Should AI Features Be Free? Both Sides of the Bundling Split Have Casualties

Microsoft folded Copilot into consumer Microsoft 365 and the annual price rose 45% in Australia. One regulator has sued, another has opened an investigation. Salesforce took the other road, metered its agents, and rewrote the meter three times in 18 months. AI bundling is not really a pricing question. It is a question about which mistake you can afford to undo.

The short answer, and why it is not about price

Should AI features be free? Bundle them when the AI defends a seat you already sell and its cost per user is small and predictable. Meter them when the work is countable, the usage is uneven, and one heavy account can consume a year of margin. The expensive mistake is bundling first and metering later.

That sequence is the whole argument. Both strategies work. Only one of them can be reversed cheaply.

Raising the price of a bundle is a negotiation. Buyers grumble, procurement pushes back, and most of them renew. Taking back something you gave away is not a negotiation. It reads as a broken promise, and buyers price it accordingly at the next renewal.

I would bundle less than most vendors currently do. Not because free is wrong, but because almost nobody writes down the exit before they take the entrance.

Bundled, metered, and the third state most vendors are actually in

Bundled means the AI is included in the plan price. The cost is recovered through list price, through a higher tier, or not at all. Metered means the AI is a separate unit on the invoice, counted in conversations, actions, credits or tokens.

Most vendors are in neither state. They are in a third one: bundled with a quota. The feature is advertised as included, an allowance sits behind it, and the allowance is the meter with the dial hidden.

That third state is where the trouble concentrates. A buyer who is told a feature is included and then hits a limit they never saw has been sold two different products. The vendors currently in front of regulators or rewriting their price lists are, almost without exception, the ones who left the quota undocumented.

The case for free: adoption is the scarce input, not revenue

The argument for bundling is simple and it is mostly correct. An AI feature nobody uses defends nothing. Price is the largest barrier to trying one, and a per seat surcharge turns every rollout into a budget conversation with someone who has never used the product.

Zoom set the reference price at zero

Zoom shipped AI Companion in September 2023, included at no additional cost for customers with paid services on their accounts. Chief product officer Smita Hashim described it as "disrupting the industry's pricing model", which was accurate in one specific sense. It set the reference price for meeting summarisation at zero.

That did more damage to standalone note takers than any feature would have. A point tool charging a monthly fee now has to justify itself against something the buyer already owns. The wider version of this pattern is covered in the piece on point tools being absorbed into agent ecosystems.

Atlassian removed the price tag to win the seat

Atlassian did the same thing 19 months later. From 9 April 2025 Rovo became available at no additional upfront cost to organisations with active cloud subscriptions, bundled into Premium and Enterprise with a limited version on Standard.

The logic is defensive. If the AI layer that sits on your data is the thing a competitor uses to displace you, giving it away is cheaper than losing the account. That is a real argument, and it is the same one behind vertical AI moving on horizontal software.

What free actually buys is three things: usage data you cannot get any other way, a habit that raises switching costs, and the removal of a line item a procurement team can cut. None of those show up as revenue. All of them show up in retention, eventually, if the feature is good.

What bundling cost Microsoft, and it was not margin

Microsoft ran the largest bundling experiment in consumer software and the bill arrived from an unexpected department. Not finance. Legal.

Australia: 2.7 million subscribers and a Federal Court filing

Copilot went into Australian consumer Microsoft 365 plans from 31 October 2024. Microsoft emailed subscribers about the change on 9 January 2025 and again on 13 April 2025. At renewal, the Personal plan went from A$109 to A$159 a year and Family from A$139 to A$179.

On 27 October 2025 the ACCC filed proceedings in the Federal Court. The allegation is not that the price rose. It is that Microsoft told about 2.7 million subscribers they could accept the increase or cancel. A cheaper Classic plan existed at the old price without Copilot. The ACCC alleges its existence was not disclosed until a customer began the cancellation flow. The regulator is seeking penalties, injunctions, declarations, consumer redress and costs.

The UK opened its own case on 29 July 2026

The Competition and Markets Authority opened an investigation into the same pattern nine months later. In the UK the Classic Personal plan runs at £59.99 a year against £84.99 for the Copilot version, and Classic Family at £79.99 against £104.99.

One detail is worth correcting, because most coverage got it wrong. This is a consumer protection case under unfair commercial practices rules, not a competition case. Hayley Fletcher, the CMA's senior director for consumer protection, framed it precisely: "Our investigation will consider whether Microsoft customers were misled and ended up paying more as a result." The CMA has reached no conclusions, and neither should anyone reading this.

Then Microsoft did it again on the commercial side. Effective 1 July 2026, and announced the previous December, base Copilot Chat capabilities folded into most suites while list prices rose across the portfolio.

The increases run from 5% on Microsoft 365 E5 to 33% on F1 across the suites Microsoft published. Some trade coverage cites a top figure of 43% on particular frontline configurations. That number is not on the list Microsoft published, so treat it as unconfirmed until the SKU is named.

The pattern underneath both cases is the same. Bundled AI is not free. It is charged to everyone, including the majority who will never open it, which is a redistribution rather than a discount.

The case for metering: the marginal cost is real this time

Software pricing was built on an assumption that stopped being true in 2023. One more user of a database costs almost nothing. One more user of a model costs tokens, every time, forever.

A 50% margin business cannot be priced like an 80% margin business

Kyle Poyar's survey of 230 software and AI companies, fielded in April and May 2026, puts the median target gross margin on AI at about 50%. Only 12% are aiming at 80% or higher, which is the range software has historically earned and the range every seat price in the industry was set against.

