Your Software Is Thinking. That’ll Be Extra.
TL;DR
Following an explosion in AI token consumption, the software subscription is getting a meter, but a bigger bill does not guarantee a better business for SaaS vendors. As AI drives up delivery costs, tech firms are charging for tokens consumed, work performed and results delivered alongside access to their products. But buyers still want bills they can budget for. Investors, meanwhile, need to know whether the extra revenue survives the token bill or merely replaces lost seat fees. Where margins fall, revenue multiples mislead. Failed attempts still cost money, and renewals will reveal whether customers think the work is worth paying for. Our Field Notes ask software companies, customers, investors, advisers, and boards what to charge, what buyers will stomach and who pockets the upside.
SaaS With a Meter
Atlassian announced on 1 September 2026 that three things bundled into its subscriptions will come with a meter from December: Rovo credits for AI, automation steps for workflows, and requests an agent resolves without a human touching them. Each still includes an allowance; the meter starts beyond it. Its developer product is already US$20 per seat per month with 2,000 credits included and $0.01 for each one thereafter.
In February, the same company was being dumped as a casualty of the “SaaSpocalypse”, on the theory that agents would clean out the seats its business model depended on.
The seat didn’t die. Instead, the work the software does got a price tag of its own. That’s the real story of software pricing in 2026. The question for investors is whether the new line on the invoice survives delivery costs, seat cannibalisation, and the first renewal as gross profit.
From Seats to Receipts
In August, Bain reviewed public pricing at about 200 B2B software companies. Among those adding AI charges, 80% chose committed capacity over pure pay-as-you-go. About 10% of new meters charged for outcomes, 35% for effort and 55% for output. A separate survey of 230 companies found 37% combining subscriptions with variable AI charges, up from 25% a year earlier.
Microsoft’s Satya Nadella says every per-user business will become per-user plus usage, with each seat entitling buyers to some consumption. Forecasting an AWS bill offers a familiar parallel. GitHub adopted the model in June, keeping Copilot plan prices but converting plans into monthly, token-based AI credit allowances. A $19 Business seat contributes 1,900 credits to a shared pool; additional credits cost a cent each. Heavy agentic users loudly protested bills far above the old flat fee.
Salesforce’s trial and error produced four Agentforce pricing models in two years: $2 per conversation; Flex Credits at $0.10 per action; a flat, multi-year licence for unlimited agent use; then $2 per resolved case. Buyers can now choose among them. Agentforce annual recurring revenue passed $1.5bn in fiscal Q2, up 240%, although its definition widened along the way, a caveat for any AI ARR claim.
The challenge is charging for usage while keeping buyers’ costs predictable.
Credits Are a Currency, Not a Business Model
Credits tend to get bad press, mainly from people who confuse the currency with the economics behind it. A useful rule comes from a16z: price at the highest layer of value you can reliably measure, attribute, and defend. Tokens if you sell access to models, outcomes only where you can see the result, and credits mapped to recognisable units of work for the large middle ground between the two.
In a well-designed credit system, the price of the pool reflects the value of the work to the customer, and the number of credits a task burns is set with delivery costs in mind. Separate the two and, provided the charge per task holds, the vendor keeps some of the benefit when model costs go down, because the customer is paying for work, not reimbursing compute.
In March, the sales-data platform Clay admitted that it had mispriced its Pro tier for years and run it at a loss, then unbundled Data Credits from Actions and began passing through the cost of volatile reasoning models at zero markup. Sydney’s agent platform Relevance AI had already built the same architecture and its documentation is explicit: a tool run, even if it fails, is an Action. An Action is not an outcome, no matter how many times you call the product an AI workforce.
That’s the whole point for anyone doing diligence. When model costs are passed on, a customer spending twice as much may generate no more gross profit. You must break out platform revenue from pass-through revenue before you can put a multiple on anything.
Where Outcome Pricing Works
Customer support is a leading use case for outcome-based pricing. A resolved ticket is discrete, binary, and high-volume; the customer already knows how much a human costs to handle it; and the vendor has enough history to bet margin on its own success rate. Bret Taylor of AI support company Sierra puts a human-handled call at $10 to $20, most of it labour. Against that, a dollar is an easy sell.
