A new IBM Institute for Business Value study published on November 17 puts a management problem behind enterprise AI adoption into unusually clear terms: many of the executives expected to deliver AI results do not control the investment decisions needed to deliver them. In a survey of 1,667 senior AI leaders, only 18% said they had full control over AI investment decisions. Just 29% said their organisations quickly adjusted decisions and resources when performance evidence revealed an opportunity to improve outcomes. At the same time, nearly half said AI agents were already initiating or executing actions across workflows, while three-quarters said governance, risk or compliance processes delayed AI-driven action. The figures describe survey respondents, not every company, but they point to a practical issue for boards and operating leaders. AI value depends on who can change budgets, processes, controls and responsibilities after a system is deployed.
Decision-maker takeaway
Give AI initiatives named business owners, measurable operating outcomes and explicit authority to scale, redesign or stop work. A senior AI title cannot compensate for decision rights that remain fragmented across finance, technology, risk and business units.
What the new IBM study actually found
IBM’s research, conducted with Oxford Economics, surveyed 1,667 senior executives responsible for enterprise AI strategy, governance, implementation and value realisation. The respondents included chief AI officers, chief data and AI officers, heads of AI and equivalent leaders. The fieldwork took place from April to June 2026 across 33 geographies and 20 industries. The study was published on November 17, so the release is current; the underlying survey is not a measurement of activity during the last week. That distinction matters when interpreting the results. It is a snapshot of executives’ reported organisational practices, not a live audit of AI deployments or a causal experiment.
The headline finding is not simply that few AI leaders control the budget. It is that authority appears to lag accountability. The same report says fewer than one-third of respondents believed AI was reshaping their business model, and only 29% said their organisations could adjust decisions and resources quickly when performance insights showed where improvement was possible. Forty per cent said opportunities were discussed but action was delayed. In other words, data may be available and leaders may recognise what should change, yet the organisation can still struggle to move people and money.
IBM also compared a group it describes as high-impact AI leaders with other respondents. The high-impact group reported measurable outcomes for 63% of AI projects, compared with 43% among the rest, and said 49% of projects moved from pilot to production, compared with 33%. They reported scaling successful initiatives enterprise-wide in eight months rather than 13.5 months. Those are self-reported group comparisons. They should not be read as proof that a particular governance structure caused the difference, but they are useful signals about the practices associated with stronger reported outcomes.
Why the authority gap is an operating-model issue
AI programmes often begin inside technology teams because those teams can procure platforms, connect data and run pilots. The business case, however, usually sits elsewhere. A customer-service assistant changes the work of service leaders. A forecasting model affects finance and supply planning. A coding agent changes engineering review and release practices. A document-processing system touches legal, compliance and customer operations. If the AI leader can recommend a change but cannot secure the process owner’s time, the finance approval or the risk decision, the initiative becomes another layer of experimentation rather than a redesigned way of working.
This is why a chief AI officer should not be judged only on model selection, deployment counts or the number of teams using a tool. The role needs a clear mandate that connects strategy to operating decisions. That mandate should specify which decisions the AI leader can make independently, which require a business sponsor, which need risk approval and which belong to the executive committee. It should also say who owns benefits after a pilot is handed into a business unit. Without those boundaries, every disagreement becomes an escalation and the organisation loses the speed that AI was meant to create.
For smaller and mid-sized organisations, the lesson is not that every company needs a new C-suite title. A COO, CIO, finance director or transformation lead can perform the coordinating role if the board makes the remit explicit. The critical design choice is to appoint one accountable owner for the portfolio and then give process owners responsibility for adoption and outcomes. A title without authority is symbolic. Authority without business ownership is equally risky.
Agents make unclear decision rights more expensive
The IBM survey says 48% of respondents reported AI agents initiating and executing actions across workflows. That is a meaningful shift from a tool that merely drafts text or recommends a next step. An agent may retrieve information, update a record, trigger a workflow, contact a customer or pass work to another system. The moment an AI output can change an operational state, the organisation needs to know what the agent is permitted to do, which conditions require human review and who is accountable when the action is wrong.
The survey also found that 73% of respondents said their organisations had defined where human oversight was non-negotiable, while 61% said AI decision rights were clearly defined and consistently applied. Those percentages suggest a gap between policy statements and consistent execution. A company may say that a person must approve a high-risk action, but the rule can be applied differently across departments, vendors and software environments. A written policy is not enough if system permissions, escalation paths and audit logs do not reflect it.
Executives should map agent permissions to business risk rather than using a single approval rule for every task. Reading a public document, preparing a draft or categorising an internal request may be low risk. Issuing a refund, changing a supplier bank account, altering a production schedule or approving a customer credit limit can have financial, legal or safety consequences. Each action should have a defined owner, limit, evidence requirement and recovery path. The organisation should be able to pause or redirect an agent without disabling an entire workflow.
Move from pilot approval to portfolio management
One of the most practical findings in the IBM release is that only 31% of all respondents considered clear thresholds for stopping AI initiatives critical to success. That is a warning about sunk-cost behaviour. Once a project has a senior sponsor, a vendor contract and a visible pilot, it can be difficult to stop even when the economics are weak. AI portfolios need the same discipline as capital expenditure: an explicit problem, a baseline, a budget, a review date and a decision rule.
Before a pilot begins, document the current cost and performance of the process. Include labour time, queue time, error rates, rework, customer impact, compliance effort and the cost of supporting the new system. Define what improvement would justify wider deployment and what failure would trigger redesign or closure. The baseline should be agreed by the business owner and finance, not created retrospectively by the project team. If the process is already changing for other reasons, record those changes so that AI is not credited for benefits it did not produce.
