AI Is Accelerating the Work. Your Bottleneck May Be Somewhere Else

Follow the work through to its outcome

The time an AI agent spends working tells leaders how much the tool is running. To judge its value, they also need to know whether the work solved the right problem, passed review, reached its intended users, and produced a useful result.

The distinction matters because AI can produce drafts and analyses much faster than an organization can put them to use. Four agents running in parallel may produce four analyses before a manager can evaluate any of them. A development team may create more code than reviewers can safely inspect. A research group may run more experiments than experts can interpret. A marketing team may generate more campaigns than the organization can approve, launch, and learn from.

These are illustrative business scenarios, not findings from the cited studies. If completed work arrives faster than the next stage can process it, queues will grow. That can delay feedback and put reviewers under pressure.

Leaders need to know how much work passes through the entire process, meets its quality requirements, and reaches someone who can use it.

The slowest step can limit the benefit

Mert Demirer, Leon Musolff, and Liyuan Yang studied more than 100,000 developers on GitHub, a platform for managing software development. Their June 21 article reports cumulative increases across successive generations of AI tools of roughly 180 percent in saved code changes, called commits; 50 percent in projects; and 30 percent in released software versions. Each percentage measures a different type of output. [2]

The researchers compared developers who adopted AI with similar developers observed a year earlier. They did not randomly assign people to use AI, so the estimates depend on how well those comparisons account for other changes. The findings should not be treated as a forecast for every business. A 30 percent increase in releases would still be substantial, even though it is much smaller than the increase in coding activity.

The pattern suggests that other steps may limit delivery, although it does not establish every cause. In Wharton's September 8 discussion, coauthor Leon Musolff points to review, integration, distribution, and work with users as possible constraints. Wharton's article discusses the same research and is not independent confirmation. [3]

For a business leader, the next step is to find where work gets held up before buying tools that produce more of it.

Account for review and the work AI creates

A February 17 research note from METR explains why time saved on individual tasks may not translate into the same improvement across a workday. It examines 5,305 records of coding-agent sessions from seven technical staff members. People tend to use AI where it helps and may take on extra work because it becomes easier. Estimates of how long that work would have taken without AI can also be unreliable. The small sample limits how widely the findings apply. [4]

For leaders, this means separating necessary work completed more efficiently from additional work made possible by lower costs. Both may create value, but they need different business cases. A tool built in an afternoon is useful only if its benefits justify the time needed to verify, maintain, and use it.

Plan who will check the additional work and how much time they will need. For software, this may involve testing whether a change works with existing systems. For an analysis, it may mean checking source data, assumptions, and calculations. The appropriate level of human review depends on the consequences of an error.

An agent can help inspect its own output, but it may repeat the same mistaken assumptions. For work that affects important decisions, define what an acceptable result must contain and check it against original records, tests, or a qualified reviewer's judgment. A different AI model may still make the same mistake.

Build a workflow scorecard

A simple workflow scorecard can show where time is saved, where work piles up, and whether the business benefits. Track five things:

• Work started: How many tasks or ideas are being pursued because AI makes them easier or cheaper?
• Work produced: How many drafts, code changes, recommendations, or test results are completed?
• Review effort: How much time and money are needed to check the work and correct mistakes?
• Work put to use: How much completed work is approved, delivered, and actually used?
• Business results: What changed in revenue, cost, delivery time, quality, risk, customer experience, or the organization's mission?

Keep these measures separate. A team can produce more work while spending even more time correcting it. It can also deliver more tools or reports that few people use.

Consider a hypothetical proposal team. AI increases drafts from 20 to 60 per week, but reviewers can approve only 20. If all 60 are submitted each week, 40 additional proposals accumulate in the queue. Buying more drafting capacity would not raise approvals. The team needs to improve review, narrow its priorities, or reduce avoidable rework while maintaining quality.

Track the cost of each completed and accepted piece of work, how long it takes, how often it passes review without corrections, and whether people use it. Include tool charges, staff time, and the cost of fixing errors. Compare work of similar difficulty: ten simple requests are not equivalent to ten complex cases.

Invest where the work gets stuck

When one stage becomes faster, check whether another stage now limits delivery. A useful pilot measures both the accelerated activity and the next handoff.

Five actions can help:

• Map the whole process before buying more licenses. Follow work from the initial request through drafting, checking, approval, delivery, and use.
• Measure results alongside activity. Track drafts and experiments together with completed decisions, delivered work, actual use, and business results.
• Give reviewers the time and tools they need. Use tests, clear checklists, source records, and qualified people to check important work.
• Limit work in progress. Do not allow cheap generation to flood a process that cannot absorb the output. Prioritize tasks with a clear owner and path to use.
• Use savings to improve the slowest step. Faster coding may require better testing and delivery. Faster analysis may require quicker decisions and follow-through. More experiments may require more time to interpret and confirm results.

The AI budget should cover the people, checks, training, and changes to existing systems needed to put the additional work to use. Those costs belong in the business case alongside the price of the tools.

Start with one process and measure the result

OpenAI's experience illustrates one organization making extensive use of agents. The opportunity for other businesses depends on their own work, constraints, and ability to turn additional output into useful results.

Start with one process, assign someone to own the result, and agree on what counts as completed work. Record its current delivery time, quality, cost, and usage before introducing AI. Then compare those measures after the change and use what you learn to decide whether to expand, adjust, or stop.

If any link remains slow or weak, more agent activity may simply create a larger queue in front of it.

AI can accelerate the work. Leadership still has to accelerate the system that turns work into value.

Questions for executives

1. Are your AI reports showing more activity, better business results, or both?
2. Where does AI-generated work wait longest before someone can use it?
3. If AI doubled the amount of work produced next quarter, which team or step would struggle to keep up?