What Are AI Credits?
AI credits are a vendor-issued unit of consumption. You buy a bundle each month, and every AI action your staff trigger draws some down. A credit is not a token and not a seat. It is the vendor's own currency, and the vendor sets what each action costs.
Here is the part most owners miss. Traditional software billed you per person. Ten staff, ten licences, one predictable number. Credits break that link entirely.
Two staff in a busy month can now cost more than ten staff in a quiet one. Your software bill has stopped tracking headcount and started tracking activity.
Why Did Software Vendors Switch From Seats to Credits?
Vendors switched because AI agents broke the maths behind per-seat pricing. When one person with an agent does the work of five, charging by the person means the vendor earns less exactly as the product becomes more valuable.
The shift is measurable, not theoretical. Pricing research summarised by Monetizely's 2026 guide to SaaS and agentic pricing found pure per-seat pricing fell from 21% to 15% of SaaS companies in twelve months, while hybrid seat-plus-usage models rose from 27% to 41%.
Two names make it concrete. GitHub announced on 27 April 2026 that Copilot would move to token-based AI Credits. HubSpot has been openly hybrid, extending HubSpot Credits to cover its Breeze Customer Agent and describing it as the first step in monetising AI.
The reason this matters to a Hong Kong owner is simple. When your renewal quote arrives, the line that used to say "10 users" may now say "10 users plus 50,000 credits", and the second number is the one that moves.
How Does a Credit Actually Get Used Up?
A credit is consumed whenever the software calls an AI model on your behalf. The size of the draw depends on how much text goes in, how much comes out, and how many steps the task takes.
Think of it like a taxi meter rather than a bus fare. A bus fare is fixed no matter how far you sit. A meter runs on distance and traffic. AI credits run on length and complexity.
Typical draws across the tools an SME already uses:
--- A short chat reply: a fraction of one credit
--- Summarising a 30-page contract: several credits, because the whole document is read first
--- One AI-generated image: often 10 to 50 credits, priced far above text
--- An agent that researches, drafts and files a report: dozens of credits, because it runs many steps in a loop
The last line is the dangerous one. An agent decides for itself how many steps a task needs. A staff member who asks for "a full competitor review" has authorised spending they cannot see.
If the underlying unit is unfamiliar, our explainer on what AI tokens are and how they are billed covers the layer beneath credits.
How Much Do AI Credits Cost a Hong Kong Small Business?
There is no single rate, because each vendor defines its own credit. What you can do is model your own worst month before you sign, which takes about twenty minutes.
Take a real shape of business. A Sham Shui Po trading company, eight staff, four of them at desks all day handling supplier emails and quotations.
The quiet month
--- 4 staff x 20 AI actions per working day x 22 days = 1,760 actions
--- At an average of 2 credits per action = 3,520 credits
--- Comfortably inside a mid-tier bundle
The month a big tender lands
--- The same four staff run document summaries and drafted responses instead of short replies
--- Average draw rises from 2 credits to roughly 12
--- 1,760 actions x 12 = 21,120 credits, about six times the quiet month
Nothing went wrong here. Nobody misused the tool. The business simply had a good month, and the bill followed the work rather than the payroll.
That is the whole lesson. Budget for your busiest month, not your average one, because the busy month is the one you actually want.
What Do Business Owners Get Wrong About AI Credits?
Four misunderstandings cause almost all of the unpleasant invoices.
"Credits and tokens are the same thing."
They are not. Tokens are the model's unit. Credits are the vendor's wrapper around it, and the exchange rate between them is set by the vendor, not by you. A vendor can quietly change how many credits an action costs while the sticker price stays the same.
"Unused credits roll over."
Usually they expire at the end of the billing period. Buying a large bundle to feel safe often means paying for capacity you burn only twice a year.
"Credits are cheaper than seats."
Only when usage is low. Below roughly a few hundred AI actions a month, a flat per-seat plan is normally cheaper and always more predictable. Credits reward light users and punish heavy ones.
"One action costs one credit."
An agentic task is a loop, not a single call. One instruction can trigger twenty model calls before it returns an answer, and every one of them draws down.
Seats, Credits or Per-Outcome: Which Bill Are You Being Offered?
Three pricing shapes now sit on the same quotation, and vendors often mix two of them. Knowing which one you are looking at tells you where the risk sits before you negotiate anything.
