How Much Should We Budget for AI in 2027?

A guide for CEOs, CFOs, CIOs and transformation leaders setting next year's number.

Short answer. There is no benchmark percentage worth using, and the evidence for that comes from the firms publishing the benchmarks. Build the number bottom up from committed outcomes, and structure it as four lines that behave differently and should be governed differently: what it costs to run what is already live, what it costs to build what is next, what it costs to adopt each of those into the way work is actually done, and what it costs to prove the result. Add a fifth, capped line for genuine experiments, so they stop hiding inside the other four. The run line now behaves like a variable utility bill. The adopt line produces almost all of the return and is the first thing cut, because it is the only line with no vendor to defend it. A 2027 budget that funds build without funding adopt has already forecast its own result.

Spending is accelerating. Returns are not.

Gartner forecast in September 2026 that worldwide AI spending would reach $2.7 trillion in 2026, up 49.5% on the previous year, and would grow a further 36.2% in 2027 to roughly $3.64 trillion. Inside that total, AI infrastructure rises from about $1.48 trillion to $1.98 trillion, AI software from $462 billion to $656 billion, and spending on AI agents and assistants grows 77% in 2026 to around $29 billion before more than doubling again in 2027 to around $65 billion.

Set that against what the money has produced. McKinsey's State of AI survey, published on 25 August 2026 and based on 1,719 respondents across 97 nations, found that 80% of respondents said AI had improved their own productivity. It also found that 37% attributed at least some positive EBIT impact to AI, essentially unchanged from the previous year, and that about 6% qualified as AI high performers, attributing 5% or more of EBIT to AI with significant impact. That 6% was unchanged too.

Adoption, meanwhile, is close to universal. Stanford HAI's 2026 AI Index, published in April 2026, puts organisational adoption at 88% in 2025, up from 78% the year before, drawing on McKinsey's own survey data rather than an independent count.

So individual productivity gains are widely felt, adoption is near-saturation, and enterprise financial impact has not moved since the previous survey.

The instinctive reading is that AI is overhyped. The data does not support that, and the more useful reading is available in the same research. McKinsey's March 2025 report tested 25 organisational attributes against EBIT impact from generative AI and found that the redesign of workflows had the single biggest effect of any of them. The August 2026 survey is consistent with it: nearly three quarters of high performers report fundamentally redesigning workflows, against about a quarter of everyone else.

That is the gap. And it is a budgeting gap before it is anything else.

If someone quotes the 95% figure at you

Somewhere in the next six weeks, someone will say that 95% of AI projects fail. It lands heavily because it sounds like permission to be cautious.

It comes from The GenAI Divide: State of AI in Business 2025, a Project NANDA paper with MIT-affiliated authors, circulated in July 2025 as a draft labelled preliminary findings. Its headline claim is that 95% of organisations in its evidence base were getting zero measurable profit-and-loss return from generative AI, rather than that 95% of pilots fail. The report is inconsistent on this, using failure-rate language elsewhere in its own text, and it never reconciles the 95% to a single denominator or says which of its three evidence bases produced it. Those bases were structured interviews with representatives of 52 organisations, survey responses from 153 senior leaders, and a review of over 300 publicly disclosed AI initiatives, with return measured six months after the pilot, a window the authors themselves flagged as possibly too short.

The widely circulated "$30 to $40 billion wasted on enterprise AI" line is a rewriting of the same paper's figure for total enterprise generative AI investment. The paper never calls that money wasted, and the two statistics are one source, not two.

Read carefully, it says that most organisations had not converted generative AI into measurable financial impact within six months. That is a finding about organisational absorption, and it is an argument for funding absorption. Used as it usually is, it becomes an argument for funding nothing.

The return comes from the work nobody invoices you for

Workflow redesign does not generate an invoice. That is the whole problem.

