China has already decided how to price AI agents. Australian software companies should pay attention.

The number is ¥198

Tencent's WorkBuddy Enterprise costs ¥198 per seat per month, with 2,000 credits per seat refreshed monthly and shared across the team (Tencent Cloud). ByteDance's Feishu sells its Doubao Workmate team subscription at the same ¥198 per seat per month (WallStreetCN via Futu News). Alibaba's QwenWork lists its Enterprise Standard plan internationally at US$30 per seat per month, also with 2,000 monthly credits pooled across the organisation (Alibaba Cloud).

That is roughly US$29 to US$30 a seat across all three (Poe Zhao). None of them charges by task completed or by business result. All of them kept the human seat as the thing customers buy, and added credits to control how much AI work each seat consumes.

Analyst Poe Zhao, who has tracked the convergence closely, puts it simply: the companies with public enterprise pricing "still charge by the human seat" (Hello China Tech).

Half the seat isn't paying for AI

The more revealing detail is what happens when you take the price apart.

Tencent sells extra credits in top-up packs. The smallest pack is 2,000 credits for ¥100 (Tencent Cloud). So the 2,000 credits included in a ¥198 seat are worth about ¥100 at Tencent's own top-up price. The other ¥98, roughly half the seat, is paying for something else.

Alibaba's numbers show the same pattern. Its add-on credit packs start at US$20 for 2,000 credits (Alibaba Cloud), so about two-thirds of a US$30 seat buys AI work and the remaining third buys the seat itself.

What's in that remainder is the admin console, seat management, usage controls, security, audit and observability. Tencent has been open that its early enterprise revenue comes from exactly these security, observability and management capabilities (WallStreetCN via Futu News).

Gartner's George Brocklehurst describes the shift well: "The seat is not completely disappearing; it is being unbundled from execution." In his words, seat count "will increasingly measure the human control layer, not the total volume of software execution" (TechTarget).

That is the real lesson from China. The seat now prices control and accountability, and the credits price the work.

The West has landed in the same place

The large Western vendors use different ratios but the same idea.

• Microsoft charges US$30 per user per month for the Microsoft 365 Copilot add-on. Everyday AI is included, and more advanced agent work is billed as it's used through Copilot Credits (Microsoft).
• Salesforce offers Agentforce either as a US$125 per user per month add-on with unmetered use, or as a US$5 per user licence plus Flex Credits at about US$0.10 per action (Salesforce). On those numbers, the flat fee only pays off for a user whose agents take more than about 1,200 actions a month.

Seat for the person, meter for the machine. Every major vendor has arrived at some version of it.

Why credits, and why now

The reason is cost. Gartner's review of nearly 1,000 case studies found fewer than 10% of agentic AI deployments are delivering robust return on investment, while inference costs per agentic workflow are predicted to rise more than five times through 2028. Reasoning models typically consume five to 30 times more tokens than basic chatbots (Gartner).

Buyers are nervous about this too. Gartner predicts that through 2027, organisations running AI agents on uncapped, token-linked pricing will exceed their initial cost expectations by more than 50%, compared with those on predictable subscription and bundled pricing (Gartner).

Credits solve both problems at once. The vendor protects its margin because heavy users pay for heavy use. The buyer gets a predictable bill because the allowance is known in advance. That is why Gartner lists credit-based entitlements among the main alternatives vendors are now exploring (Gartner).

Launch prices aren't final prices

One more detail deserves attention. Tencent introduced a new enterprise billing plan on 15 May 2026 (Tencent Cloud). Poe Zhao reports the per-seat price went from about US$11.50 to US$29 at that point, which fits a familiar pattern of subsidising adoption and then repricing once usage and costs are clearer, though Tencent hasn't given that reason (Poe Zhao).

Outside China, Tencent's WorkBuddy Team plan costs US$40 per seat per month with 1,000 base credits, currently doubled for a limited time (Tencent Cloud). Tencent's own cloud chief says the company hasn't yet set specific revenue targets for WorkBuddy (WallStreetCN via Futu News).

In other words, even the biggest players are still finding the right price. Expect more changes, and expect them to go up as usage grows.

The catch: nobody knows what a credit buys

There's an uncomfortable truth inside this model. A credit isn't a unit of work. It's a currency, and the vendor sets the exchange rate.

Tencent says WorkBuddy's credit use depends on the model chosen, how complex the task is and how long the conversation runs (WorkBuddy). Its own buying guide admits that WorkBuddy doesn't publish per-task credit costs, so "the practical way to assess sufficiency is to run one month, observe the consumption trend, then adjust your plan or top up" (Tencent Cloud). Tencent has even built a check that pauses an agent and asks for confirmation when it detects a task may be stuck in a loop (WorkBuddy).

Microsoft publishes its rates, which shows how wide the range can be. A generative answer costs 2 Copilot Credits, but if the agent uses a reasoning model, Microsoft adds 10 credits per 1,000 tokens on top (Microsoft Learn). A single answer that reasons through 2,000 tokens therefore costs 22 credits, eleven times the base rate, for what looks to the user like the same question.

