Which workflow changed, and what did it earn?
Imagine a CFO looking at AI spend. There are licences, pilots, platform bills and a growing team. Employees can show faster drafts and better answers. The finance question is still incomplete: which operating outcome changed, and what did the complete change cost?
That is the enterprise impact I want to examine. Adoption tells us something about use. It does not establish a financial return. Time saved can be valuable without reducing an expense, and a production interface can exist before anyone has shown that the process works better.
The investment decision becomes clearer when the business separates three questions: is there a useful completed outcome, does value survive the full cost, and what evidence justifies scaling?
Read market evidence as a guide, not your business case
McKinsey’s November 2025 survey reported regular AI use in at least one function among 88 percent of respondents’ organisations, while 39 percent attributed enterprise EBIT impact to AI. About 6 percent met its demanding high-performer definition. These categories do not make everyone outside the last group unsuccessful. They distinguish use, reported financial impact and a particular high-performance threshold.
The report associates stronger outcomes with workflow redesign. That is a useful direction for investigation, not proof that a prescribed redesign causes a return. It is also a dated survey, not an audited census of the current market.
Widely repeated pilot-failure percentages, vendor ROI multiples and executive surveys often measure different things. A disclosed-initiative study and a company survey cannot be treated as the outer bounds of a single market success rate. I would not use that apparent precision to approve or reject a local investment.
Use outside evidence to choose questions and plausible mechanisms. Then ask the owner of the actual workflow to show the baseline, intended change and result.
Decision one: fund a completed outcome
The strongest early candidate may look ordinary: retrieving the right internal reference, preparing a service response or checking a recurring document. The important property is that the business can tell what a useful finish looks like and who receives it.
OpenAI’s Morgan Stanley case describes expert evaluations for retrieval and summarisation, followed by meeting-note preparation with adviser review. It reports active use by more than 98 percent of adviser teams. This is vendor-published adoption and implementation evidence, not a disclosed firm-wide ROI calculation.
The useful lesson is the connection between a defined task, evaluation and operating handover. It does not require a claim that automation must replace the adviser or that every company will achieve the same economics.
For an industrial example, consider a hypothetical quotation assistant. It can draft faster while engineering and operations still hold the case. A credible outcome might be a usable quotation reaching the customer sooner with the applicable configuration, terms and delivery approval intact. That is different from counting drafts.
Ask which constraint changes and who is empowered to change it. If capacity is released, choose its destination: more demand served, an avoidable expense reduced, better quality or a sustainable workload. Each has value; each needs evidence suited to the claim. The companion essay Your AI Saved an Hour. Who Gets It? follows that allocation decision.
A pilot can also be funded for learning. State what uncertainty it will resolve and which decision follows. Learning is a legitimate outcome when it can be inspected; it should not be silently reported as realised profit.
Decision two: compare full delivered cost
A model quote is one cost line. The delivered process may also require integration, document ingestion, permissions, retrieval, evaluation, monitoring, support, training and human review. Those costs can dominate a modest internal workload even when inference is cheap.
Compare designs that meet the same outcome, quality and operating conditions. A seat licence and a token-metered API may include different services and require different internal work. Comparing their headline prices alone can make an expensive implementation look like the cheapest option.
The same applies to model tiers. Lower token prices do not establish better value if a design needs repeated attempts, longer context or more checking to meet the standard. A more expensive model may be economical on a hard task; it may be wasteful on a routine one. Test the intended workload rather than assume either result.
Available throughput quotas describe capacity. Actual requests, token consumption, cache behaviour and operating utilisation describe use. A published limit is not evidence that an enterprise consumes that amount. If a process needs steady throughput or low latency, make capacity planning a separate design question.
In the quotation example, compare the cost per usable case, including review, repairs and unresolved exceptions. Include the transition work and ongoing ownership. Avoid counting the same freed hour both as cost saved and as revenue-generating capacity.
Then test how the economics change with volume, source changes and exceptions. A favourable pilot mix may exclude the cases that consume engineering time. A quiet month may conceal the support load at quarter end. The business case should make those dependencies visible rather than bury them in an average.
Decision three: scale on evidence that survives the change
Scaling is a permission, volume and operating-capacity decision. More use can expose different cases, more dependencies and larger consequences. A successful narrow test is a reason to design the next step, not an automatic authority to roll out everywhere.
Compare similar cases with the previous process where possible. Use holdouts or staged rollout when feasible, and document changes in demand, staffing, rules and source quality. If AI arrived alongside process repair, report the combined improvement honestly rather than assigning the entire gain to the model.
Quality should travel with the outcome measure. A support system that shortens handling time but increases repeat contact has changed where the work occurs. A coding tool that accelerates submission but adds downstream defects has not established the same value as faster accepted work. These are illustrative mechanisms, not reported results from a new study.
Controls need operating owners. Someone maintains authoritative sources, evaluation cases, permissions and recovery. Someone can intervene when the process fails. The right organisation need not be identical for every company, but the responsibilities cannot be left implicit because the vendor manages the model.
Cost controls such as output limits, caching, batching and routing can be useful where they preserve the required outcome. Track their actual effect. Reducing generation cost while increasing reviewer time is a tradeoff, not an unqualified saving.
The funding conversation
The AI Opportunity Map can help compare a workflow problem with process repair, integration or rules. The AI ROI Calculator can make the economic assumptions explicit. Their outputs are scenarios to test, not substitutes for observed results.
Bring the three largest AI line items to the next investment review. For each, name the workflow owner, the baseline, the principal failure mode and the cost guardrail. If a field is missing, assign the work needed to close it and define the funding decision it affects. A missing baseline is a measurement gap; it does not make a live system fictional.
Some initiatives should remain inexpensive assistance. Some should continue as bounded learning. Some may justify a change to the complete process, and some should stop when their quality or economics fail. The distinctions are more useful than a portfolio sorted by how advanced the technology sounds.
If you bring a workflow to a conversation, start with its volume, owner, current outcome and the decision you are trying to make. A transformation proposal becomes reviewable when people can explain the work, the full cost and the evidence that would change their minds.