Prices are already falling. Zendesk charges approximately $2 per resolution on demand, and $1.50 on committed volume. Salesforce’s Fin charges $0.99 per outcome. In April, HubSpot cut its customer agent pricing from $1 per conversation handled to $0.50 per conversation resolved. Sydney’s Lorikeet, an AI support agent for fintechs, prices at about $1 and gives the customer a veto on what counts. Headline prices for a resolved ticket now run from $0.50 to $2. Salesforce now sells resolutions at two prices: $2 a resolution through Agentforce, and $0.99 through Fin, for which it paid about $3.6bn. Price won’t be the differentiator for much longer. Verification will, because “resolved” is a contractual term.
Fin bills a resolution when the customer gets an answer and walks away without asking for more; if the customer comes back to the same conversation, the charge is reversed. Zendesk runs a separate evaluation model over every conversation, and any case that a human had to close is excluded. Both are serious efforts at the same problem. Neither is proof. A shopper who gives up and goes elsewhere is indistinguishable from a satisfied shopper in the logs. Silence is not satisfaction, and an LLM grading an LLM is a measurement, not an audit.
Meanwhile, buyers still go for the boring option. The AI support vendor Decagon has both meters, and by its own account the vast majority of its customers have opted for per-conversation pricing, because a cost-centre manager is judged on predictable spend before automation rates. Gartner’s Tom Coshow calls the boom in outcome-based pricing “more buzz than reality”. Only 19% of services buyers are using outcome-based arrangements, and Gartner sees fewer than a quarter of technology services contracts doing so through 2031.
A final note: financial outcomes are easier to price when a third party does the counting. Where the vendor grades its own work, the price is a negotiation over attribution.
The Margin Is the Message
Another caution: failed attempts still incur costs, even when they cannot be billed.
Bessemer Venture Partners’ Talia Goldberg says “COGS is the new CAC”. The constraint on a SaaS business was the cost of winning a customer. The constraint on an AI business is the cost of serving one.
Say you charge $1 per accepted outcome and pay $0.20 per attempt. If you succeed 80% of the time, you have a 75% contribution margin. The margin is 60% on 50% success. At 25% success it’s 20%, before anyone has paid for hosting, human review, or the exceptions queue. The hard cases eat a disproportionate share of compute and people.
This is supported by the industry figures. The growth investor ICONIQ puts AI-product gross margins at 45% in 2025, rising to a projected 53% in 2026, and 59% in 2027, with pure application-layer companies at the lower end of the range. The maths is brutal for a traditional SaaS company. Take a third of your revenue and move it from 80% gross margin to 50% and the blended number drops ten points.
Revenue multiples have quietly stopped being comparable. Take two businesses each with $16m gross profit, one with $20m revenue at 80% and one with $32m revenue at 50%. At an enterprise value of $160m they trade on 8x and 5x revenue, and on 10x gross profit each. The buyer of the latter is paying 5x for a compute bill and a software company.
The pricing model also affects revenue recognition. Snowflake’s customers commit to capacity, but its filings recognise product revenue when it is consumed, not rateably over the term. Prepaid usage credits generally create deferred revenue, recognised as services are delivered, with qualifying unused balances recognised under applicable breakage rules. Cash collected upfront is not automatically revenue earned.
Nearer to Home
In ANZ, several of the ASX’s biggest software names were never priced per seat. They sold off in February as if they were.
Pro Medicus prices its radiology contracts per scan, with minimums. When it renewed Chicago’s Northwestern Medicine in April, it raised both minimums and scan fees. Despite losing two-thirds of its stock value by late February, the company reported FY26 revenue growth of 22.9% and booked six five-year renewals worth A$141m, maintaining its historical 100% renewal rate.
TechnologyOne includes implementation in fixed pricing, with 114% net revenue retention and 0.6% churn. Its shares turned positive for the year by April while the sector struggled. Meanwhile, Xero bundled its JAX agent into existing plans, raised prices and added payments. Revenue per customer rose 23%; gross margin fell 5.1 points, largely from payments mix rather than inference. Canva capped AI credits in seat plans, charged for heavy use and cut its growth forecast in August as AI costs bit.
Cannibalisation depends on the workflow, not the sector.
The AI natives are even more instructive, because between them Sydney and Melbourne have produced all three archetypes. Lorikeet is deliberately outcome-priced, in regulated financial services where the audit trail is the product. Relevance AI uses dual-track credits. The AI medical scribe Heidi charges per clinician, has increased its price and reportedly closed a US$340m funding package in September. The automation of a doctor’s notes does not replace the doctor. Where AI is augmenting a licensed human, rather than replacing one, the seat continues to exist, and the vendor can charge more for it.