That gap is the entire commercial argument for a meter. It is also why 29% of those companies already use AI credits and 33% said they intend to introduce them within 6 to 12 months. Among companies above $150 million in annual recurring revenue, roughly one in two said they plan to add credits this year. The cost mechanics behind that decision sit in the analysis of inference costs and AI margins.

Metering also does something bundling cannot. It gives the buyer a number to compare against a wage. When Salesforce prices a standard Agentforce action at 20 Flex Credits, or $0.10 against a $500 block of 100,000, a buyer can work out what the agent costs per resolved case. That comparison is the strongest sales argument in enterprise AI, and a bundle destroys it by hiding the unit.

What metering cost the vendors who chose it

The bill for metering does not arrive from a regulator. It arrives as rework, and rework at pricing scale is expensive in a way that does not appear in any budget line.

Salesforce now sells the same agent four ways

Agentforce launched at $2 per conversation. In May 2025 Salesforce added Flex Credits at $500 per 100,000, with a standard action costing 20 credits and a voice action 30. Then came per user licensing. As of August 2026 the published pricing page carries all of it at once.

Prices are Salesforce's own published list, read in August 2026. Industry specific add-ons run at $150 per user per month. Enterprise discounting means few customers pay list.

Read the fourth row again. Having spent 18 months arguing that agents should be metered because they are digital labour, Salesforce now sells an unmetered per seat version of the same thing. That is not incoherence. It is a company that has run the experiment and found that buyers will not standardise on one model, so it sells all of them and lets the customer carry the choice. Whether that survives contact with a full renewal cycle is the open question in the wider read on Salesforce in the agent era.

Atlassian bundled first and had to build the meter afterwards

Atlassian took the other route and ended up in the same place by a longer path. Rovo went out with no additional upfront cost in April 2025. By 2026 the published usage allowance reads like a meter, because it is one.

Standard gets 25 Rovo credits per user per month, Premium 70 and Enterprise 150 on the Jira, Confluence and service collections. A Rovo agent request costs 10 credits. Deep Research costs 100. Search, summaries and definitions cost nothing. Credits pool across the organisation and do not roll over.

Now look at the guarantees around it. Atlassian states it is not currently billing for usage above the allowance. It commits to at least 90 days notice before that changes. It also says no customer will be billed for overage without an explicit opt in. Credits became visible in the administration dashboard from August 2026.

Those three commitments are the price of going second. A vendor that metered on day one never has to promise a notice period, because there is nothing to take back. Atlassian is paying for the goodwill it bought in 2025, in instalments, and the instalments are contractual.

None of this is unusual. Three in four software companies in the Poyar survey changed pricing or packaging within the previous year. What is unusual is doing it on a feature customers were told was included.

How to choose, in two questions

Skip the frameworks. Two questions settle most cases, and a third one settles the rest.

Does the AI defend revenue you already collect, or create revenue you do not? Defensive AI belongs in the bundle. It exists to stop a competitor using the AI layer as a wedge, and a surcharge that suppresses adoption defeats the purpose. Additive AI, the kind that does work a customer would otherwise pay a person to do, belongs on a meter. The buyer is comparing it to a wage, so give them a unit.

Is usage flat or concentrated? If your heaviest 5% of accounts use roughly what the median account uses, bundle it and forget it. If the top decile uses 40 times the median, a bundle transfers money from light users to heavy ones and eventually breaks. The tell is not average usage. It is the ratio between the 90th percentile and the median, and most vendors have never calculated it.

The third question is the one nobody asks. If you have to reprice this in 18 months, what does the email say? Write that email now. If you cannot write it without sounding like the renewal notices two regulators are currently examining, you have your answer. It is to publish the limit on day one instead.

Publishing the quota costs you almost nothing and buys the right to enforce it later. Microsoft's difficulty is not that Copilot cost money. It is that the cheaper option existed and was not in the email. Buyers forgive a limit. They do not forgive discovering one.

Where this argument is weakest

Three places, and the first is the load bearing one.

The regulator cases are about disclosure, not bundling

Neither the ACCC nor the CMA has said that bundling AI into a subscription is unlawful. Both are examining how the change was communicated to customers who had already bought something else. If Microsoft had named the Classic plan in the first line of the renewal email, there is a reasonable chance neither case exists.

So the evidence I have leaned on constrains how you bundle, not whether. Anyone citing these cases as proof that bundled AI is a bad strategy is over-reading them, and the allegations remain untested in court.

Free may be working, and the numbers cannot show it

Zoom has given AI away for nearly three years. In fiscal 2026 it reported total revenue of $4,868.8 million, up 4.4%, with enterprise revenue up 6.5%. That is not the inflection a free AI layer was supposed to produce.

It is also not evidence against the strategy. Defensive spending shows up as an absence, and there is no counterfactual Zoom to measure. Bundled AI may be holding accounts that would otherwise have gone, and a retention effect never appears in a growth rate. I would not use that 4.4% to argue anything, and neither should you.

The wider difficulty is the one running through every attempt to measure AI return. Almost nobody recorded the baseline before deploying, which makes the after picture uninterpretable in either direction.

Frequently asked questions

Where to start this week

If you sell software, run one query. Pull AI feature usage per account for the last 90 days and calculate the ratio between the 90th percentile and the median. Under 5 and you can bundle safely. Over 20 and a bundle is already transferring money from your quiet customers to your loudest ones, whether or not anyone has noticed.

If you buy software, open your three largest renewals and find the AI clause. Two things to establish: what the allowance is in units you can count, and what happens when you exceed it. If the contract says the vendor may introduce charges at its discretion, ask for a notice period in writing. Atlassian has publicly committed to 90 days, so that is now a reasonable thing to ask for. It is also the fastest way to find out whether your vendor has decided anything at all. The same discipline applies across the rest of the stack, as the work on rationalising app sprawl sets out.

Related reading