Per-seat
--- You pay a fixed monthly fee for each named user
--- Best for: predictable teams where everyone uses the tool a similar amount
--- The risk: you pay for the four people who logged in twice
Credits or consumption
--- You buy a bundle of units and draw them down by activity
--- Best for: uneven or seasonal work, and for testing before you commit
--- The risk: your best month is also your most expensive month
Per-outcome
--- You pay only when the AI completes a defined result, such as a resolved support ticket
--- Best for: one clearly countable job with an agreed definition of done
--- The risk: arguing over what counts as resolved
Most 2026 contracts an SME will see are hybrids, typically a small seat fee plus a credit bundle. That is not a trap by itself. The trap is signing the seat number carefully and skimming the credit number.
One question settles it. Ask the salesperson: "Show me this quote at three times my expected usage." A vendor who can answer that in a minute is quoting you honestly. A vendor who needs to check is telling you the credit line is where the margin lives.
How Do You Keep a Credit-Based Bill Predictable?
You cannot fix the price, but you can remove the surprise. Five controls do most of the work, and none of them require an IT team.
--- Ask for the conversion table in writing. How many credits does a chat reply, a document summary and an image cost? If the vendor will not put it in the contract, that is your answer.
--- Set a hard cap, not a soft alert. An alert tells you after the money is gone. A cap stops the spend. Confirm which one you are getting.
--- Check the overage rate before the base rate. The bundle price is advertised. The per-credit overage price is where a busy month becomes an expensive one.
--- Name one owner for the account. One person watches the meter weekly. Without a named owner nobody notices until the invoice arrives.
--- Run one month on the smallest bundle first. Real usage data beats any vendor estimate, and you can always size up.
This is not exotic discipline. It is the same thinking you already apply to an electricity bill or a mobile data plan.
Is a Credit Model Right for a Small Business at All?
For many Hong Kong SMEs, yes, and for a reason that has nothing to do with price. Credits let you start small without committing every employee to a licence.
That matters given where local businesses actually are. A Dah Sing Bank survey of more than 340 Hong Kong SMEs, published in July 2026, found 23% had adopted AI and 32% planned to within one to two years. Among those who had not started, 57% cited a lack of relevant knowledge or skills and 38% pointed to unclear return on investment.
Credits speak directly to that second objection. A small bundle lets one department test a real workflow for a month and produce a number, instead of a ten-seat annual contract signed on a hunch.
The judgement call is where you are on the usage curve. Light and exploratory, credits are your friend. Heavy and routine, a fixed plan or a scoped deployment will almost always cost less. If you are weighing a subscription against something purpose-built, our comparison of an AI chatbot subscription versus a deployed AI employee works through that decision.
Frequently Asked Questions About AI Credits
Do AI credits expire?
In most plans, yes, at the end of each billing cycle. A minority of vendors allow limited rollover on annual contracts. Confirm it in writing, because it is rarely on the pricing page.
Can a vendor change how many credits an action costs?
Yes, and this is the structural risk in credit pricing. Your monthly price can stay identical while the work it buys shrinks. Ask whether the conversion rate is contractually fixed for the term.
What happens when we run out mid-month?
Either the feature stops, or it keeps running at an overage rate. Both are survivable, but only if you knew which one you signed up for.
Should I buy the biggest bundle to get the discount?
Not in month one. Buy the smallest bundle, measure a real month, then size up with evidence. The discount on capacity you never use is not a discount.
The Takeaway
AI credits are not a trick. They are software pricing catching up with software that now does work instead of just displaying screens.
The change asks one new thing of you as an owner. You used to budget by counting people. Now you budget by understanding work, which is a harder question but a more honest one.
Get the conversion table, set a hard cap, measure one real month. That is the entire discipline, and it is well within reach of any business without an IT department.
Nobody should have to become a pricing analyst to buy a piece of software. We understand AI. UD stands with you.
Not Sure Where Your Business Should Start?
Before you compare credit bundles, it helps to know which parts of your business are actually ready for AI. UD's free AI Ready Check gives you that picture in three minutes, and if you want to go further, we will walk you through every step from assessment to going live.
Reviewed by the UD AI team, Hong Kong. Published 5 August 2026.