After launch, review results at fixed intervals. Scale when the process produces repeatable gains without unacceptable quality, privacy, security or customer risks. Redesign when the model is useful but the surrounding workflow is inefficient. Stop when the benefit does not justify the cost or when the risk cannot be controlled. The decision should be recorded, including who approved it and what evidence was used. This creates a portfolio that learns rather than a collection of pilots that never quite finish.
Redesign the workflow before buying more capability
IBM says high-impact respondents were more likely to embed AI directly into workflows, while many organisations still use AI to strengthen an existing model rather than change how work is done. That distinction is important. Adding a chatbot to an old approval chain may reduce one task while leaving the queue, duplicate data entry and unclear hand-offs untouched. The business can end up paying for a new tool without changing the total time required to serve a customer or complete a transaction.
Start with the work itself. Map the process from the first request to the final outcome. Identify repeated data capture, waiting points, exceptions, decisions that require judgement and controls that must remain. Then decide which steps can be automated, which should be assisted and which should stay human-led. The process owner should define the service standard and the exception path. Technology should fit that design rather than becoming the design by default.
Workflow redesign also changes workforce planning. If an agent completes routine preparation, employees may spend more time on exception handling, relationship management, analysis or quality assurance. That shift requires training, new role descriptions and updated performance measures. Productivity gains should not be assumed to become cash savings automatically. They may first appear as capacity that can be redeployed to reduce backlogs, improve service or support growth. Finance and operations should agree how those gains will be measured and where they will be reinvested.
A practical 90-day governance reset
In the first 30 days, build a single inventory of AI use cases, including informal tools and vendor-embedded features. Record the business process, owner, model or provider, data used, level of autonomy, users affected, cost and current status. Mark systems that can take action without a human click. Do not rely only on procurement records: shadow AI often appears in team workflows before it appears in the technology register. Ask business leaders to confirm the inventory and identify where a system is already influencing a customer, employee or financial decision.
During days 31 to 60, assign decision rights and risk tiers. For each material use case, name the business owner, technical owner, risk reviewer and executive sponsor. Define what the system may read, write, recommend and execute. Set thresholds for human approval, exceptions, monitoring and shutdown. Review contracts for data use, subcontractors, retention, incident notification and exit assistance. If a supplier cannot explain how an AI feature works within the service, treat that uncertainty as a procurement issue rather than assuming the vendor’s standard terms resolve it.
During days 61 to 90, review the first portfolio decisions. Compare pilot results with the baseline, identify bottlenecks and decide which initiatives should scale, change or stop. Publish a concise dashboard for leadership showing spend, process outcomes, adoption, incidents, exceptions and benefits realised. The purpose is not to create a large bureaucracy. It is to shorten the time between evidence and action while preserving the controls needed for trust.
Measure outcomes, not AI activity
Useful measures depend on the process, but a balanced scorecard should include business outcomes, operational quality, adoption and control. For customer service, measure resolution time, repeat contacts, customer satisfaction and escalation rates. For finance, measure cycle time, reconciliation exceptions, forecast accuracy and the number of manual adjustments. For software engineering, measure lead time, defects, rework, security findings and successful releases. For supply operations, measure forecast error, stock availability, waste and the time required to resolve exceptions.
Track the cost of the complete system, not just the model subscription. Include integration, data preparation, human review, change management, monitoring, security, vendor support and the cost of correcting errors. An initiative can look inexpensive at the licence level and still be costly when the organisation has to maintain parallel manual processes. Conversely, a system that appears expensive may create value if it removes a major bottleneck or supports more volume without sacrificing quality.
Also measure the speed of decisions. How long does it take to approve a change in scope? How quickly can the business redirect funding? How long does it take to pause an agent after an unexpected result? These are not technical vanity metrics. They reveal whether governance is helping the organisation respond to evidence or slowing it down.
What business leaders should take from the findings
The IBM study is a useful prompt, not a universal benchmark. It is based on executive responses, and IBM is a technology provider with an interest in enterprise AI adoption. The high-impact group comparisons are associations rather than proof of causation, and the study’s global sample should not be mistaken for a representative survey of every US company or every sector. Leaders should use the findings to ask better questions about their own operating model, not to copy a percentage or buy a platform.
The central question is whether the organisation can act on what it learns. If AI identifies a faster way to complete a process, can the process owner change the workflow? If the benefit is weak, can finance stop the spend? If an agent takes an unexpected action, can operations pause it and restore the previous state? If a supplier changes the model or data flow, does the company know who must approve the change? Clear answers are a stronger foundation than a broad AI ambition statement.
Enterprise AI is moving from experimentation into the operating fabric of companies. That makes decision rights, workflow ownership and investment discipline central to value creation. Organisations that connect those elements can move more confidently from pilots to repeatable outcomes. Those that leave authority fragmented may continue to deploy technology while struggling to convert it into better service, lower cost or growth.
Frequently asked questions
Does IBM’s study prove that giving AI leaders more budget authority causes better results?
No. The survey reports associations between organisational practices and outcomes reported by executives. It does not establish that budget authority alone causes better performance. Companies should test whether clearer decision rights improve their own speed, cost, quality and risk outcomes.
Does every company need a Chief AI Officer?
No. A dedicated role may be appropriate in a large or complex organisation, but smaller businesses can assign the coordinating responsibility to an existing executive. What matters is a clear remit, access to decision-makers and authority to coordinate business, technology, finance and risk owners.
What should a company do before allowing AI agents to execute business actions?
Define the permitted actions, data access, approval thresholds, exception handling, audit trail and accountable owner. Test the ability to pause or reverse actions, monitor performance and restore service. High-impact decisions should retain appropriate human oversight.



