Related Service
AX Consulting and HR Consulting
HCG Consulting provides AX Consulting that supports HR AI transformation and HR Consulting that designs performance, organization, and culture systems.
Solve Complex HR Challenges with HCG
Talk to our experts
Insights
The success of AI transformation depends on how work and decisions are designed, not just the tools deployed.
“Complete this task, including the criteria you will use to verify that it has been done correctly.”
Adding this instruction to an AI agent changes how work gets done. Instead of watching and checking every output, a person can let the agent execute, test its work against defined criteria, and iterate. But real change takes more than a better prompt. Organizations must design what the agent can consult, what standards it must meet, which decisions it can make, and when it must hand a decision to a person. Adding AI to existing workflows and becoming an organization that works effectively with AI are very different undertakings.
AWS compared the effects of AI tools and changes in working practices on delivery productivity across more than 50 development teams in Amazon Stores. The study examined ordinary teams doing their existing work, rather than a specially selected group of engineers or unusually favorable projects.
The findings were striking. The 25 teams that adopted new AI tools and changed how they worked achieved a median 4.5× productivity improvement, with some recording gains of more than 10×. Teams that simply added AI tools to their existing practices did not see comparable results. The difference suggests that the decisive variable lay beyond the technology itself.
As AWS puts it, the issue is not necessarily using the wrong tools. It is using the right tools within the wrong ways of working.
The strongest teams treated AI as more than a coding assistant. They changed their development environment and practices so agents could carry out work with greater independence.
They converted developers’ tacit knowledge and team conventions into explicit resources such as skill files, coding standards, test criteria, and guides to navigating the codebase. This let them accumulate organizational knowledge and decision rules so agents would not repeat the same mistakes. They also removed outdated rules as models and working conditions changed, keeping the context relevant.
Even the high-performing teams did not become faster immediately. They first improved error messages so agents could diagnose failures, connected the necessary tools and systems, and reorganized code and documentation into forms AI could use. AWS notes that teams can slow down during the first two weeks. Those that invested in the environment then saw gains compound; teams expecting instant results tended to revert to their old practices.
Watching an agent work and issuing corrections after each result limits its independence. High-performing teams specified the conditions for completion alongside the task itself.
The instruction defines what “done” means, not merely what the agent should do.
Generating a large amount of code against unclear requirements and revising it later can add cost. Strong teams first agreed on a specification covering the objective, requirements, constraints, exceptions, and acceptance criteria. Clarifying a sentence in a specification is easier than revising multiple code paths and systems after implementation. The faster AI produces output, the more valuable this early agreement becomes.
To work independently, an agent must be able to fail quickly and correct itself. Strong teams built local integration testing and automated guardrails that did not depend on connecting to external systems. Agents could then use test results to revise and verify their work without asking a person at every failure.
These habits point to three operating foundations, rather than five isolated productivity tips.
Structured work context: Turn the knowledge, rules, policies, and decision criteria AI needs into clear documentation and data.
Redesigned execution and verification: Build into processes and systems what AI may do, how its output will be checked, and how it should respond to failure.
Clear decision boundaries: Define what AI may handle independently, which judgments require a person, who has final approval, and how exceptions are managed.
The high-performing teams encoded acceptance decisions in specifications and tests that agents could apply. People still decided what to build, how far to automate, and what risks to accept. AI transformation requires a deliberate division of execution and judgment between people and AI.
Faster code generation brings a new challenge. Research from Smartsheet cited by the Stack Overflow Blog found a 55% year-over-year increase in automation intensity among enterprise users and a 46% increase in overall work activity. People edited 80% of AI-generated content before it was finalized.
AI can multiply output without increasing the time people have to evaluate it. The working day stays the same while the density of tasks and decisions grows. Even if implementation that once took nine to twelve months takes one or two, an organization will not move much faster if deciding what to build and securing stakeholder approval still takes two months. As execution becomes less costly, decisions become the bottleneck.
A development team’s advantage increasingly depends on distinguishing reversible decisions from difficult-to-reverse ones, then making the judgments that matter accurately and promptly.
This pattern extends beyond software development. According to Gartner, 88% of employees with access to employer-provided AI also use personal AI tools at work. In a separate survey of executives, only 27% of organizations reported having a comprehensive AI strategy, and just 20% considered their workforce ready to use AI. Access to tools should not be confused with organizational readiness.
AI can analyze large volumes of applications and recommend candidates in seconds. But decisions will not accelerate unless the organization defines when AI may screen someone out, when a person must review a case, and who can override a recommendation. Poorly designed processes may amplify bias or force people to recheck everything, creating duplicate work.
AI can synthesize a year of performance data and feedback into draft review comments. Yet a new bottleneck emerges if it is unclear whether that output is reference material or an accepted draft, and whether a line manager or HR can adjust the final rating. Drafting may get faster while review and calibration meetings multiply.
AI can model pay bands and increase proposals using market data, internal equity, and performance. Without defined tolerances, exception rules, principles for handling budget overruns, and final approval authority, it generates more options without speeding decisions. Across all three examples, the unresolved question is who reviews and approves AI outputs, when, and against which criteria. Faster execution exposes that gap more clearly.
For HCG, AI transformation (AX) means structuring the work context AI needs, creating processes in which it can execute and verify work, and redesigning decision boundaries between people and AI. HR brings together policies, rules, data, systems, authorities, and exceptions in particularly complex ways. Deciding what to automate and which judgments people should retain requires an understanding of how HR actually operates.
HCG AX Consulting designs for AI to work within day-to-day operations through five stages:
HCG designs the point where strategy, HR policies, processes, systems, and people’s roles meet in actual work, going beyond the strategy report.
Deploying tools is getting easier. Structuring the context AI needs, enabling it to execute and verify work, and defining where people retain decision authority take deliberate design. That is why the same tools can produce very different results.
The results of AI transformation depend less on tool features than on how thoroughly an organization redesigns its way of working. HCG’s view of AX connects strategy, policies, processes, systems, and people’s roles for the AI era.