In research announced on 17 November 2026, IBM says only 18% of surveyed enterprise AI leaders report full control over AI investment decisions. For business sponsors reviewing an AI pilot, this guide recommends naming an approver with sufficient budget authority, informed by delivery, finance and risk evidence. An AI job title alone does not determine who can authorize further spending.
The report was originally published on 16 November. IBM and Oxford Economics surveyed 1,667 enterprise AI leaders in April–June 2026, including chief AI officers and people in equivalent roles. Full control is narrower than influence or shared authority: the finding does not mean everyone else has no say.
IBM also reports that 29% of respondents say their organizations adjust decisions and resources quickly when AI performance evidence identifies opportunities to improve outcomes; 40% say discussion happens but action is delayed. These are reported responses to performance evidence, not cancellation rates for unsuccessful pilots.
This is vendor-sponsored, self-reported research. Its associations do not prove that centralizing budget authority or cancelling projects improves returns, and it cannot identify the right approver for an individual business.
Who can approve the next funding decision?
Start with existing spending authority. Name the business sponsor who can approve the proposed amount, record their limit and identify the person or committee that decides anything beyond it. If approval is shared, name the required approvers and who resolves disagreement. Shared authority need not be a problem; an unresolved approval route is the issue to fix.
Keep recommendations distinct from approval. Delivery staff and subject experts should explain whether the pilot completes useful work acceptably. Finance should check costs and the basis for claimed benefits. The operational and risk owners should specify when use must pause, who can act and who maintains the fallback process.
That separation adapts NIST’s guidance on responsibilities and delegated authority, which also recommends separating management of development from testing. It does not require a chief AI officer or finance chief to control every pilot. In a small business, one person may hold several roles; record each responsibility and seek a second review where feasible. IBM’s enterprise survey does not establish small-business results.
What evidence should reach the approver?
Ask for a short comparison against the intended task and the current way of doing it, including a non-AI alternative. NIST’s AI Risk Management Framework includes considering viable non-AI approaches and deciding whether development or deployment should proceed. It is risk guidance, not a formula for financial returns.
The evidence owner should state the review period, work completed, quality problems and human checking or rework required. Separate observed outcomes from estimates and unresolved questions. For a closer look at interpreting outcome claims, see how Amplitude checks whether AI-suggested product changes help. Keep that evidence attached to the funding request, not in a separate dashboard nobody has to act on.
For the next spending period, distinguish money already spent, unavoidable commitments and spending the decision can still avoid. Include subscriptions, usage charges, integration, supervision, support and exit work. Treat estimated staff time released as capacity, not automatically as a cash saving.
The UK government’s AI procurement guidance addresses whole-life costs, decision checkpoints and evaluation timeframes. Those are useful planning principles here, not procurement rules imposed on private businesses. The proposed decision is whether the next commitment is justified against feasible alternatives; past spending alone does not answer it.
When should you continue, revise or stop?
Agree outcomes, acceptable risks and review conditions before committing more money. None of these sources supplies a universal return-on-investment percentage, accuracy cutoff or pilot duration. Use criteria suited to the task, and do not quietly relax them after disappointing results.
Continue at the agreed scope. Evidence supports useful outcomes and the next period’s costs, risks remain acceptable and someone owns ongoing operation. Passing a pilot review is not automatic approval to scale.
Revise or extend narrowly. A specific fix or missing piece of evidence could change the decision. If continued use is acceptable, authorize a bounded step with a cost cap, named owner, evidence due and review date. Inconclusive evidence is neither success nor proof of failure; an extension should buy an answer, not simply more time.
Stop further funding or pause use. Agreed conditions are not met and no credible improvement path justifies another commitment, or an operational problem requires suspension. Record which action is needed: a budget decision and an immediate safety intervention are not interchangeable.
These are editorial decision criteria to adapt, not findings that IBM tested or a mandatory NIST checklist. If the review date arrives without sufficient evidence, explicitly approve a bounded extension or decline the next commitment; do not let the decision become an automatic renewal.
Put the decision in one record
Use the following as RohitAI’s proposed decision record, not an official IBM or NIST form. Record whether to continue, revise or stop, who approved it and what happens next.
Question and scope: Name the pilot, intended users, business task and current alternative. State the decision being requested.
Evidence and limits: Attach results for the stated period, previously agreed outcomes and risk limits, and the uncertainties that could change the decision.
Next-period commitment: Set out expected benefits and costs with assumptions. Separate past expenditure, unavoidable commitments and avoidable spending.
Authority and review: Name the evidence reviewers, authorized approver, spending limit and escalation route. Record unresolved disagreements.
Decision and follow-up: Record the approved action, reason, amount or cap, scope restrictions and approver’s sign-off. Assign follow-up work and the next review date, with triggers for an earlier review.
Pause or exit arrangements: Assign service handover, access shutdown, data and record handling, supplier obligations and any restart conditions to named owners.
What happens after a stop decision?
Stopping new expenditure, stopping execution and ending a supplier contract are separate actions. NIST’s intervention guidance calls for assigned responsibility to disengage systems whose performance or outcomes conflict with intended use. The UK Government AI Playbook separately emphasizes operating responsibilities, escalation and releases that can be withdrawn.
For a managed stop, agree who takes over the work, checks outstanding tasks and removes unnecessary access. Ask the contract owner what commitments remain and what data or records must be retained or returned; the UK procurement guidance explicitly addresses end-of-contract roles. A funding decision does not itself complete those steps.
For agent-based pilots, the related guide to CrowdStrike’s AI kill-switch coverage limits examines why a technical cutoff needs its own scope check. Keep that operational review separate from permission to spend.
Before the next renewal, bring the evidence owners and authorized approver together around the same record. Leave with a decision, a named person responsible for carrying it out and a date for checking what happened.
