Why AI Spending Rarely Reaches Earnings

Executive summary

• In McKinsey’s State of AI survey reported in August 2026, 80% of respondents using AI reported individual productivity gains, 37% attributed at least some EBIT impact to AI, and only 6% qualified as high performers, meaning they attributed at least 5% of EBIT to AI and called the impact significant. That 6% has been flat for a year.
• The common explanations, weak governance and poor data, are real but incomplete. They do not explain why productivity keeps climbing while earnings impact stays still.
• The structural answer is that most AI budgets buy efficiency inside an unchanged revenue model, and efficiency that any competitor can purchase from the same vendor gets passed to the buyer as price.
• Every technology investment opens one of three doors: efficiency, capability, or revenue model. Only the third compounds, because copying it requires rebuilding a commercial system rather than signing a vendor contract.
• Three questions separate the three doors before you approve the spend. If a program changes no price, no billing trigger and no payer, it is an efficiency purchase, and it should be budgeted and defended as one.
• Four revenue model moves are now commercially live in B2B: outcome pricing, consumption pricing, data-back monetization, and service-to-product conversion. Each one breaks something specific, and this article names what.

The number that should be moving and is not

Enterprise AI adoption is no longer the bottleneck. Money is being spent, tools are deployed, and people report that the tools help them personally.

What has not changed is the earnings line. In the McKinsey State of AI survey of 1,719 professionals reported on August 25, 2026, 80% of respondents using AI reported individual productivity gains. Only 37% attributed any EBIT impact to AI, roughly the same share as a year earlier, and most of those said AI accounted for less than 5% of EBIT. Just 6% met the bar for high performers, and that figure was flat year over year.

MIT’s GenAI Divide study, published in 2025 and built on 52 executive interviews, 153 leader surveys and an analysis of 300 public deployments, reached a compatible conclusion from a different direction: 95% of generative AI pilots delivered no measurable P&L impact.

Two independent studies, two methods, one result. Individual productivity is rising. Company earnings are not following.

The standard explanations point at execution: governance gaps, dirty data, workflow integration, change management. Those diagnoses are correct as far as they go, and I have argued them myself in why AI investments fail to generate revenue. But they cannot be the whole answer, because if execution were the only obstacle, the best-run companies would already be showing the gain. The high performer share would be climbing. It is not.

Something is absorbing the productivity before it reaches the income statement.

Where the gain goes

Here is the mechanism, and it is not complicated once stated plainly.

Most AI investment is purchased to make an existing revenue model cheaper or faster. Same product, same buyer, same contract, same billing. Lower cost to serve.

Your competitor buys a comparable model, from a comparable vendor, at a comparable price, on roughly your timeline. The capability is not scarce. It is a line item on a price list available to anyone with a purchase order.

When cost to serve falls across an industry, price follows it down. The productivity gain is real and it is temporary, because it was never yours exclusively. It moves through your P&L and lands in your customer’s.

This is not an implementation failure. It is the expected fate of an advantage that anyone can buy. Efficiency reaches EBIT and stays there only under two conditions: competitors cannot obtain it, or you convert it into something structural before the market prices it away.

Almost nobody plans the conversion. That is the gap.

The three doors

Every technology investment opens one of three doors. Naming the door before you approve the budget is the single highest-leverage decision in an AI program, and it takes about five minutes.

Door one: efficiency. Same product, same buyer, same billing, lower cost to serve. The value is real. It decays at the speed your competitors adopt the same tool. Treat it as maintenance spending, not strategy.

Door two: capability. Same revenue model, materially better product. The advantage lasts as long as the capability gap does, which in a market where foundation models improve on a public schedule is usually shorter than the business case assumed.

Door three: revenue model. What you sell changes, or who pays changes, or when the money arrives changes. This is the only door where advantage compounds, because a competitor cannot copy it by signing a contract. They have to rebuild pricing, contracts, finance and compensation, and most will not attempt it while the current model still works for them.

The pattern I see most often in board materials is door one spending presented in door three language. The tooling is efficiency. The slide says transformation. Nobody is lying. The vocabulary simply drifted, and the drift is expensive, because it means nobody builds the plan that door three actually requires.