The risk for buyers is real. Organisations are seeing AI budgets exceed forecasts by two to three times within a single year, and one reportedly used up its annual AI budget by March (TechTarget). The risk for vendors is real too. When Cursor moved its US$20 plan from 500 requests a month to a pool of usage credits in June 2025, many users only understood the change when surprise charges arrived. The CEO apologised and refunded the charges (Darkroom).

Credits don't create this uncertainty. Under flat seat pricing, the vendor quietly absorbed it in its margin. Credits hand it back to the customer, and the customer's Finance team now has to manage it.

Who approves the top-up?

This is the question most businesses haven't answered yet. When the pool runs dry mid-month and someone asks for more credits, on what basis does Finance say yes or no?

The FinOps Foundation, the industry body for cloud cost management, describes what's working: Finance sets the envelope and the tolerance, engineering turns that into enforced limits, and someone owns the evidence. As one version puts it, "the budget is open and flexible, but we do not want to see it double" (FinOps Foundation).

Turned into approval rules, that gives five tests for any top-up request:

1. Is it more work, or more expensive work? If the volume of real work has gone up, the top-up is growth, so approve it. If the cost per task has gone up, something has changed, such as a pricier model, longer conversations or a loop, so investigate first. One practitioner puts it neatly: total spend rising with usage is growth, but unit cost rising is a change someone made (Frugal).
2. Does it have an owner and a purpose? Credits should belong to a team and a workflow, not a general pool nobody answers for.
3. Is it an experiment or business as usual? Experiments get a fixed budget, success criteria and an end date, and standard requests can be approved automatically (FinOps Foundation). Business-as-usual work gets a run rate that can be forecast.
4. Is it cheaper than the alternative? Compare the cost of the top-up with what the same work would cost in people's time. If AI is still cheaper, the request is usually sound.
5. Is it inside the agreed tolerance? Agree a band in advance, say up to 20% over the monthly allowance, approved automatically within it and escalated beyond it.

None of this is complicated, but almost nobody has written it down yet. The businesses that do will spend less time arguing about top-ups and more time deciding where AI is worth it.

What this means for Australian software companies

1. Your buyers will compare you with seat plus credits.

It's now the model enterprise buyers see from Microsoft, Salesforce and the largest Asian platforms. If your AI pricing is purely by outcome, expect to spend part of every sales cycle explaining it. If your AI is free, you're teaching customers that AI costs nothing, which makes every future price rise harder to justify.

2. Price the seat for control and the credits for work.

The part of a seat enterprise buyers will happily pay for is admin, permissions, audit trails, budget caps and visibility of usage. If your AI tier is just "the same product with AI switched on", you're leaving the most defensible part of the price on the table.

3. Know your cost per AI action for your heaviest users.

Pooled credits exist so that light users subsidise heavy ones within a customer, and heavy customers pay for their consumption. If you can't name your cost per AI action for your top 10% of users, you don't yet know whether your AI tier makes money.

4. Leave yourself room to reprice.

Tencent's May change shows what's coming. The companies that will move most easily from launch price to sustainable price are those with clear usage allowances from day one. "Unlimited AI" is a promise that will be expensive to walk back.

5. Keep outcome pricing for the deals that suit it.

Outcome pricing works in large deals where the result can be measured and the buyer trusts the measurement. For most mid-market software, it adds friction without adding margin. Gartner advises buyers to compare products by the cost of completing the work, not the advertised unit price (TechTarget), so be ready to show yours.

6. Make your credits easy to approve.

Your customer's Finance team will be the one approving top-ups, so design for them. Publish what typical tasks cost in credits, show usage by team and workflow, send alerts well before the pool runs out, let admins set caps, and give a forecast of month-end use. The vendor whose credits are easiest to approve will have an easier renewal.

Six questions to ask about your own AI pricing

1. What does one AI action cost us to deliver, on average and for our heaviest 10% of users?
2. Which parts of our product are about control (admin, security, audit, usage limits), and are we charging for them?
3. Do our customers have a known monthly allowance, or an open-ended meter?
4. If inference costs doubled next year, could we change our pricing without breaking contracts?
5. Could a buyer's Finance team compare our price with Microsoft's or Salesforce's in under five minutes?
6. When a customer asks for a top-up, can they see what the credits were spent on and what they'll get for the next batch?

If any of these is hard to answer, that's the place to start.

Coming next

Rev-IQ Advisory will be in Shanghai and Bangkok over the New Year, looking at how technology companies in both markets are turning AI into products people pay for, and will report back in January.

In the meantime, our earlier Insight, Everyone's using AI. Almost nobody's getting paid for it, covers the wider value gap. If you want to test whether your business is ready to charge for AI, the free AI Readiness Check is a good place to start.

This article is general information only, not pricing or financial advice. Vendor prices, credit allowances and promotions change often; check current terms before relying on them. The half-the-seat and 22-credit figures are Rev-IQ calculations from published vendor prices, and the 20% tolerance band is an illustration. Prices checked 5 November 2026.