AI is the first business technology that gets more expensive the better it works. Costs rise with use rather than headcount, and most organisations budgeted as though it were ordinary software. This analysis explains why the bills keep climbing even as prices fall, and what leaders can do about it. Written for finance and technology decision-makers.
Every business technology before this one got cheaper per person as you grew. AI works the other way round. The more useful it becomes, the more people use it, and the more it costs to run.
That surprise is now showing up in budgets. In one recent survey of enterprise organisations, more than nine in ten had overshot their AI spending. Prices per unit have fallen sharply, and bills have gone up anyway.
This matters most to finance leaders, technology leaders and anyone who signed off an AI budget that is not going to hold. The main takeaway: this needs a permanent management routine, not a one-off round of cost-cutting.
Estimated reading time: 11 minutes
Key Takeaways
AI costs scale with how much work it does, not with how many people have a licence, which breaks most existing budgeting habits.
The price of AI has fallen dramatically while total spending has risen, because the way businesses use it has changed faster than the price.
Most of the cost of an automated task is the checking and fixing afterwards, not the first answer.
The same task can cost wildly different amounts depending on the route the software takes, so budgeting on averages does not work.
A large share of AI spending in a typical business is not tracked at all, sitting in cloud bills, software features and departmental purchases.
The savings from managing this properly are substantial and well documented, and most of them do not require new technology.
The Strategic Challenge
There is a moment described in recent McKinsey research that will feel familiar to a lot of leadership teams. A group approves a new AI capability. Everyone is enthusiastic. Then somebody asks what it will cost to run.
Nobody in the room knows.
That silence is the strategic problem. Not because the costs are enormous — often they are not — but because nobody can predict them, attribute them, or explain them afterwards.
To understand why, it helps to know one piece of terminology. AI systems charge by the token, which is roughly a fragment of a word. Every question you ask and every answer you get back is billed in tokens. That is the whole of what you need to know for this article.
Here is the part that catches people out. Traditional business software charged per person. If you knew your headcount, you knew your bill. AI charges for work done. Which means a genuinely useful AI system costs more than a mediocre one, because people use it more.
McKinsey’s Enterprise AI FinOps survey, run in May 2026 across 120 enterprise participants with 75 qualified responses, found that 93% of respondents had exceeded their AI budgets. It also found that moving from a few isolated uses to organisation-wide adoption pushes AI spending up roughly fourfold. Most respondents expected another increase of at least 25% over the following year.
What happens if you ignore this? Two things, and they both hurt. Budgets get exhausted partway through the year, forcing awkward conversations with vendors and finance. And then the reflex kicks in — cut it back — which tends to strangle the uses that were actually working.
Callout — Why this is different from every previous IT cost problem Cloud computing costs scaled with the infrastructure you provisioned, which you controlled. AI costs scale with how much work people ask it to do, which you do not. That is why the old cost playbook does not transfer cleanly.
Who This Strategy Is For
This is most urgent for organisations where AI has moved past a handful of enthusiasts.
Act on this now if you are:
A finance leader who cannot currently answer what AI costs the business per month
A technology leader whose AI spending is spread across several vendors and departments
Any organisation running AI agents that take actions rather than just drafting text
A business where AI use is growing quickly and nobody has set a ceiling
Anyone approaching a renewal on an AI-related contract
This is less urgent if you are:
A small business paying a fixed monthly subscription for a handful of AI tools. Fixed pricing means fixed exposure, and building cost machinery around a predictable bill would cost more than it saves.
Still in early experiments with a small number of users and a capped spend.
The line to watch is the shift from fixed subscriptions to usage-based billing. That is the point at which your exposure changes character, and it often happens quietly at a renewal.
The Current Landscape
Something genuinely odd has happened over the past three years, and it is worth sitting with because it defies normal intuition about technology.
The price of AI has collapsed. Stanford’s AI Index recorded the cost of running a particular level of AI capability falling from $20 to $0.07 per million tokens through 2024. That is not a discount. That is a different order of magnitude.
Over roughly the same stretch, business spending on these systems tripled.
Cheaper AI produced bigger bills. That is not a contradiction — it is what happens when the way you use something changes faster than its price falls. Businesses moved from asking AI single questions to handing it whole jobs, and a whole job consumes vastly more than a question.
The management response is only now catching up. The FinOps Foundation, which sets standards for managing technology spending, publishes an annual survey of practitioners. In its 2026 edition, covering 1,192 respondents, 98% said they now manage AI spending — up from 31% two years earlier.