Cloud capacity generates an invoice. Model access generates an invoice. Licences, platforms, integration partners, security tooling and data pipelines all arrive with a vendor, a quote and a contract. They are easy to cost, easy to approve and easy to defend, because somebody outside the organisation is asking to be paid.

The management time spent deciding which steps of a process disappear generates nothing. Neither does the training that teaches an underwriting team to work differently, the handover protocol between a person and an agent, the rewritten role profiles, the new quality checks, or the six weeks in which throughput dips while people learn. Those costs are real and substantial, they are borne internally, and they are invisible to a procurement-shaped budget process.

Boston Consulting Group has a rule of thumb for this, which it calls 10-20-70: roughly 10% of effort to algorithms, 20% to technology and data, and 70% to people and processes, so that the changes stick. BCG has repeated it across several publications to make the point that the value curve bends only when organisations invest beyond the tool itself.

The rule is widely quoted and almost as widely misapplied. Two cautions.

BCG itself states it two different ways. In its January 2026 piece on scaling AI, it is a guide to effort allocation, with algorithms first. In a June 2026 publication written for CIOs, it describes where value comes from, with the first two proportions reversed: roughly 10% of value from the technology, 20% from data and algorithms, 70% from people, process and operating model change. Those are different claims about different things. Treat the ratio as directional.

The second caution matters more. The rule describes transformation effort, not your total AI bill. Take an AI budget that includes cloud capacity, inference, security and platform licences, force it into a 10-20-70 shape, and you will produce a nonsense number, because much of that budget is infrastructure you pay for whatever your transformation ambition is. Conflating the two is a common way a sound principle produces an indefensible budget.

The usable version is narrower. For every use case you intend to put into production, the effort and money required to change how people work around it will typically be several times what the technology costs. Budget each use case that way, and keep the money in a line of its own where it cannot be quietly reallocated.

Three things changed between the 2026 budget and the 2027 budget

A 2027 AI budget is not a 2026 budget with a growth rate applied. Three structural shifts make last year's method unreliable.

The bill became variable

For three years, enterprise AI costs behaved like software costs: largely fixed, broadly predictable, negotiated annually. That is ending.

Gartner reported in August 2026 that 2026 is the first year in which spending on inference surpasses spending on training within AI-optimised infrastructure as a service, at roughly $23.3 billion against $19 billion, with inference forecast to take 59% of that market in 2027. That segment is a narrow slice of the infrastructure total quoted earlier, which is overwhelmingly servers, semiconductors, networking and AI-capable devices rather than cloud inference capacity, but the direction it shows is the one that lands in your budget. Inference is not the cost of building a model. It is the cost of using one, every time, and it scales with how much work the system does.

In August 2026 Gartner forecast that inference cost per agentic workflow would rise more than fivefold through 2028. In March 2026 it had separately forecast that by 2030 the cost to generative AI providers of running inference on a one-trillion-parameter model would be over 90% lower than in 2025, while warning that those provider savings would not be fully passed on to enterprise customers. The same March analysis put agentic models at roughly five to thirty times the tokens per task of a standard chatbot exchange, and observed that token consumption is growing faster than token costs are falling.

Read those together and the conclusion for a budget is uncomfortable. Unit costs are falling, the savings are only partly reaching you, and the thing you are buying is getting larger faster than its price is falling. A line item priced on 2026 chat-style usage will be wrong if that workload becomes agentic during 2027, and it will be wrong by a multiple rather than by a margin. Any 2027 budget containing agentic workflows that states inference as a single fixed annual figure is a forecast with no error bars on its most volatile item.

One further wrinkle for anyone budgeting outside the United States: inference is priced in US dollars. If your budget is in Singapore dollars, dirhams or rupees, the run line carries an unhedged currency exposure that nobody in the planning cycle usually owns.

