AI Adoption Should Be Measured in Changed Workflows

Analysis | Commercial Bank of Dubai’s use of Copilot illustrates why measuring AI adoption starts with what a team does differently and ends with what that change delivers.

In a June 2025 customer story, Microsoft reported that Commercial Bank of Dubai saved 39,000 hours a year through its use of Microsoft 365 Copilot. The same story described smaller changes that readers can examine on their own terms. The bank’s internal audit team used the tool to turn information into structured risk matrices. Analysts in strategy and operations used it to organise material for data management policies, with final oversight staying with the analysts. The headline gives the bank’s reported scale of savings. The examples tell us where work changed and which people still had responsibility for it. Microsoft’s Commercial Bank of Dubai customer story

An adoption measure earns its place when it explains something about the work. Licences can show who has access. Usage can show that staff found reasons to open a tool. A workflow measure asks what the team now does differently, whether the change has held up and whether it helped the people who depend on the output. The strongest case ties those observations together without treating them as interchangeable.

Read the bank’s figures at the scale they describe

CBD reported holding a Copilot Promptathon with Microsoft Innovation Hub in January 2025. Microsoft’s later account describes several such events and a library of prompts that teams could share. This is evidence of how the bank tried to help employees find repeat uses. A shared prompt can reduce the effort needed to start a task or make a method easier to teach. How widely it was used and whether its outputs were consistently useful require separate observations. Commercial Bank of Dubai’s January 2025 Promptathon release

The Microsoft story also reports that the IT project management office generated 3,200 meeting summaries and action points in the month it describes. This is an activity count with a stated time window. A team looking for the effect on project delivery would want to know how many outputs were useful, whether owners accepted the assigned actions and whether those actions were completed sooner or with fewer clarifications. Some meetings might have had detailed notes already. Others might have had none. A count of outputs leaves those starting points combined.

The case story describes time savings for some tasks but does not provide a full reconciliation of the annual total to measured cases, task mix, rework and the baseline. The 39,000-hour figure is best read as a reported result from the bank and its supplier. It gives a reader a reason to examine which tasks contributed and what happened to the time. Converting it directly into payroll savings, or assuming that a business with similar staffing could repeat it, would go beyond the evidence.

Name the handoff that changes

For a UAE project office adopting meeting summaries, the changed workflow might be that team members stop writing notes individually and use one approved action list. The tool can draft that list. A meeting owner still needs to resolve ambiguous wording, identify who accepted responsibility and decide which decisions belong in the official record. The adoption evidence is that people receive and use a common record as part of ordinary work. The performance evidence is whether fewer actions need clarification or are completed late.

Write down the handoff in terms both teams recognise. Identify who provides the input, what a usable output contains, who accepts it and where errors are sent. If a department uses AI to summarise supplier correspondence, decide whether the receiving team expects a summary or a verified record that can be acted on. A summary can be produced quickly while the buyer still has to reread the correspondence to verify delivery dates. The buyer’s remaining work belongs in the assessment.

This also helps identify where learning happened. One employee may get good at steering the tool through a tricky case. A team has adopted a repeatable method when colleagues can handle comparable cases without that employee sitting beside them. The UAE Government’s Instagram carousel announcing the UAE Skills Platform provides separate context for the importance of workforce skills. For an individual employer, the relevant evidence is whether the people doing the job can apply and explain the new procedure, including when they should set it aside.

Follow the time after the draft

Consider a hypothetical Dubai procurement team processing 240 comparable requests a month. In the old workflow the buyer spends 12 minutes on each initial review and the approver spends another 4 minutes. That is 16 minutes a request, or 64 hours in total. In the new workflow the buyer spends 7 minutes, the approver spends 6 and corrections add an average of 2 minutes across all requests. That is 15 minutes a request, or 60 hours. The buyer’s first step saves 20 hours, while total handling falls by 4 hours. The figures are illustrative.

Both figures can be correct because they answer different questions. The buyer has capacity to start more requests, but the approver now has more to do. If there is one approver and requests arrive in clusters, the time customers wait could rise even as total handling falls. Measure elapsed time to an approved decision alongside active staff time. Look at unusually long cases as well as the typical one. A median can improve while difficult requests spend longer waiting for attention.

The calculation can also miss a change in quality. If the new workflow flags a missing document before a request reaches a supplier, extra review time may be worthwhile. If reviewers have to reconstruct the same facts repeatedly because an output omits its source, the team has an opportunity to improve the handoff. Compare the result that the work is supposed to deliver, including errors discovered later, before deciding that additional minutes are waste.

Then check what the freed time did. A team might work through a backlog, review requests it previously ignored or respond sooner at peak times. It could also save small amounts of time across many people without creating a block that can be reassigned. Both can be useful. A claim about additional output needs an observable result such as more cases completed to the same standard, shorter customer waits or time specifically assigned to another named activity.

Check whether the change lasts without a campaign

A burst of interest during training makes adoption visible but can overstate how much work has changed. Take another look after the launch activity ends and the workflow goes through a busy period. Ask whether the same case types still go through it, whether new staff are taught how to use it and whether people know when to fall back to the old process. A workflow that functions only during a demonstration or with an enthusiastic colleague nearby is still dependent on those conditions.

Keep exceptional cases in view. If a team sends straightforward requests through the tool and keeps ambiguous ones in email, the assisted queue may look excellent because it has less difficult work. Compare how each kind of request is handled and record whether the split was intentional. Where case volumes, staffing or policy changed during the comparison, report the change and make a narrower claim. A staged rollout across similar teams can help separate the effect of the new workflow from a busy month or an unrelated improvement.

Assign one person responsibility for the result after launch. That person needs the authority to change the procedure when approvals pile up, reviewers disagree or an update changes what the tool returns. Keep the approved procedure, evidence about exceptions and the reason for material changes together. If the company later compares two departments, it can then tell whether they used the same workflow or just had access to the same software.

If the data are sparse, begin with a small sample traced from input to outcome. For a meeting record, compare the draft with the accepted list and check a week later whether the assigned actions entered the project system. For an audit matrix, ask a second reviewer to compare selected entries with the underlying material and note any missing references. These samples do not support a bank-wide estimate. They can show where logging is missing and which review steps are doing useful work. A management team can then decide which observations to collect routinely before using the result to set targets or renew spending.

For the next management review, describe a changed workflow in one sentence that names the work, its owner and the outcome. Then attach the figures that show frequency, quality and consequences. A bank’s announcement can establish that staff used AI to prepare risk matrices or capture meeting actions. A durable operating result needs evidence that people used the output, that the entire job improved and that someone can maintain the new way of working when conditions change.