Blog

When AI Spending Leaves the IT Budget

The business functions closest to the work are increasingly buying AI themselves. The strategic challenge is to preserve that speed without allowing the company's architecture, costs and accountability to fragment around it.

Listen to the summary

1:36 · Read by an AI voice

An outlined square with three navy squares inside and more squares tumbling out of it in an arc, one of them lime, on pale blue.

The AI budget has already moved.

Boston Consulting Group's 2026 Applied AI Index estimates that corporate AI spending has reached 3.3% of revenue, with 80% of it now sitting outside the enterprise IT budget. Yet Gartner found that only 22% of organizations have successfully scaled AI across multiple business units. BCG: The Formula for Agentic AI Value · Gartner: Scaling AI across business units

These findings initially appear contradictory.

Companies are spending across the organization, but relatively few have turned that activity into an enterprise capability.

The explanation is that AI purchasing has decentralized faster than AI architecture.

Marketing acquires a content platform. Customer service adds intelligence to its support system. Finance licenses a document-analysis tool. Operations builds an internal assistant. Individual employees create workflows using APIs, automation platforms and software already approved for another purpose.

Each decision may be reasonable on its own.

Together, they can produce a technology estate that no one intended, no one can fully see and no one is clearly responsible for operating.

AI is becoming everyone's budget before it becomes anyone's architecture.

Distributed spending is a rational response

The movement of AI spending outside IT should not automatically be interpreted as poor governance.

Much of AI's value is highly specific to a workflow.

A customer-service team knows which requests create avoidable work. A finance team understands where document review delays a decision. A sales team can identify which research, qualification and follow-up tasks consume time without improving the customer relationship.

These functions are often better placed than a central technology department to recognize a valuable use case.

AI tools have also become unusually easy to acquire.

A capability may arrive as a premium feature inside existing software, a direct subscription, an API, a cloud service, a consultant-built workflow or an application assembled internally by an employee. It may require no large implementation project and no obvious addition to the conventional technology estate.

This lowers the cost of experimentation.

It also allows teams to test ideas without waiting for an enterprise-wide program to define every requirement in advance.

That autonomy is valuable. Requiring every small trial to pass through a lengthy central approval process would preserve order by suppressing learning.

The objective should not be to move all AI spending back into IT.

It should be to distinguish between local experimentation and infrastructure the company is beginning to depend on.

A local purchase can create enterprise dependencies

An AI tool may be selected by one department, but it rarely remains isolated from the wider organization.

To become useful, it may need access to customer records, product documentation, internal policies, employee accounts or operational systems. It may generate information that another department relies on. It may send messages, update records or influence a customer-facing decision.

The budget may be local.

The dependencies are not.

Consider a sales team that purchases an AI research assistant.

At first, the tool summarizes public information about prospective customers. The risk and integration requirements are limited.

Its value becomes more interesting when it can read the company's customer relationship management system, identify opportunities, draft outreach and record activity. It may eventually influence which prospects employees prioritize.

The tool has not merely become more capable.

It has become part of the sales operating model.

It now depends on identity, permissions, data quality, integration reliability and rules governing how customer information may be used. Its output may affect revenue attribution, brand communication and the quality of records shared with other teams.

The same transition occurs when a support assistant moves from drafting responses to issuing refunds, or when a finance tool moves from extracting information to recommending whether an invoice should be approved.

At that point, the company is no longer managing a departmental subscription.

It is operating software.

The ledger does not reveal the architecture

Decentralized AI is difficult to manage because the company's financial and technical views rarely describe the same thing.

Finance sees payments assigned to cost centres, departments and vendors.

Technology teams see applications, APIs, data stores and integrations.

Business leaders see use cases.

None of those views necessarily reveals the complete system.

One AI workflow may appear across several categories:

  • A SaaS subscription paid by the business function.
  • Cloud usage charged to a central account.
  • Integration work funded as a project.
  • Data preparation performed by internal employees.
  • Model calls billed through another platform.
  • Human review absorbed inside an operational team.
  • External support covered by a consulting agreement.

Conversely, one vendor contract may support many unrelated use cases with different owners, risks and commercial outcomes.

This makes "How much are we spending on AI?" surprisingly difficult to answer.

