Faster implementation can shorten a customer's wait, make overdue maintenance affordable, or give a team more time to explore a difficult problem. It can also move effort into review and correction without improving the result. The investment is worthwhile when it helps the organization achieve something it values at a cost it can justify.
A return-on-investment discussion starts with that improvement and the resources needed to achieve it. Usage reporting can help explain what the work cost; understanding the return means examining whether that work achieved the improvement your organization wanted. A comparison with the team's starting point helps separate that result from expectations about what AI might do.
Decide which improvement would justify the investment
The starting point is a business or delivery problem, not a target for AI adoption. A team might want customer fixes to reach production sooner, recurring maintenance to consume less attention, or an expensive modernization effort to become practical. Those goals lead to different choices about what to measure and what a worthwhile result would look like.
The people affected by the problem should help describe that result. A developer's shorter implementation time may not improve a customer's wait if the change then spends longer in review. Conversely, an agent-assisted investigation that rules out an unnecessary migration can create value without producing code.
An improvement also needs someone who can act on it. If released capacity is the expected benefit, what work will receive that capacity? If faster delivery is the goal, what happens when the next stage becomes the constraint? Answering those questions gives the investment a path to value beyond buying access to a tool.
Keep the different kinds of benefit separate
Some benefits can be valued financially; others first appear in capacity, quality, or the ability to pursue new work. Keeping them distinct makes the conclusion easier to trust.
| Benefit | What would support the claim | What the claim doesn't establish by itself |
|---|---|---|
| Less effort on recurring work | Comparable work takes less total human effort, including review and correction. | The organization's expenditure fell. |
| More delivery capacity | The team completes additional valuable work with similar staffing and acceptable quality. | Every reported hour saved became productive capacity. |
| Better quality or service | Fewer recurring defects, faster recovery, or a better customer result in the work examined. | The full change was caused by AI. |
| Work becomes affordable | A previously deferred opportunity can now be pursued within an acceptable cost and time. | The anticipated business benefit has already been realized. |
Financial ROI compares the benefit left after costs with the total investment over the same period. Turning a capacity improvement into that calculation requires an explanation of how the capacity creates financial value or avoids expenditure. Finance and the relevant business owner can help with that interpretation. Multiplying reported time savings by salary isn't enough to establish cash savings.
Related benefits can also overlap. If faster work allows additional delivery, counting both the full value of saved time and the full value of that delivery may count the same improvement twice. Showing the assumptions separately is more informative than hiding them inside a single impressive percentage.
Establish a comparison that reflects the work
Recent work can provide a starting point: its elapsed time, human effort, corrections, costs, and eventual result. The period should include the variation that matters to the decision. Frequent maintenance tasks and infrequent releases won't need the same observation period.
When historical records are incomplete, the team can establish a baseline prospectively or compare similar work during an initial rollout. Missing information should remain visible rather than becoming a zero. There is still value in an early assessment if its limitations are clear.
The comparison needs to account for the work being selected. Agents may receive easier tasks because teams are beginning cautiously, or harder tasks because people seek help where they are stuck. Comparing all AI-associated changes with all other changes can confuse that selection with the effect of AI.
Work type, complexity, team, and release path can help identify a more comparable set. Changes to staffing, CI performance, or review practices should be considered alongside the AI introduction. A trend can show that the overall approach improved while leaving uncertainty about how much each part contributed.
Follow the work beyond the agent's completion
A single improvement may involve several sessions, failed attempts, human corrections, and a later release. Assessing only the successful execution leaves much of the effort out. Assessing every session as a separate delivered result can count the same work several times.
The work item, investigation, or delivered change can provide a way to bring those attempts together. The appropriate endpoint depends on the purpose. An accepted explanation may complete an investigation; a code change intended to resolve a customer problem needs follow-through to whether that problem was resolved.
Including open, abandoned, and rejected work helps keep the comparison honest. Otherwise, a report may improve simply because difficult cases haven't finished. Looking at actual work and talking with the people involved can explain why an apparent success still required substantial intervention.
For implementation work, cost per accepted change can help compare the effort needed to produce a result the team is willing to use. The cost includes preparation, checks, review, correction, and unsuccessful attempts, not just the final agent run. Comparing similar kinds of change keeps a small documentation fix from being treated as equivalent to a complex migration.
For comparable work, a lower cost per accepted change indicates improved delivery efficiency, not the return on investment by itself. A fix still needs to resolve the problem it was intended to address. An investigation may create value by avoiding an unnecessary change altogether. The question remains what the organization gained, not how many changes it accepted.
Account for the full cost of the approach
The investment includes tool and model charges, build and test resources, preparation, context maintenance, evaluation, training, review, correction, integrations, and support. Some costs are initial investments; others recur as the approach expands or changes.
For GitLab Duo Agent Platform, GitLab Credits describe product consumption under GitLab's billing rules. They aren't a universal token conversion or a measure of value. Charges from a model provider and the cost of running supporting systems also contribute to the investment. A rise in consumption means more resources were used; whether that was worthwhile depends on what the work achieved.
A comparison between approaches should include unsuccessful attempts and the work required to accept the result. A lower model bill can be outweighed by additional correction, just as a more capable model can be unnecessary for a routine task. Representative work provides a better basis for that choice than the price of an isolated request.
An investment decision can include a spending limit and a point at which someone reconsiders the approach. An alert gives that person information; an enforced limit stops particular activity. Knowing which is available, and what happens to work already running, helps the team avoid spending beyond its decision while waiting to assess the result. Staying within a budget doesn't establish that the work was worthwhile.
Decide what to continue, change, or stop
Delivery records and conversations with the people affected help explain what changed. A faster response to customer problems may justify continued investment; more completed tasks with no improvement in the customer experience may not. Knowing where the money went helps assess the cost, but doesn't answer that question on its own.
At a review point, the team can compare the intended improvement with the actual outcome, total cost, quality, and experience of the people involved. Mixed results suggest a more specific next action. Faster implementation with a growing review queue may call for coordination changes. A costly approach that produces better outcomes may still be worthwhile. An approach that creates more work than it removes may deserve to be narrowed or stopped.
The decision should include what the organization will do with the benefit and what would prompt another review. Measures are there to support that decision, not to rank people by their AI activity. As usage grows or the technology changes, the team can revisit the economics without assuming either that the original result lasts forever or that every new capability requires starting the assessment from nothing.
Further Reading
- Choosing measures that guide improvement helps select measures that can change a decision, establish a starting point, and recognize unwanted effects.
- Reliable DORA and value-stream data helps check what delivery records actually measure before using them in a comparison.
- GitLab Credits and usage billing explains the consumption rules to use when interpreting GitLab charges.
- Coordinating human and agent work across delivery addresses handoffs and capacity when faster execution moves the delay elsewhere.