That is one of the fastest expansions the discipline has recorded, and it did not happen because AI became a boardroom priority. It happened because the invoices arrived.
Where organisations stand
Figure
Exceeding their AI budgets
93%
Spending increase from isolated use to organisation-wide
~4x
With mature practices for managing AI spend
20–25%
AI spending that is typically untracked
20–30%
Already limiting AI use because of running costs
~20%
FinOps practitioners now managing AI spend
98% (from 31%)
Sources: McKinsey Enterprise AI FinOps survey (May 2026) and State of AI (Aug 2026); FinOps Foundation State of FinOps 2026.
There is a second gap hiding in that table. Only a fifth to a quarter of companies have mature practices for managing this, and somewhere between a fifth and a third of AI spending is not tracked at all — scattered across cloud bills, model contracts, AI features bundled into software you already bought, and purchases made directly by departments.
Most organisations do not have an AI cost problem. They have an AI visibility problem that becomes a cost problem.
Why the Bills Keep Climbing
McKinsey identifies six reasons automated work costs more than people expect. They are worth understanding individually, because each one has a different fix.
The system keeps re-reading its own notes. AI models have no memory between steps. So when software works through a multi-step job, it resends everything that came before at each stage. Research on automated coding work found these tasks can consume around a thousand times more than a simple back-and-forth conversation.
Most of the cost is the rework. This is the finding that surprises people most. The expensive part is not producing the first answer — it is the checking, correcting and re-verifying that follows. Roughly 60% of the cost of an automated task goes on refining the answer.
Think about what that means. If you are budgeting for AI to do a job, you are mostly budgeting for it to check its own homework.
The same job costs different amounts every time. Because the software chooses its own route, it may take a different path, call different tools or retry in different ways. On identical programming tasks, researchers found a thirty-fold variation in cost between runs.
This is the single most important point for anyone building a budget. AI cost is not a price. It is a range. Budgeting on the average will be wrong roughly half the time, and wrong in the expensive direction often enough to matter.
Expensive thinking gets used for easy jobs. The most capable AI models cost considerably more to run. People default to them anyway, because it is not obvious when a cheaper one would do. Deep reasoning is worth paying for on a hard problem and is pure waste on a simple one.
How the work is organised changes what it costs. The way a job gets broken up and passed between steps can change costs dramatically without changing the result at all.
How you write things affects what you pay. Prompt length, formatting, and even language matter. Text in languages other than English breaks into more tokens for the same meaning, so the identical conversation costs more.
Future Signal Tip If your business operates in a language other than English, your AI costs are structurally higher than a comparable English-language business for the same work. Worth factoring into any benchmark you are given by a vendor.
Continue Reading with The Signal Pro
Get the full analysis, decision framework, and recommendations with The Signal Pro.
The change creates real openings, but only for buyers who prepare.
You can pay less for work you were already buying. Customer support is the most developed category. Resolution rates are measurable, several vendors compete openly, and published prices range from roughly $0.40 to $2.00 per resolution. If you handle high ticket volumes and have decent help content, the maths often works in your favour.
Your renewal is more negotiable than it looks. Procurement platform Tropic, drawing on more than $15 billion of software spend, found that structured negotiation cuts a vendor’s opening ask by around 55%. The final price still tends to land about 12% above the old baseline — the increase is negotiable, not avoidable.
Timing alone is worth money. The same data shows businesses starting renewal conversations six months ahead save considerably more than those starting a month out.
Unused licences become bargaining chips. If AI has taken over a workflow, the licences attached to it are still being paid for. Find them before your vendor’s account manager does and they become leverage. Find them afterwards and they are simply money you wasted.
You can trade commitment for protection. Vendors part-way through changing their pricing want reference customers. They will often agree to price locks, waivers on mandatory AI bundles, hard spending ceilings and credit rollover in exchange for a volume commitment. That is a very different conversation from asking for a discount.
Future Signal Tip
Before your next renewal, ask the vendor to show you the return on their AI features using your usage data — not a case study from another company. If they cannot produce it, you have just found your strongest negotiating point. If they can, you have learned something useful either way.
Strategic Risks
Every one of these risks has already caught out a well-run company.
Spending runs away before value arrives. Uber encouraged staff to use AI coding tools heavily and ranked usage on internal leaderboards. It exhausted its entire 2026 budget for those tools in roughly four months. Bloomberg reported the fix in June: a cap of $1,500 per employee per month, per tool, with a usage dashboard and a process for requesting more.