A smaller McKinsey survey found that 93% of respondents had exceeded their AI budgets, while 62% had already moved beyond experimentation into active deployment. The report also noted that business units were purchasing capabilities independently and that employee-built applications could create substantial consumption outside central visibility. McKinsey: The cost of intelligence

The issue is not only uncontrolled expenditure.

It is that the accounting structure can conceal the system the organization is assembling.

Sanctioned AI can still become shadow infrastructure

The phrase "shadow AI" often suggests employees secretly using unauthorized tools.

That is only one form of the problem.

A tool may be contractually approved, securely configured and purchased by an authorized leader while still remaining invisible as part of the company's architecture.

The procurement team knows about the vendor.

The business function knows about the workflow.

The technology team may know about an integration.

But no one has connected these fragments into a complete operating picture.

This is how authorized purchases become shadow infrastructure.

The organization may not realize that three departments are separately paying for similar document-processing capabilities. Several teams may create their own connections to the same customer data. Multiple vendors may retain copies of substantially the same information.

Each initiative can pass its local review while the company accumulates duplication, inconsistent controls and avoidable dependency.

Visibility therefore cannot stop at a list of approved software.

The organization must understand what each capability does, which information it uses, what actions it may take, who depends on its output and what would happen if it stopped working.

The full cost appears only after adoption

A pilot often enters the budget as a license or a limited amount of model usage.

The operating cost emerges later.

There are integrations to maintain, evaluations to run, employees to support and exceptions to review. Data sources change. Vendors alter pricing or product behaviour. Someone must investigate failures and decide whether a new model version can be introduced safely.

The workflow may also create costs elsewhere.

A system that produces more sales opportunities may increase the qualification burden. A service assistant that escalates too frequently may transfer work rather than remove it. A document tool may accelerate extraction while creating additional verification downstream.

The local business case can remain positive while the complete process becomes more complicated.

This is why AI initiatives should not be evaluated only through the budget that originally funded them.

Their costs and benefits need to be measured across the workflow they affect.

The relevant question is not whether one department saved time.

It is whether the company produced the intended outcome more efficiently, reliably or profitably.

Companies are accumulating initiatives without managing a portfolio

Most organizations have a method for starting an AI project.

Fewer have a disciplined method for combining, scaling or ending one.

An experiment receives a sponsor, a small budget and access to a tool. If it performs well enough, usage gradually expands. More information is connected. Additional employees depend on it.

The initiative enters production without a clear moment at which the company deliberately decided to operate it as a durable capability.

Weak initiatives can persist in the same way.

The original sponsor moves roles. Usage falls but does not disappear. A contract renews automatically. An integration remains active because no one is certain whether another process still depends on it.

The company becomes good at adding AI and poor at subtracting it.

Gartner found that stronger performers treat AI initiatives as a portfolio: they monitor returns, reassess performance and reallocate or stop investments that do not deliver. These organizations reported positive returns from 81% of their AI initiatives, while lower-performing organizations frequently could not determine whether an initiative had produced a return at all. Gartner: Treating AI as a portfolio of value

Portfolio management requires every material initiative to have a few explicit properties:

  • A business owner.
  • A defined outcome.
  • A record of the data, models and vendors involved.
  • A view of recurring and variable costs.
  • A production and support owner.
  • Evidence of adoption and operational effect.
  • A review date.
  • A condition under which it will be expanded, combined or stopped.

If no one can explain what evidence would cause an initiative to end, the pilot is becoming permanent through inertia.

Centralizing every decision would solve the wrong problem

The obvious response to fragmented spending is to create a central AI committee and require every project to receive approval.

That may improve visibility.

It can also recreate the conditions that caused business teams to work around central technology functions in the first place.

A central group cannot understand every workflow as well as the people operating it. It should not decide whether a claims process, procurement review or customer-service escalation presents the best opportunity for improvement.

The business function should continue to own the problem, adoption and expected commercial result.

What the enterprise needs to centralize is not every use case.

It is the infrastructure that allows those use cases to be built and operated coherently.