Further north, a customer of the Malaysian messaging platform respond.io pays a subscription, then per user, then per contact, then for AI credits and then Meta’s WhatsApp fees on top. And in Manila, the contact centres AI was expected to hollow out are being pushed to charge for results rather than hours, as shorter AI-assisted calls reduce hourly revenue.
The Unpleasant Bit
There are two reasons why this analysis might seem premature.
First, the renewal cliff. Bessemer’s take is that while 2025 was all about adoption at all costs with almost no price sensitivity, we will get a clearer picture once the first real renewals fall due in late 2026 and early 2027. Second, the customer’s wallet. In one survey, 70% of software vendors said their customers were still paying for AI out of technology budgets, not payroll. The pitch that AI opens the whole labour line to software vendors is ahead of the evidence.
What We’d Look For
We would apply six tests.
Follow the meter, not the label. What burns a credit, what happens to unused balances, and does a failed attempt burn one? A $120,000 minimum credited against usage is not a $120,000 platform fee plus usage. The difference is the whole forecast.
Demand the bridge. Contract, attempted work, accepted outcomes, invoice, delivery cost, cash. A vendor that cannot reproduce its own invoices from event records is not ready for outcome pricing, whatever the term sheet says.
Gross profit by pricing cohort. Separate platform revenue from pass-through, and margins by workload and delivery model. Rising customer spend can simply be rising third-party consumption.
Audit the auditor. Reopened tickets, credit-back rates, dispute rates, and who grades the outcome. If the grader is the vendor’s own model, price that in.
Net the cannibalisation, workflow by workflow. If an account goes from $10,000 in seat revenue per month to $7,000 and adds $5,000 of AI revenue, it’s up $2,000, not $5,000.
Underwrite the second year. Adoption shows interest. Renewal proves the economics.
We previously wrote that SaaS was getting a brain. It did. The bill for it is now coming due, and it is being split three ways: per outcome where the outcome can be counted, per unit of work where it can’t, and per seat for as long as a human still needs a login. The vendors that hold their pricing will be the ones that can demonstrate the work got done. The rest will spend 2027 debating attribution.
Software isn’t eating the world anymore. It’s sending it an itemised bill.
FIELD NOTES
The same pricing change looks different to the person selling it, the person paying for it, and the person putting a value on the business.
SOFTWARE COMPANIES
RMS Cloud
RMS Cloud has moved from user licences to per-key pricing, with tiered packages designed to open up more of the product. CFO Chetan Arjun says customers retain their choice of integrations while gaining features that previously cost extra: “Overall, it’s simpler, more transparent pricing that customers can budget for and rely upon.”
AI may also become part of those packages. With clients piloting features including guest insights and an AI concierge, Arjun says: “Some of our packages will look to have AI inclusions as part of the standard or premium module range when we start rolling this out.”
SiteMinder (ASX:SDR)
SiteMinder CEO and Managing Director Sankar Narayan sees AI as a growth investment, not simply a cost centre: “We see AI as an unlock of growth and genuine value-add to our customers, rather than a new technology to be used for usage sake.”
Its Dynamic Commerce Engine executes AI-driven hotel pricing and distribution recommendations within controls set by users. With revenue from subscriptions and transaction-based fees, SiteMinder already has a way to share in that growth: “And, as hoteliers do more business through our platform, our model allows us to grow with them. For this reason, we have never priced per seat.”
Adveritas (ASX:AV1)
Adveritas’s TrafficGuard spares advertising budgets from invalid traffic and puts it back to work. It links enterprise pricing to advertising spend and usage, not user seats. The value comes from getting more genuine customers out of the same marketing budget. CEO and Co-Founder Mat Ratty connected customer benefits to pricing in 2024: “As the significant benefits to our clients have become increasingly evident, TrafficGuard’s pricing has been adjusted accordingly”.
That gives it a different exposure to AI from a vendor dependent on selling more licences. “We see AI as an accelerant to our business, not a threat,” Ratty told investors in February 2026. Its accounts describe contracts that can combine a fixed fee, a usage allowance, and overage charges. The point for the pricing debate is that a vendor can demonstrate value in dollars without making every invoice contingent on an agreed saving.
Salesforce (NYSE:CRM)
Salesforce’s changing price list reflects a simple problem: buyers do not all want to pay for AI the same way. At Dreamforce, Chair, CEO, and Co-Founder Marc Benioff put the case for keeping seats plainly: “Some still want pure user pricing. It’s predictable. They understand it.” His answer: “So we’re not fixed to any one particular pricing model. We let that customer have whatever they need.”