You pay for AI features you never use. Vendors bundle AI into existing products and charge for it whether or not anyone touches it. Tropic’s contract data puts these increases at 20–37%, against a historical 3–9% annual uplift. CFO Dive reported the same finding: prices are rising faster than the value buyers can see.
Comparison becomes almost impossible. Credit-based pricing is not comparable between vendors. A credit at one company is not a credit at another, which quietly removes your ability to benchmark.
Prepaid credits can expire. Salesforce Flex Credits must be used before the order period ends, with no rollover. Testing in a sandbox consumes them too. Money you have already paid can simply evaporate.
The vendor writes the definition of success. Fin’s documentation describes a conversation as resolved when the agent answers and the customer either confirms it helped or leaves without asking anything further. A frustrated customer who gives up can bill identically to a happy one.
Your current pricing is not permanent. HubSpot’s April 2026 change applied to every customer with no opt-out and no grandfathered tier.
Acquisitions reset the terms. Salesforce signed an agreement to buy Fin for around $3.6 billion in June 2026, expected to close in its 2027 financial year. Roadmap, packaging and pricing can all shift after a deal closes.
Efficiency gains often disappear. Time saved tends to be absorbed by employees as a better working week rather than converted into capacity you can redeploy. It is a real effect, and almost nobody models it.
Decision Box
Act now if: you have a renewal within twelve months, you already run AI agents without spending caps, or your support and sales workflows have visibly changed.
Pilot first if: you are considering outcome-priced tools but have no baseline data yet. Start with one measurable workflow, ideally customer support.
Prepare if: your contracts run beyond eighteen months. Use the time to audit unused licences and build measurement.
Monitor if: your software spend is small and fixed, and you have no AI rollout planned this year. Revisit before your next contract.
Whatever you choose, do not sign a consumption-based contract without a written definition of what counts as billable.
Implementation Roadmap
A practical sequence for the next six to twelve months.
STAGE 1 STAGE 2 STAGE 3 STAGE 4 STAGE 5
Assess → Control → Measure → Renegotiate → Reallocate
Weeks 1-4 Weeks 4-10 Weeks 6-14 Months 3-9 Months 6-18
List every Set spending Record where Open renewals Retire unused
tool by how caps and you started 6 months tools; move
it charges usage limits before you early, armed savings into a
you per team deploy AI with evidence managed budget
Stage 1 — Assess what you actually have. List every tool by pricing model: per seat, per usage, credits, per outcome, or hybrid. Flag every renewal falling in the next twelve months. Identify licences attached to workflows that AI has already taken over.
Stage 2 — Put controls in place before you need them. Set spending limits per person, per team and per tool, with a clear route to request more. Give budget owners and end users the same usage dashboard. Decide which approvals are needed for which type of work — reviewing a contract and running an autonomous investigation cost very different amounts.
Stage 3 — Record your starting point. For each AI use case, capture what the work costs today, how long it takes, how much of it there is, and how often it goes wrong. Agree in advance which financial line this work is expected to move, and name one person accountable for it. A baseline taken after deployment is worthless.
Stage 4 — Renegotiate from evidence. Start six months out. Ask explicitly and in writing for your existing pricing. Then negotiate the terms that actually protect you — price locks, no forced AI bundles, hard spending ceilings, credit rollover, exemption for testing environments, and a written definition of a billable outcome with a process for disputing one.
Stage 5 — Reallocate deliberately. Retire the tools your audit uncovered. Move the savings into a managed AI budget rather than letting them disappear into general costs. Review quarterly: what does each outcome cost, is that cost stable, and has any saved time actually been redeployed?
Common Mistakes
Treating the AI increase as part of a routine renewal. Vendors prefer this, because one number is harder to challenge than two. Price the base platform and the AI capability as separate decisions.
Signing before defining the billable event. This is the most expensive mistake available, and it happens because the definition sits in documentation rather than in the contract. Ask for it in writing, including what happens with partial outcomes and how you dispute a charge.
Rolling out AI tools with no spending limits. Uber is the well-documented version of a very common failure. Enthusiastic adoption plus per-use billing plus no caps equals a budget gone in a third of the year. Caps are not a brake on adoption — they are what makes wide adoption survivable.
Measuring the wrong thing. Counting logins, prompts or hours saved feels like progress and proves nothing. Boards are now asking which line of the profit and loss account changed. Very few organisations can answer.
