¥0‑¥198 / month (free tier often sufficient).

AI Business Failures and Safety

The China Consumers Association disclosed two incidents: an AI‑managed second‑hand marketplace sold a 1 000 CNY sound card for 400 CNY; a food‑delivery rider’s AI‑assisted complaint promised a refund that was later retracted after human review.

Takeaways: employee‑facing AI should provide explanations only; final decisions must be made by humans. The disclaimer “AI answer for reference only” is harmful in customer‑facing dialogs and can increase liability. Public AI use on company data creates data‑leak risks.

AI Cost Realities

Meta’s “OT” project attempted to replace thousands of jobs with AI agents. After one year: code changes +220 %; new user‑facing features +36 %; major incidents +40 %; engineer fire‑fighting time +70 %; morale fell from 74 % to 55 %. Forrester reports 55 % of AI‑related layoffs are later regretted; 32 % of those firms re‑hire the same roles with 25‑28 % higher salaries.

Uber exhausted its annual Claude Code budget in four months, forcing a $1,500 per user‑tool monthly cap.

AI Pilot Success Factors

Studies from MIT, S&P, and RAND show less than 5 % of AI pilots deliver ROI; 42 % of firms have cut AI projects; only 17 % of 2024 AI investments reach production. The primary cause of failure is leadership, not technology.

Wrong sequence – buying tools before identifying scenarios.

Scenario mismatch – applying AI to visible problems instead of the most painful ones.

Lack of metrics – no success definition before launch.

Attach a quantifiable metric to every AI project; without it, the project is not considered delivered.

AI Subsidy Programs (8‑City Compute Vouchers)

Beijing, Shanghai, Shenzhen, Suzhou, Wuhan, Chengdu, Jinan, and Xiamen issue AI compute/model vouchers. “Maximum quota” differs from the amount each applicant can actually receive. Qualification criteria vary (e.g., Wuhan defines “AI + super‑individual”, Guangzhou requires a single‑natural‑person shareholder).

Check whether the program is first‑come‑first‑served or qualification‑based.

Key dates: Shenzhen voucher starts 2022‑09‑01 (deadline 2022‑11‑10); Beijing Economic‑Development Zone deadline 2022‑09‑20; Hangzhou deadline 2022‑09‑30 (first‑come). Implementation difficulty ★★★☆☆ (paperwork and application).

Recommendations by Company Size

10‑50 employees: prioritize AI customer‑service self‑test → AI knowledge‑base → AI office tools → monitor subsidy windows.

50‑200 employees: focus on AI service‑metric correction → AI workflow automation → enterprise knowledge‑base → broader AI‑enabled business systems.

200‑1 000 employees: build an enterprise AI platform → agent layer → AI engineering → governance framework.

Key Takeaway

AI’s true value lies in re‑architecting existing business processes—replace deceptive metrics like “interception rate” with genuine outcomes such as “resolution rate”.