Three questions that tell you which door you are in

Before approving any AI investment, ask three questions. Assume the technology works perfectly, then answer.

1. Does our price list change?
2. Does our contract length or billing trigger change?
3. Does anyone new start paying us?

Three noes means you bought efficiency. That is a legitimate and often necessary purchase. Budget it as efficiency, expect the margin to be temporary, and take it off the strategy page.

One or more yeses means you are attempting a revenue model change, and the work is different in kind. Deployment is the small part. Pricing, contract terms, revenue recognition and sales compensation are the large part, and none of them belong to the technology function.

The test costs nothing. Its main value is diagnostic: it reveals, quickly and without argument, whether a program has been oversold internally.

The four moves that are commercially live in B2B

Four revenue model changes are no longer theoretical. Each is running in production somewhere, and each one breaks something you currently rely on.

1. Outcome pricing

Charge for the result rather than for access.

Intercom prices its Fin agent at $0.99 per resolution, where a resolution means the customer confirms the issue is solved, or does not ask for more help, or the agent completes a defined workflow including handoff. If Fin does not resolve the conversation, there is no charge.

The shift is far enough along that it has reached the accounting profession. In June 2026, Deloitte published guidance on accounting for outcome-based pricing in agentic AI products, framing the central judgment as whether the vendor’s promise is a stand-ready obligation to provide continuous access or an obligation to deliver a specified quantity of successful outcomes. When accounting firms publish revenue recognition guidance for a pricing model, that model has left the conference stage.

The underlying logic is simple and uncomfortable for software vendors: per-seat pricing is structurally broken for autonomous software. The better your agent performs, the fewer seats your customer needs. You are being paid to under-deliver.

What it breaks: your forecast. Revenue becomes contingent on events you do not fully control, and, per Deloitte’s framing, on judgments about the nature of your performance obligation. Your CFO inherits an estimation problem your commercial team created. If forecast accuracy is already a live issue in your business, fix that first. Outcome pricing on top of an unreliable forecast is not a strategy, it is an accelerant.

2. Consumption pricing

Charge per decision, per document, per inference, or per unit of work completed.

Futurum’s 1H 2026 enterprise software buyer research found 43% of buyers preferred consumption-based models and 27% preferred outcome-based structures, with fewer than one in five still preferring classic per-user pricing. Directionally, the buyer has already moved.

What it breaks: procurement. A buyer who cannot put a fixed number in a budget line has to escalate, and escalation is where deals stall. This is the specific reason AI deals that demo brilliantly die in the approval step, a pattern I cover in more detail for fractional CRO work with AI companies. The fix is not to abandon consumption pricing. It is to sell a cap, a floor, or a committed tier alongside it, so the buyer has a number to defend internally.

3. Data-back monetization

The operational data your business already produces becomes a product that someone else pays for.

In industrial, distribution and agribusiness companies this is usually the least explored of the four and often the most defensible, because the data is a byproduct of an operation a competitor cannot replicate. Yield data, equipment telemetry, dealer sell-through, cold chain conditions, and route performance are all assets that already exist and are typically stored, not sold.

What it breaks: contracts and trust. If your customer agreements do not grant you the rights, you do not have a product, you have exposure. And the first customer to learn about this from your marketing is the one who calls their counsel. Legal review comes before the business case, not after it.

4. Service-to-product conversion

Work you currently deliver as billable hours becomes a product with software economics. AI lowers the scale at which this becomes viable, which is why it is now available to companies that could not have attempted it three years ago.

What it breaks: your first year. You trade recognized revenue today for deferred revenue and a better gross margin later. Boards approve this readily in the abstract and react badly to the first quarterly print. Model the dip, present the dip, and get agreement on the dip before you start.

Where to start

The right door depends on your position, not on what is fashionable.