Agents moved into the operating column, unevenly

McKinsey found that 40% of respondents at organisations above $1 billion in revenue were scaling AI agents in at least one business function, up from 27% a year earlier. Deloitte's 2026 State of AI in the Enterprise, based on 3,235 business and IT leaders across 24 countries, found 74% expecting their organisations to be using AI agents at least moderately by 2027. BCG's AI Radar found over 30% of 2026 AI investment directed at agentic AI, with 90% of CEOs expecting measurable return from agents within the year.

Those numbers invite a confident agentic line. Three others should temper it.

The same Deloitte survey found that only 21% of organisations report a mature governance model for agentic AI. Agents are being scaled roughly three times faster than the controls around them, which is the precondition for the cancellations below rather than a separate problem.

Stanford's 2026 AI Index reports that scaled agent deployment remained in the single digits across nearly all business functions, which suggests much of what is reported as scaling agents is closer to assisted workflow than to autonomy. And Gartner predicted in June 2025 that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls, alongside a pattern it called agent washing, in which existing assistants, chatbots and robotic process automation are rebranded as agentic. Gartner estimated only around 130 of the thousands of vendors claiming agentic capability were delivering it.

The implication is not to avoid agents. It is to fund them as a portfolio carrying an expected failure rate, in a line you are willing to stop, instead of as a committed programme you will be embarrassed to cancel.

For regulated firms in Asia, governance acquired a date

This is the shift most likely to be missing from a 2027 plan, and it is the one with a deadline.

On 7 October 2026 the Monetary Authority of Singapore issued its Guidelines on Artificial Intelligence Risk Management. They apply to all financial institutions, in a manner proportionate to size and risk profile, and to all AI that derives outputs from learned premises, with generative AI and AI agents expressly included. Rule-based systems, tools whose outputs come from manually specified formulae, and robotic process automation without learning or adaptation fall outside the definition, so an old scorecard running deterministic rules is not caught. The Guidelines take effect on 7 October 2027, with the oversight and risk management expectations in Sections 3 and 4 applying from that date and the lifecycle control and capability expectations in Sections 5 and 6 by 7 October 2028.

Read the requirements as budget lines, with their dates attached, and the picture is specific. By October 2027, boards and senior management must provide effective oversight with defined roles, risk appetite and frameworks, and firms must maintain an inventory of AI use cases with materiality assessment. By October 2028, they must apply controls across the development, deployment and maintenance lifecycle, covering data governance, testing, human oversight, cybersecurity, monitoring and change management, and they must hold the capabilities and capacity to run AI safely, which MAS frames as both skilled people and resilient infrastructure. On third-party AI, where residual risk cannot be kept within the firm's risk appetite, MAS expects the firm to consider limiting or suspending the service, or replacing the provider.

For a MAS-regulated institution this converts the prove line from a sensible discipline into a dated obligation falling inside the 2027 financial year, and it converts AI capability building from a development perk into a supervisory expectation. Three things follow for the budget.

The split across the two dates sets the sequence. The inventory falls in the first tranche and the controls in the second, and you cannot scope the controls until the inventory exists, so the 2027 number carries the inventory work and the start of the capability build rather than the whole programme. The capability requirement is a training and hiring line, not a technology line. And the third-party provision means the AI embedded in software you already bought is in scope, which is usually the part nobody has counted.

Firms regulated elsewhere in Asia and the Gulf should not read this as a Singapore curiosity. It is the clearest dated example of a direction that is general, and it is a reasonable template for what a 2027 prove line needs to contain.

The five lines of a 2027 AI budget

A defensible AI budget separates spending that behaves differently, fails differently and should be governed differently. Presenting it as one number is the most common structural error, because it lets the most important line be cut invisibly.

Four of the five lines carry a warning worth stating plainly.

Run is the line that will surprise you, because it is the only one whose variable half can move by a multiple. It should carry an expected case, a disclosed ceiling, and a written trigger naming who acts when consumption crosses a threshold.

Build is the line everyone negotiates and the line that matters least to the outcome. It is visible, it is where vendors concentrate, and it absorbs most of the planning energy. On the evidence above, it is not where returns are decided.

