The challenge
Generating an answer is only one part of engineering work. A useful result also needs to fit the requirement, respect the technical environment and stand up to review. That makes AI adoption a workflow problem as much as a model-selection problem.
My focus is on how AI assistance can support professional services in practice: helping teams work with technical context, carry out bounded tasks and evaluate the output before it becomes part of a customer solution.
My role
I help shape the technical AI direction for services, including the tooling and engineering practices around adoption. The existing work spans agentic frameworks, multi-agent orchestration and approaches to measuring AI-driven productivity.
The connection between those areas matters: a promising tool needs a place in the delivery process, and an adoption decision needs evidence beyond the quality of a demonstration.
The approach
Choose a bounded task
Start with a task whose output can be checked: explaining unfamiliar code, preparing a first draft of documentation, proposing tests or assisting with a defined implementation. Make the expected output and the review criteria clear.
Supply the right context
Relevant requirements, repository guidance, coding conventions and product constraints help make the task specific. Retrieval and reusable instructions are useful when they bring the information needed for that task into reach.
Keep review visible
AI output remains a proposal until it has been checked. A developer or reviewer needs to assess correctness, fit and potential side effects, supported by the normal testing and source-control process.
Use orchestration where it earns its place
Agent workflows can split a larger piece of work into focused tasks. Clear boundaries and a defined integration step matter; additional agents also introduce coordination and review overhead.
The important trade-offs
- Speed and quality. Time saved during generation can be lost during correction. Review effort belongs in the evaluation.
- Capability and cost. Match the tool to the task, taking account of response quality, context needs and running cost.
- Autonomy and accountability. Define which actions an assistant can take and which decisions require a person.
- Experimentation and repeatability. Capture useful patterns so an individual success can become something others can apply.
What success should look like
I’m interested in the full path from starting a task to accepting the result. Useful signals include completion time, review effort, defects and whether people continue to use the workflow after the initial trial.
Comparing similar tasks with and without assistance helps distinguish a convincing first impression from an improvement that lasts. These are evaluation criteria, not published performance results.