One sequencing rule matters more than the table. Measurement precedes monetization. Every one of these four moves depends on your ability to measure something reliably: an outcome, a unit of consumption, a data asset, a delivery process. Companies that change the pricing model before they can measure the thing they are pricing end up in disputes with their best customers. If your instrumentation is not ready, the honest sequence is to build the measurement first and change the model in the following cycle. That work is covered in turning AI investment into measurable revenue.

Five ways this goes wrong

1. Repricing before you can measure. The fastest route to a customer dispute is charging for an outcome you cannot evidence.
2. Letting the vendor’s pricing model set yours. Your vendor prices to its cost structure and its investors. Neither is your business.
3. Running the change without the CFO from day one. Revenue recognition, forecast variance and deferred revenue are not downstream consequences. They are design constraints.
4. Presenting door one as door three. The most common failure, and the one that quietly guarantees the follow-on investment never gets built.
5. Assuming the buyer wants it. Procurement frequently values predictability above fairness. A pricing model that is better for the customer in theory can still lose to one that is easier to approve.

What this means for the next budget cycle

The 6% figure is the useful one. It has been flat for a year while spending rose, which tells you that the constraint is not effort, money or tooling.

The companies in that 6% are not running better pilots. They changed something structural about how money enters the business, and they did it with the CFO in the room, on a measurement foundation they built first.

For most companies, the practical next step is not another use case. It is to take the AI programs already approved, run the three questions against each one, and separate the efficiency purchases from the model changes. Fund both. Describe them accurately. Stop expecting the first category to produce the results of the second.

That separation is usually a one-afternoon exercise, and it changes what the next budget cycle looks like.

Frequently asked questions

What is an AI revenue model?

An AI revenue model is the commercial structure through which a company earns money from an AI capability. It is distinct from an AI use case. A use case describes what the technology does. A revenue model describes what changes about your price, your billing trigger, or who pays you. Most AI programs have a use case and no revenue model, which is why the productivity gain does not reach earnings.

Why do AI investments improve productivity without improving profit?

Because productivity purchased from a vendor is available to competitors on the same terms. When cost to serve falls across an industry, competition passes the saving to buyers as price. The gain is real and temporary. It reaches EBIT durably only when competitors cannot obtain it, or when the company converts it into a structural change before the market prices it away.

Is outcome-based pricing for AI actually being used?

Yes. Intercom charges $0.99 per resolution for its Fin agent and nothing when the agent fails to resolve the conversation. Deloitte published revenue recognition guidance on outcome-based pricing for agentic AI products in June 2026, which indicates the model is in commercial production rather than at the proposal stage.

What is the biggest risk in moving to consumption or outcome pricing?

Two risks, in order. First, pricing an outcome you cannot measure reliably, which creates disputes with the customers you can least afford to lose. Second, forecast instability, because revenue becomes contingent on events outside your direct control. Both are manageable, and both require the CFO involved in the design rather than informed of the result.

Should a mid-market industrial company attempt a revenue model change at all?

Often yes, and frequently through data-back monetization rather than pricing. Industrial, distribution and agribusiness operations generate data that competitors cannot replicate because they do not run the same operation. That asset usually already exists and is rarely sold. The first step is a contract rights review, not a product plan.

Sources

1. McKinsey and Company, State of AI survey, figures reported August 25, 2026, survey of 1,719 professionals. Reported by The Register. McKinsey report page: mckinsey.com
2. MIT NANDA, “The GenAI Divide: State of AI in Business 2025.” 52 executive interviews, 153 leader surveys, 300 public AI deployments.
3. Intercom, Fin AI agent pricing: intercom.com/pricing
4. Deloitte, “Technology Spotlight: Accounting for Outcome-Based Pricing in an Agentic AI Software Product,” June 4, 2026: dart.deloitte.com
5. Futurum Group, 1H 2026 Enterprise Software Decision Makers survey, May 12, 2026: futurumgroup.com

About the author

Andre Magrini is a chief revenue officer and fractional CRO based in the Greater Chicago Area. He led North America for Ag Growth International and, as general manager in Brazil, scaled an operation more than 4x in three years. He served as Vice President of the Marketing and Communications Committee at the American Feed Industry Association, and is the author of seven books, including five on sales, marketing analytics and corporate governance.