Adopt is the line with no external advocate. Nobody outside the organisation will call to chase it, which is exactly why it needs ring-fencing and a named business owner instead of a percentage buried inside a project cost. It is also the hardest of the five for an internal team to size alone, which is the problem Massivue's AI Workforce Transformation work is built around.

Prove is the line that protects next year's budget. If 37% of organisations can attribute any EBIT impact to AI, it is reasonable to suspect that some of the remainder are not failing so much as failing to measure. No published study separates those two groups, so treat that as a hypothesis rather than a finding. What is not hypothetical is the consequence: an organisation that cannot produce a credible number when the board asks will find next year's request discounted regardless of what it actually delivered.

One gate, applied to every build line

If you take a single operating rule from this article into your planning process, take this one.

No build line is approved without an adopt line attached to it, owned by the function whose work changes, and sized by that function rather than by technology.

The gate is simple and unusually effective. It forces a conversation that otherwise never happens: a sponsor asked to name the owner and cost of the process change will either produce one, in which case the use case is real, or will discover nobody has thought about what actually changes, in which case it is not ready. Both outcomes are useful and both are cheaper to reach in a planning meeting than in production. It puts the money where the evidence says the value is. And it creates a natural stopping point, because a function unwilling to fund the change required to use something is telling you it does not want it badly enough to deliver the benefits attached to it.

How to build the number from commitments

Executives want a percentage. There is not one worth having, and the clearest evidence comes from the firm publishing the most widely cited figures.

In its AI Radar survey published in January 2026, covering roughly 2,400 executives, BCG reported that companies expected to roughly double AI spending in 2026 to about 1.7% of revenue, up from 0.8%. In its Applied AI Index released on 30 September 2026, covering 1,330 executives, BCG reported AI spending at 3.3% of revenue, risen from about 1.7% in late 2025, and added the finding that explains much of the movement: more than 80% of AI spending now sits outside the enterprise IT budget.

On BCG's own reporting, then, the figure runs 0.8%, then 1.7%, then 3.3% in little more than a year, and most of it is no longer where a technology budget would find it. The two surveys do not even agree on what 1.7% was: the January report has it as a 2026 expectation, the September one as a late-2025 actual. A benchmark that doubles inside a year, disagrees with itself about its own starting point, and mostly sits outside the budget you are looking in cannot tell you what to spend. It can tell you whether you are unusual, which is a different question and rarely an important one.

Build the number from five questions instead.

1. What have we already committed to running? Everything live today has a cost next year whether or not you approve anything new. Establish that floor, with the variable component modelled as a range. It is almost always larger than expected, because inference has typically been growing quietly inside a cloud bill rather than appearing as an AI line.

2. Which outcomes are we committing to in 2027? Not themes, not capabilities, not an ambition to be AI-first. Specific outcomes, each with an owning function, a measure and a number. Three or four committed outcomes produce a better budget than thirty aspirations, because each can be costed and each has somebody answerable.

3. What does the technology cost for each of those outcomes? Build, integration, and the data remediation that specific use case requires.

4. What does the change cost for each, sized by the receiving function? Process redesign, role and authority design, training, management time, and the temporary dip in throughput. This is a separate question from the third on purpose. Merge them and you have recreated the single undifferentiated number this article is arguing against.

5. What would it cost to prove the result? Baseline measurement before deployment, a defined attribution method, independent challenge for anything material or regulated, and, where a supervisory deadline applies, the evidence that deadline requires.

Add those and you have a number built from commitments, which means you can defend every line of it. You can also answer the question that always follows, which is not "why this much" but "what happens if we approve half". A budget built this way answers by naming which outcomes come off the list. A budget built from a percentage cannot answer it at all.

If you have not established where your organisation currently sits, Massivue's AI Maturity Assessment evaluates AI readiness across five dimensions and gives a starting position for the first question.

