How Much Does AI Really Cost a 50-Person Company?

The prices below come from the DeepSeek API pricing page (api-docs.deepseek.com, fetched August 2026, RMB per million tokens):

(The pricing page also lists cache-hit input at ¥0.02 per million tokens for chat — repeated identical content is nearly free — and off-peak pricing. Always check the live page for the current numbers.)

Assumptions: 50-person company, 200 Q&A interactions per day (employees/customers), 1,000 tokens input + 300 tokens output per conversation, 22 working days per month:

cost per call = 1000 × ¥1/M + 300 × ¥2/M
             = 0.001 + 0.0006 = ¥0.0016
per day       = 200 × 0.0016 = ¥0.32
per month     ≈ 0.32 × 22 ≈ ¥7

Even at 4× the volume (800 calls/day with 3,000-token context), it stays under ¥30/month. Note this is API call fees only — integration time is not included.

Assumptions: 30 documents per day, 5,000 tokens input (~7,000 characters) + 800 tokens output each, using the reasoner model for quality:

cost per call = 5000 × ¥3/M + 800 × ¥6/M
             = 0.015 + 0.0048 = ¥0.0198
per day       = 30 × 0.0198 ≈ ¥0.6
per month     ≈ 0.6 × 22 ≈ ¥13 (3× volume ≈ ¥40/month)

Assumptions: 20 pieces per day, 800 tokens input + 1,500 tokens output each (output-heavy, chat model):

cost per call = 800 × ¥1/M + 1500 × ¥2/M
             = 0.0008 + 0.003 = ¥0.0038
per day       = 20 × 0.0038 ≈ ¥0.08
per month     ≈ 0.08 × 22 ≈ ¥2 (5× volume ≈ ¥10/month)

Taking the "scaled up" upper bound of each scenario: support 30 + documents 60 + content 50 = about ¥140/month, roughly ¥1,700/year. Even doubling the volume, using only the reasoner model with full history, it is hard to exceed ¥5,000/year. Against any staff salary, that is an order of magnitude less.

⚠️ But this does not mean "AI equals 0.01 of an employee." It replaces repetitive workload, not a whole person. The honest framing: AI can take over 60–80% of the repetitive Q&A / sorting / first-draft work in a role (an industry rule of thumb, not a guarantee), leaving humans with the work that needs judgment.

API fees are cheap, but wiring AI into daily workflows takes time:

1. Selection and integration: connecting IM (DingTalk/WeCom/Feishu), connecting APIs, getting the whole pipeline working — hours to a couple of days for someone experienced, 1–2 weeks following tutorials without prior experience
2. Knowledge-base work: turning company documents into a format AI can retrieve — the most time-consuming part, but built once and reused
3. Maintenance: model upgrades, API changes, spot-checking answer quality — a few hours a week

This is the fork in the road between self-hosting and a platform: self-hosting saves subscription fees but you carry all three tasks; a platform (like Mule Agent, an IM+AI middle layer) saves integration time, keeps API costs transparent, and typically runs at RMB 100–500/month.

Worth doing:

• High-frequency repetitive Q&A (support, HR policy questions, IT helpdesk) — high volume but cheap per call, clearest ROI
• Document drafts / summaries / translation — model quality is good enough, humans just review
• Adding an AI entry point to an existing process (e.g. @ a bot in a group to check inventory)

Hold off:

• Scenarios involving customer privacy / regulated data — confirm compliance boundaries first (model provider terms, data masking)
• High-cost-of-error decisions (medical advice, loan approvals) — AI assists, humans approve
• Companies with no documentation at all — build the knowledge base first, otherwise the AI has nothing to work with