A team can have plenty of ideas for agents without knowing which would improve its work. The most visible task isn't always the source of delay, and the easiest demonstration may have little in common with what people struggle with each week. Choosing where to begin is a delivery decision before it's a tooling decision.
There are several legitimate results: less repetitive work, a shorter wait for customers, better decisions, or the ability to attempt work that was previously too expensive. Starting with one of those outcomes helps the team compare AI with other ways to achieve it.
Find the problem behind the task
A recent piece of work can reveal where the effort actually went. The people involved may remember reconstructing a decision, searching for the right project, waiting for access, or repairing an incomplete change. Those explanations often suggest a different opportunity from simply accelerating code generation.
Suppose maintenance changes take weeks to reach production. An agent that prepares them in minutes may help, but only if preparation accounts for much of that delay. If most of the time is spent waiting for a release window, the next improvement may concern release practices. If reviewers repeatedly discover missing context, better preparation could make both human and agent work easier.
It helps to describe the problem without naming a solution first. “Reduce the effort needed to understand an unfamiliar service” leaves room for assisted investigation, better documentation, or both. “Deploy an investigation agent” makes adoption look like success even if understanding doesn't improve.
Compare assistance, delegation, and conventional automation
Different approaches suit different sources of difficulty. A repeatable transformation with clear rules may be cheaper and easier to maintain as ordinary automation. Work that involves interpreting unfamiliar information may benefit from an assistant. Delegation becomes more attractive when the intended result, available context, and checks are clear enough for work to proceed without continuous direction.
Three questions help a team judge how much to delegate:
- Does the agent have enough context to make progress without someone repeatedly filling in missing information?
- Can checks assess the result without a person having to inspect every detail?
- If the agent makes a mistake, will those checks catch it before it affects other work, or does that still depend on a person noticing?
The answers can differ within one task. A person might work through an uncertain design with an assistant, then delegate an implementation covered by well-understood checks. If correctness still depends on a reviewer tracing every change, that review needs time in the plan. The questions help choose where people stay involved; they don't rank teams or make autonomy the goal.
Managed GitLab capabilities are worth considering before creating a custom workflow. The comparison is whether they handle the team's task, not whether they reproduce every detail of an existing process. Customization can close a real gap, but it also creates instructions, integrations, and behavior that someone must maintain. The ongoing effort is part of the choice.
Look for the conditions that let the approach work
Missing context can be a reason to improve the environment rather than reject the opportunity. A team might first connect the design decision to the work item or document a working test command. Conversely, giving an agent more access won't resolve disagreement about what should be built.
The people expected to use or review the capability need time to learn it and a route to help. A successful demonstration by its creator doesn't tell you whether another engineer can direct the same work, recognize a mistake, and recover. Including those people early reveals costs that a technical prototype can miss.
Learn from representative work
A first introduction should include enough variation to test the idea behind the investment. Routine successes explain where the approach fits; ambiguous requests, incomplete context, and unsuccessful attempts explain what support it needs. Choosing only the easiest cases can create an expectation the broader rollout won't meet.
Before beginning, the team can agree on the improvement it hopes to see and the effects it will watch alongside it. Less implementation effort is encouraging if review, quality, and delivery time remain acceptable. A lower effort estimate is less persuasive when the same work repeatedly returns for correction.
The comparison doesn't have to settle every question about future adoption. It should be strong enough to support the next decision: continue with similar work, change the approach, address a missing prerequisite, or stop. Keeping the normal path available gives people a way to complete their work while the team learns.
Expand toward value, not a target for autonomy
A successful use can reveal adjacent opportunities, but it doesn't establish that every project has the same context or constraints. Expansion is a chance to ask which conditions made the first result possible and whether they remain true elsewhere.
Some opportunities will change the economics of the work rather than make an existing task faster. A team might be able to examine neglected services or explore more design alternatives. That can be valuable, provided someone decides how the resulting knowledge or changes will be used.
An adoption plan can therefore contain several working modes and different stopping points. The organization benefits when people can choose an effective approach and explain its result, without pressure to replace a successful assisted workflow with unattended execution.
Further Reading
- Evaluating business value and the cost of delivery helps connect the initial choice to a comparison and investment decision.
- Operating and evolving the team's AI capabilities covers the support and maintenance involved after adoption.
- A value-stream improvement cycle helps investigate delivery delays when the cause isn't yet clear.