Three ratios that catch a bad budget

A bottom-up number cannot be validated against a percentage. It can be validated against its own shape. These are judgement tests rather than researched thresholds, but they catch more bad AI budgets than any revenue benchmark will.

If the adopt line is under a third of the build line, the change cost has almost certainly been estimated by the technology function rather than by the function whose work changes.

If the run line is under a third of the total in an organisation with more than three live use cases, the variable consumption has probably not been found yet. It is in the cloud bill.

If the prove line is zero, the budget is undefendable next November, and in a regulated firm it may also be non-compliant.

Cost per workflow, and where it stops working

Most AI budgets are denominated in the wrong unit: per seat, per licence, per project or per token. None of those connects to anything a business leader cares about.

The unit that does is cost per completed workflow, fully loaded. What does it cost, all in, to process one claim, open one account, resolve one service request, produce one credit memo? Measured before deployment and after, that single number does more work than any other metric available.

Two of its properties matter most. It survives contact with a CFO, because it is the form in which operational economics are already discussed. And it exposes the agentic multiplier early, because a workflow costing a few cents as a single model call and thirty cents as a multi-step agent loop shows the difference in a test rather than in a production bill three months after go-live.

It also changes the board conversation. "We spent $4 million on AI" invites scepticism. "Our cost per processed claim fell from X to Y across 180,000 claims, and here is the independent check on that number" invites a question about which process is next.

The unit does not fit everything, and the exceptions are instructive. Coding assistance and general-purpose summarisation have no countable completed workflow, which is precisely why their value cases are permanently contested and why they should not be the use cases carrying a budget increase. If a use case cannot be denominated in a unit, fund it from the experiments line and stop claiming a return for it.

One practical warning. Baselines cannot be reconstructed retrospectively. A 2027 use case deployed without one will have a value that is a matter of opinion for its entire life.

The costs that appear in no vendor quote

A first agentic deployment tends to go over budget in a predictable sequence, and none of the overruns appear in the original quote.

It begins with data. Not a data strategy, the specific remediation this one use case requires, which only becomes visible when somebody tries to use the data for something real and frequently turns out to be the largest single element of the build. Then integration, because connecting an agent to a core banking platform or a policy administration system is not an API call, and Gartner's agentic research flags integration into existing systems as a source of workflow disruption and costly change, even though it is not one of the three causes of cancellation Gartner names.

Then the throughput dip, which arrives after go-live and is the most misread cost in the sequence. For a period after a workflow changes, the team is slower. This is normal and temporary, and organisations that have not budgeted for it tend to interpret it as failure and abandon the change at the point it was about to pay back.

Running underneath all of it is management attention, the scarcest resource in any transformation and the only one that cannot be purchased. A programme requiring sustained attention from four senior leaders is competing against everything else those leaders could be doing, and if that is not acknowledged in planning it will lose that competition silently.

Finally, the evidence trail: a defensible record of what a system is permitted to do, what it was tested against, how it performed and who approved it. Massivue's AI CoE Playbook sets out how organisations separate delivery from independent assurance, which is the structural choice that determines how expensive producing that evidence becomes later.

Size these explicitly. Do not add a generic contingency, which invites negotiation in a way that a named cost with an owner does not.

A $4 million budget with a $6 million ceiling

The following is illustrative, not a benchmark. The figures are constructed to demonstrate the method and the relationships between the lines. They are not derived from survey data. Your numbers will differ, and the point is the reasoning, not the totals.

Consider a mid-sized MAS-regulated bank with 4,000 employees. Six AI use cases are live: document summarisation in credit, a service agent in the contact centre, coding assistance in technology, and three smaller ones. Two more are proposed for 2027, one of them agentic: an automated first-pass review in trade finance document checking.