Others want to pay for completed tasks or business results. Accommodating both has meant more than adding a meter: “We’ve had to throw away our price point and come up with a whole new pricing mechanism that delivers the right pricing for the right customer at the right time.”
CUSTOMERS / SOFTWARE BUYERS
Fisher & Paykel Appliances
For Rudi Khoury, Chief Digital Officer at Fisher & Paykel Appliances, the attraction of consumption pricing is being able to “manage our overheads in line with our revenue.” His business moves with the economic cycle; a variable software bill need not be a problem.
The proviso is control: “as long as you get the economics right at the start, and you’ve got the right guardrails around the agent”. He is less comfortable with how outcome-based fees might divide the gains: “I think technology providers will want to take a disproportionate return”. Being willing to pay for consumption is not the same as agreeing how to share the upside.
T2
The issue of pricing - and budgeting - is a live one for specialty retailer and lifestyle brand T2, according to the firm’s marketing director Peter Randeria.
“What I’d say is, I don’t think our marketers are clear on the unit economics of where we’re heading from an AI perspective: The cost per piece of content, the cost per image, the actual physical cost that things will start to translate into. And they stack up.”
It’s an issue that needs ongoing navigation, he said. “We’ve moved the AI cost centre out of the IT line and the P&L into its own standalone P&L line to give it transparency.”
He said the approach has helped managers focus on how the firm is learning and managing the cost of AI.
INVESTORS
Potentia Capital
Potentia Capital Managing Director Andrew Gray puts the value proposition ahead of the pricing mechanism. His challenge to portfolio companies is to explain clearly what their AI solutions deliver for customers: “Get that right, then charging on outcomes should not be an issue. Get it wrong, then customers may well revolt and look elsewhere for solutions.”
Advent Partners
Symon Vegter, Partner at Advent Partners, expects much of today’s AI functionality, including search, summaries, and drafting assistance, to be absorbed into the core subscription: “Those features improve the product but rarely justify a separate charge.”
His test for an additional fee is financial impact: “Separate pricing will be reserved for AI that has a material impact on a customer’s revenue or costs. People don’t pay for productivity. They pay for value.”
Five V Capital
Five V sees outcome pricing taking hold first where buyers already pay per result, particularly in Customer Experience. But charging per action brings a tension: it must be balanced against “the incentive for the customer to use the software as much as possible so as to become ingrained, while also avoiding ‘bill shock’ at the end of the quarter.”
The investment case still rests on how useful, and how embedded, the software is: “Software that is system of record, is delivering daily value, and is deeply interwoven into business processes will be able to extract fair economic value, but that mechanism is evolving.”
ADVISERS
GP Bullhound
GP Bullhound Partner Mike Lippert sees AI “weakening the link between user headcount and the value a product delivers.” As agents take on work once done by employees, he says, “vendors are moving away from pure per-seat licensing towards hybrid models that combine a fixed platform fee with variable charges linked to consumption or outcomes.”
Constellation Research
Liz Miller, VP and Principal Analyst at Constellation Research, separates the work software performs from the business result it produces: “So there's a business outcome and there's an operational outcome. Those are not the same.” That distinction matters when agreeing what to pay for. Completing a task is not the same as proving how much money it made or saved.
BOARDS
Roger Sharp
Roger Sharp, founder at North Ridge Partners and Chair of WebBeds, Iress, and Nitro, learnt the limits of sharing the upside 20 years ago. His business charged a low seat fee and a high transaction fee, carrying most of the early risk and losing money for years. “Just as we began recovering our investment, the customer demanded a 50% price cut. Sharing the risk was popular; sharing the rewards less so.”
His prescription: “So I favour a hybrid: dependable recurring revenue, plus upside from usage or outcomes. The recurring base needn’t be per seat. And agree upfront how prices step down as volumes grow, so success doesn’t trigger a renegotiation.”
The cost of delivery needs equal attention: “Every outcome delivered by AI carries a token bill, and that bill shrinks your gross margin. Be careful what you offer, especially at a fixed price.” His advice before putting established revenue at risk: “Prove the economics on new customers before touching your installed base.”
For enquiries, please contact:
Gerry Gimenez, Director of B2B Software & AI at North Ridge Partners, gg@northridgepartners.com