Assuming outcome pricing is automatically cheaper. It is cheaper when your outcomes are simple and your content is good. It is expensive when your queries are complex, because you pay for every one the agent handles.
Confusing outcome pricing with transparent pricing. Fin publishes $0.99 on its website. Sierra publishes nothing at all, and independent estimates put a first year in the low hundreds of thousands. Both charge per outcome.
Success Metrics
Measure a small number of things properly rather than everything badly.
What to track
Why it matters
A reasonable target
Cost per outcome, and its spread
Averages hide the expensive cases that break your budget
Stable month to month, with the spread narrowing
Renewal increase vs the vendor’s opening ask
Tells you whether your negotiation worked
Around 55% reduction from the ask, landing near 12% above your old price
Share of software spend that is variable
Shows how much of your budget you can no longer forecast
Known and deliberate, not accidental
Consumption forecast accuracy
Early warning for the next budget overrun
Within 10% of plan each month
Prepaid credits expiring unused
Straightforward waste
Zero
Use cases with a recorded baseline
You cannot prove a return without one
100% before spending is approved
Verified resolution rate
Headline rates come from the vendor’s own definition
Sample and check resolved items yourself
Reopened cases after a billed resolution
The clearest sign your outcome definition is wrong
Falling over time
If you run a software business yourself, add one more: the share of new revenue that does not depend on your customers’ headcount. ServiceNow’s disclosed figure of 50% is a useful benchmark.
Future Outlook
What is already confirmed. Gartner expects consumption-based pricing to account for more than 35% of new corporate legal technology spending by 2028, and warned general counsel on 3 September 2026 that unprepared teams face budget shocks. Deloitte’s 2026 predictions describe the same move toward hybrid pricing across business software generally, and expect as many as 75% of companies to invest in AI agents by the end of this year. Both Deloitte and EY have published formal accounting guidance on how to recognise revenue from outcome-based AI pricing — a good sign that this is now a permanent structural change rather than an experiment.
What the industry expects. Hybrid pricing continues to spread as the default. Consolidation continues, with large platforms buying AI-native capability. The buyer’s negotiating window stays open for several more renewal cycles, since Gartner’s exposure estimate runs to 2030 rather than next year.
Our own reading, offered as interpretation rather than fact. Expect arguments about what counts as a billable outcome to become a visible commercial issue during 2027, as per-resolution billing reaches scale. Expect standard contract language for measuring outcomes to emerge, in the way that service-level agreements standardised for cloud hosting. And expect at least one vendor to retreat from pure outcome pricing back toward hybrid, because a flat fee per outcome only works while the mix of easy and difficult cases stays stable.
The Future Signal
The stock market spent early 2026 asking the wrong question. It asked whether AI would destroy software companies. The more useful question was always what software companies would start charging for instead.
The answer turned out to be unglamorous. Not a revolution in pay-for-results, but a steady move to a base fee plus a meter. That is closer to how you buy electricity than how you bought software.
The strategic implication runs deeper than procurement. When you paid per person, your software cost was a function of your hiring decisions — visible, predictable and set once a year. When you pay per unit of work, your software cost becomes a function of your operations, changing daily, driven by decisions made by people who have never seen a contract.
Very few businesses are set up for that. Most finance functions can tell you the headcount to the person and cannot tell you what a support ticket costs to resolve.
Leaders should focus on two things: knowing what a unit of work costs before they sign anything that bills by the unit, and making sure someone owns that number. What can safely be ignored is the noise about whether software is finished as an industry. It is not. It is changing what it sells, and the businesses that understand what they are buying will pay less for it.
What Businesses Should Do Next
Take these in order. The first three cost nothing but time.
List your software by how it charges you. Per seat, per usage, credits, per outcome, hybrid. Most businesses have never done this and are surprised by the result.
Mark every renewal in the next twelve months. Anything more than six months away goes to the top of the list, because that is where the savings are.
Find your unused licences. Look specifically at workflows where AI has changed how the work gets done.
Put spending caps on any AI tool billed by use — per person and per tool, with a way to request more. Do this before the first surprise, not after.
Record a baseline for every AI use case before you deploy: current cost, current time, current volume, current error rate.
Ask for the outcome definition in writing on any consumption or outcome-priced contract, including partial outcomes and how you dispute a charge.
Name an owner for each AI use case — someone accountable for a specific financial result, not a project sponsor.
Review quarterly. Cost per outcome, forecast accuracy, and whether saved time has actually been redeployed.