Run: expected $1.85m, ceiling $3.83m. The six live use cases cost $1.1m a year today, $700k of fixed platform and licence cost and $400k of consumption. Usage is growing about 40% a year as more staff use the tools, an assumption to be replaced with your own measured rate. On that basis alone the 2027 line is $1.26m.

But the contact-centre agent, which carries about $90k of today's consumption, is scheduled to move from assisted to agentic during the year. Gartner's published range of five to thirty times the tokens per task puts that one use case somewhere between $450k and $2.7m. The 40% growth is not applied on top of it, because that growth comes from more staff using the tools, while the contact centre's volume is set by customer contacts rather than by internal adoption. The other five cases, grown at 40%, come to $434k.

The reasoning that produces a usable number matters more than the number. The thirty-times bound describes an open-ended multi-agent loop. This deployment is bounded: a fixed number of tool calls per case and a human reviewer on every escalation, which puts realistic consumption nearer eight times. The expected case is therefore $700k fixed plus $434k plus $720k, or $1.85m. The full thirty-times outcome is held as a disclosed ceiling of $3.83m, not because it is likely but because the honest position is that nobody knows yet. The trigger is written into the plan: if monthly consumption exceeds a stated threshold for two consecutive months, the agentic rollout pauses pending a cost-per-workflow review.

Build: $600k. Trade finance at $380k, of which $240k is data remediation. Credit memo drafting at $220k.

Adopt: $1.28m. For trade finance, which changes how 60 people work: process redesign and mapping $260k, role profiles and authority limits and the human handover protocol $140k, training to a defined standard $75k, new quality checks and a sampling regime for the first six months $160k, six weeks of reduced throughput $340k, and management time $125k. That is $1.10m against a $380k technology build, a ratio of roughly three to one. The ratio is not an error. It is the point. The credit memo case carries a further $180k.

Prove: $260k. Baselines on both new use cases, a defined attribution method, independent assessment of the agentic case before go-live, and the inventory and materiality work the MAS Guidelines require ahead of October 2027. Of this, $150k attaches to the trade finance case specifically, which matters when the budget is cut.

Experiments: $150k, capped. Written off by default, reviewed in June.

The expected total is $4.14m, with a disclosed ceiling of $6.12m.

One thing this budget deliberately does not contain is worth saying out loud. Both new use cases go live late in the year, so their own running costs barely register in 2027 and land in full in 2028. That is not an oversight, it is the shape of the problem: the first full year of run cost for anything you build this year arrives in next year's budget, by which time the project is finished, the team has moved on, and nobody is expecting it. Build with that in mind, and note it in the plan so your successor is not ambushed.

Four things about this budget matter more than its size.

First, the run line is 45% of the expected total. Nearly half of this budget buys nothing new. That is the normal condition of a maturing AI estate, and most organisations discover it in March rather than in November.

Second, the request answers its own cut. Asked what would come out to reach $2.5m, the answer is immediate: the trade finance case, taking its build, adopt and prove costs with it, landing the plan at $2.51m.

Third, and least obvious, that cut does nothing to the risk. The run ceiling is unchanged, because the volatility sits in a use case that is already live and already scheduled to become agentic. The discretionary part of the budget and the risky part of the budget are not the same part. Most organisations learn this after the cut has been made.

Fourth, a reader doing the arithmetic against their own revenue line will notice that $4.14m is likely to land well below the 1.7% to 3.3% BCG reports. That is what a bottom-up number built from a handful of committed outcomes looks like against a benchmark that sweeps in AI embedded in products, hiring, and spending scattered across business units. The gap is real and it is worth confronting in the budget meeting rather than discovering in it. If you want to close it, the honest route is more committed outcomes, not a larger number against the same ones.

What to stop funding

Budget season is the only reliable opportunity to stop things, and most organisations waste it. Gartner's expectation that more than 40% of agentic AI projects will be cancelled by the end of 2027 reads better as a planning instruction than as a warning. Some of that 40% is in your portfolio now, and it is cheaper to find it in November than in August.