This can include:

  • Identity and permission standards.
  • Approved methods for accessing company data.
  • Reusable integrations with core systems.
  • Model and vendor gateways.
  • Evaluation and monitoring capabilities.
  • Cost attribution.
  • Logging and audit records.
  • Security and incident-response processes.
  • Requirements for moving from experimentation into production.

The distinction is important.

The business unit should own the outcome. The enterprise should own the rails.

A shared execution layer preserves autonomy

A federated model allows business teams to experiment without creating an entirely separate technical stack for each idea.

A department may select the workflow and fund the initiative. But when the capability begins using sensitive information, influencing operations or serving a larger population, it moves onto shared infrastructure.

This creates visibility without requiring uniformity.

The company does not need every use case to call the same model or use the same application. It needs common mechanisms for identity, data access, monitoring, cost allocation and control.

That shared layer can also make experimentation faster.

A new team does not need to rebuild customer authentication, document retrieval or audit logging. It can use components that have already been tested. Evaluations developed for one workflow may help another team assess a similar capability.

Successful initiatives begin to produce reusable assets.

Without a shared layer, every experiment starts from zero.

BCG found that more than two-thirds of the most AI-mature companies were committing to a common enterprise architecture and control plane, rather than relying on a single-vendor stack. These companies also concentrated investment on fewer, higher-value workflows and were substantially more likely to fund a coherent multiyear program. BCG: Enterprise architecture and control planes

The objective is therefore not technical standardization for its own sake.

It is to make useful capabilities easier to reuse and easier to replace.

Portfolio visibility creates strategic options

A complete view of AI activity is often described as a cost-control measure.

Its larger value is optionality.

When the organization knows which workflows use which models, data sources and integrations, it can make better decisions about where to invest.

It may discover that several departmental tools should become one shared product.

It may find that a promising workflow is constrained by poor data rather than model performance. It may identify a capability that should remain local because its requirements are genuinely unique.

It can also change suppliers more effectively.

If every workflow is directly connected to a vendor's proprietary interface, replacing that vendor becomes an organization-wide migration. If access is mediated through shared components, alternatives can be tested without redesigning every process.

Portfolio visibility therefore supports both consolidation and competition.

It allows the company to recognize where standardization creates leverage and where diversity preserves resilience.

Finance and technology need the same view

Distributed budget ownership can remain in place, but the enterprise still needs a consolidated economic picture.

That view should connect expenditure to the workflow producing it.

A finance team does not need to inspect every model request. It does need to understand the committed cost, variable exposure, implementation burden and measured return of each material capability.

Technology teams need the same information for a different reason.

They must know which applications are becoming critical, which integrations require support and where several teams are recreating the same infrastructure.

The shared view should therefore combine:

  • What the capability is intended to improve.
  • Which function owns the result.
  • What it costs to build and operate.
  • Which enterprise systems it depends on.
  • How its effect is measured.
  • Whether it should remain local, become shared or be retired.

BCG found that companies tracking AI directly through the P&L or meaningful proxy measures reported substantially more value than those without formal measurement. The relevant lesson is not that every experiment needs a perfect financial model. BCG: Measuring AI through the P&L

It is that scaled capabilities must eventually become visible in the same management system used to evaluate the rest of the business.

The AI budget is becoming an operating model

When most AI spending sits outside IT, the company cannot manage it as a conventional technology program.

It becomes a question of how authority is distributed.

Business functions need enough freedom to identify opportunities and test them close to the work. Technology leaders need enough visibility to prevent the underlying architecture from fragmenting. Finance needs enough information to distinguish valuable investment from accumulating consumption.

None of these objectives is achieved by assigning ownership to one department alone.

They require an operating model in which local initiative and enterprise coherence reinforce one another.

At BraidNest, we see this as an execution-layer challenge.

The first step is often to establish what already exists: which tools have been purchased, which workflows have been built, what information they use, what they cost and who depends on them.

From there, overlapping initiatives can be combined, successful components can become shared infrastructure and promising experiments can be given a credible production path. Investments that no longer justify their complexity can be retired.

The aim is not to centralize innovation.

It is to turn distributed experimentation into a portfolio the business can understand, operate and improve.

AI spending has already left the IT budget. The next task is to ensure that the company's architecture, accountability and ability to create value do not leave with it.

  • AI strategy

All articles