Pre-Renewal Checklist
Work through this before your next software contract conversation.
I know how this vendor charges — per seat, per use, per credit, or per outcome
I know what happens to unused credits at the end of the term
I have the definition of a billable outcome in writing
I know how partial or disputed outcomes are handled
I have asked for my existing pricing explicitly, in writing
I have asked for a spending ceiling to be written into the contract
I have checked whether testing environments consume paid usage
I have identified licences we no longer need
I have asked the vendor for return-on-investment evidence based on our own usage
I started this conversation at least six months before renewal
Frequently Asked Questions
Is per-seat pricing actually finished? No. It remains a good fit for software that holds your records and manages permissions — your CRM, finance and HR systems. What is fading is per-seat pricing for tools that simply perform tasks. Most vendors are moving to a mix of both rather than abandoning seats.
Will outcome-based pricing save my business money? Sometimes. It works well when your work is high-volume and reasonably simple, and when you have good documentation for an AI agent to draw on. It works badly when your cases are complex, because you pay for every one the agent handles. Model your real volumes against each pricing option before you decide.
What is the single most important thing to negotiate? The definition of a billable outcome. The headline price matters far less than what triggers a charge. Get it in the contract, not just the documentation.
Our vendor is raising prices for AI features we do not use. Is that normal? It is common. Vendors bundle AI into existing products and price accordingly. It is also negotiable — procurement data shows structured pushback substantially reduces the opening ask.
How do I stop AI spending running away? Caps and visibility, installed before the rollout. Set a monthly limit per person and per tool, with a documented way to request more, and give people a dashboard showing what they are spending.
Should a small business worry about this? Less urgently, but the same principles apply and the outcome-priced tools are often the most accessible ones. A support agent at under a dollar per resolution can be genuinely affordable for a small team — as long as you understand what triggers the charge.
Can we prove a return on AI investment? Only if you recorded a baseline before deploying. This is the most common gap. Very few organisations can currently connect AI spending to a change in profit, which is exactly why AI budgets are coming under scrutiny.
Our vendor is being acquired. What should we do? Treat your current terms as current rather than permanent. Ask your account team what is guaranteed through your renewal period, in writing, and get it before the deal closes.
Conclusion
The challenge is straightforward to state and awkward to solve. Software used to charge you for people. It is starting to charge you for work. Your budgeting, your contracts and your approval processes were all built for the first model.
The recommended approach is not to resist the change or rush into it, but to prepare properly for it. Know how each of your tools charges you. Put spending limits in place before your first overrun rather than after. Record where you started before you deploy anything. Then negotiate from evidence, six months early.
The biggest takeaway is that the money is not saved in the negotiation. It is saved in the preparation that makes the negotiation possible.
The next logical step is the smallest one: spend an hour listing your software by how it bills you, and mark every renewal falling in the next twelve months. Almost every business that does this finds something it did not expect — and finding it before your vendor does is worth more than any discount you will be offered.
The Future Signal
An independent AI intelligence publication helping business leaders make smarter technology decisions through trusted research, practical comparisons, and curated AI tools.
Related AI Tools
Clay
8.4
The spreadsheet that finds, researches and delivers the customer data your sales team is missing.
AI changes fast, but your decisions shouldn’t rely on headlines or hype. Explore more expert comparisons and practical guides—or join The Future Signal to receive one curated AI briefing each week that highlights what actually matters for business.
We use cookies and similar technologies to improve your browsing experience, analyze website traffic, and remember your preferences. With your consent, we may also use cookies to measure the performance of our content and affiliate partnerships. You can accept all cookies, reject non-essential cookies, or customize your preferences at any time. For more information, please see our Cookie Policy and Privacy Policy.
Functional
Always active
The technical storage or access is strictly necessary for the legitimate purpose of enabling the use of a specific service explicitly requested by the subscriber or user, or for the sole purpose of carrying out the transmission of a communication over an electronic communications network.
Preferences
The technical storage or access is necessary for the legitimate purpose of storing preferences that are not requested by the subscriber or user.
Statistics
The technical storage or access that is used exclusively for statistical purposes.The technical storage or access that is used exclusively for anonymous statistical purposes. Without a subpoena, voluntary compliance on the part of your Internet Service Provider, or additional records from a third party, information stored or retrieved for this purpose alone cannot usually be used to identify you.
Marketing
The technical storage or access is required to create user profiles to send advertising, or to track the user on a website or across several websites for similar marketing purposes.