Apply a three-part funding test to every existing and proposed initiative.

1. A named owner in the business, not in technology, answerable for the outcome rather than for the delivery.
2. A success test defined before the work started, capable of returning a negative result.
3. A stop condition, a stated circumstance in which the organisation would end it.

An initiative that fails any of the three is not an investment. It is an experiment, which is a legitimate thing to run, and it belongs in the capped experiments line with a review date rather than in the main budget where it will quietly consume resource for another year.

Two further controls belong with the work rather than with the funding decision, and should not be confused with it: whether the action the system takes is reversible, and if not, what approval gate sits in front of it; and whether a baseline exists against which any claimed improvement can be measured.

How to defend the number

Three conversations are coming, and they need different answers.

The CFO will ask what last year's money bought. Answer in operational units and be specific about attribution. Cost per workflow before and after, volume processed, and an honest statement of what cannot be attributed. Resist the temptation to claim a productivity figure derived from hours saved multiplied by a loaded hourly rate, unless those hours were genuinely removed from the cost base or redeployed to something you can name. CFOs have seen that calculation and discount it heavily, usually correctly.

The board will ask whether the organisation is moving fast enough. That deserves a comparative answer, not a defensive one, and the structure of the budget supplies it: here is what we are committed to running, here is what we are building, here is what we are changing about how the organisation works, here is how we will know. A board that can see the adopt line understands immediately why this is a transformation and not a procurement. It is worth knowing, going in, that BCG's AI Radar found 94% of organisations intending to continue or increase AI investment even if current initiatives do not pay off, with only 6% saying they would pull back. The real question in most boardrooms is not whether to spend. It is whether anyone can say what the spending is for.

Somebody will ask why the number is not a percentage. Answer directly: because the most cited benchmark doubled within a year and more than 80% of it sits outside the technology budget. Then show the lines.

For the finance, strategy and technology leaders who have to construct and defend that case in financial language, Massivue Academy's Total Economic Impact of AI micro-credential covers total cost of ownership including adoption drag, three-year models with a value realisation curve, and the NPV, payback and sensitivity analysis a board-level case requires.

Underneath all three conversations sits a question that is not about money at all: who holds decision rights over AI, where capability sits, and who is permitted to stop a system. That is operating model territory, and it is what Protum, Massivue's AI operating model framework, exists to settle. A budget is downstream of an operating model. If nobody can say who is allowed to stop an AI system, no budget structure will save you.

Where to start

Four things, in order, before the plan locks.

Find the run floor. Separate variable AI consumption from fixed platform cost inside your cloud bill and establish its growth rate. Until this is known, every other number is speculative.

Force the outcome list down to three or four commitments with named owners and measures. Removing things is the work. Massivue's free AI Business Strategy Canvas is a practical way to run that reduction with a leadership team in a single session.

Make each business owner size their own adopt line. If an owner cannot describe what changes about how their team works, the use case is not ready for the 2027 budget.

Baseline the processes you intend to change. This is the only item on the list that becomes impossible later.

The organisations that spend well in 2027 will not be the ones with the largest AI budgets. They will be the ones that can say, out loud and with evidence, which of last year's initiatives they stopped and why.

Sandeep Joshi, Massivue's Founder and Managing Director, takes these conversations directly. If your 2027 number is being built this month, you can book a consultation with him.

Frequently asked questions

How much should we budget for AI in 2027?

There is no reliable benchmark percentage. BCG reported AI spending at about 1.7% of revenue in January 2026 and 3.3% in September 2026, and found that more than 80% of it sits outside the enterprise IT budget. Build the number bottom up instead: what it costs to run what is live, the technology and change cost of each committed outcome, and the cost of proving the result.

What percentage of the IT budget should go to AI?

The wrong frame, for a specific reason. BCG's finding that more than 80% of AI spending now sits outside enterprise IT means an IT-budget percentage systematically undercounts what you are actually spending. It also cannot answer the only question that matters in a negotiation, which is what comes off the list if the number is cut.

What should an AI budget actually include?

Five lines with different owners and rules: run (platform, licences, inference, monitoring and support for live systems), build (new use cases, integration, data work), adopt (workflow redesign, role design, training, manager time, the throughput dip), prove (baselines, measurement, governance evidence, independent assurance), and a capped experiments line for work with no committed outcome yet.

Should we invest in AI agents in 2027?

Selectively, and funded as a portfolio rather than a committed programme. Gartner predicted that more than 40% of agentic AI projects would be cancelled by the end of 2027 on cost, value and risk-control grounds, and has forecast inference cost per agentic workflow rising more than fivefold through 2028. Budget agentic use cases with a cost range, a consumption trigger and a stop condition.

How much should we allocate to AI training?

Less than you probably think as a share of the adopt line, and more than you currently do in absolute terms. Formal training is the most visible element of adoption and usually the cheapest; process redesign and management time are the expensive parts. Training not attached to a redesigned workflow rarely changes behaviour, which is why training-only AI programmes seldom appear in financial results. For MAS-regulated firms, note that the AI Risk Management Guidelines treat capability and capacity as a supervisory expectation rather than a development perk.

How do we avoid funding AI experiments that never reach production?

Separate the budget. Run a capped experiments line with review dates, and hold the main budget to a higher standard: a named business owner, a success test defined in advance that can return a negative result, and a stop condition. Anything failing those three is an experiment and should be funded as one.

Sources

• Gartner, Gartner Forecasts Worldwide AI Spending to Grow 49.5% in 2026, 16 September 2026.
• Gartner, Gartner Forecasts Worldwide AI-Optimized IaaS Spending to Grow 96% in 2026, 10 August 2026.
• Gartner, Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028, 17 August 2026.
• Gartner, Gartner Predicts That By 2030, Performing Inference on an LLM With 1 Trillion Parameters Will Cost GenAI Providers Over 90% Less Than in 2025, 25 March 2026. Source of the five to thirty times agentic token range.
• Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025.
• McKinsey & Company, The state of AI in 2026: On the road to ROI, 25 August 2026. 1,719 respondents across 97 nations.
• McKinsey & Company, The state of AI: How organizations are rewiring to capture value, March 2025. Source of the 25-attribute finding on workflow redesign.
• Boston Consulting Group, Scaling AI Requires New Processes, Not Just New Tools, January 2026, and How CIOs Can Prove the Value of Technology in the Age of AI, June 2026, on the 10-20-70 rule.
• Boston Consulting Group, As AI Investments Surge, CEOs Take the Lead, AI Radar, January 2026. Roughly 2,400 executives.
• Boston Consulting Group, AI Is Starting to Pay Off. Almost 50% of Companies Now Generate Value with It., Applied AI Index 2026, 30 September 2026. 1,330 executives.
• Deloitte, Business and IT leaders report AI agents are scaling faster than their guardrails, 2026 State of AI in the Enterprise. 3,235 respondents across 24 countries.
• Stanford Institute for Human-Centered AI, The 2026 AI Index Report, 13 April 2026. Adoption and agent-deployment figures draw on McKinsey survey data.
• Monetary Authority of Singapore, MAS Sets Out Supervisory Expectations on Responsible AI Adoption by Financial Institutions, 7 October 2026, and the Guidelines on Artificial Intelligence Risk Management themselves.
• Project NANDA (MIT-affiliated authors), The GenAI Divide: State of AI in Business 2025, July 2025, circulated as a preliminary draft.

This article reflects Massivue's interpretation of published research and observations from client work. It is not legal, compliance or financial advice. Figures in the worked example are illustrative and are not benchmarks. Analyst forecasts are estimates and are revised frequently, and regulatory instruments change: confirm the current position with the issuing organisation before relying on it for a decision.