Share

Share

A leader can get a market analysis, a product proposal, and a draft plan before lunch. That speed is useful. It also creates a new problem: a polished answer can reach a decision-maker before anyone has checked whether its assumptions are sound.

Leadership now requires a different kind of attention. We have to decide where AI belongs in the work, what evidence we need before we trust it, and who is responsible when a recommendation becomes an action.

The leader becomes a designer of work

In its 2026 Work Trend Index, Microsoft surveyed 20,000 AI users across ten countries. Only 26% said their leadership was clearly and consistently aligned on AI. The study also found that organizational factors, including culture, manager support, and talent practices, were more strongly associated with employees’ reported AI benefits than individual attitudes and behaviors. Those are survey associations, so they don’t prove that a particular management change will cause better results. They do point to a responsibility leaders can act on: creating conditions in which people know how to use AI well.

That starts with the work itself. If I ask a team to use AI but keep the same approval process, deadlines, and measures of success, I haven’t given them much room to change anything. I need to identify the steps AI can help with, the points where a person should review its work, and the result we’re trying to improve.

Consider a customer service team. AI might draft responses in seconds. The leader still has to ask whether customers got their problems resolved, whether the responses were accurate, and whether employees can recognize cases that need a human conversation. Faster replies are valuable only if they help the customer.

Judgment becomes more visible

AI can give a leader more options. It can challenge a forecast, summarize competing proposals, or reveal a pattern buried in a large set of reports. None of that makes the final choice for me.

IBM’s 2026 CEO Study surveyed 2,000 Chief Executives and equivalent senior leaders. Sixty-four percent said they were comfortable making major strategic decisions based on AI-generated input. In the same study, 85% said functional leaders need to become technology experts in their own domains. I read those findings together: leaders are willing to use AI in consequential decisions, and they know they need enough understanding to question what it gives them.

For me, that means asking better questions before accepting a convincing answer. What information went into it? What has the system left out? What would change my mind? Who will feel the effects if the recommendation is wrong?

A good leader has always needed judgment. AI makes that judgment easier to test, because we can examine more possibilities. It also makes it easier to skip the test when an answer arrives quickly and looks complete.

People need a reason to trust the change

There’s a human side to this that leaders can’t delegate. Employees want to know what AI means for their jobs, where their expertise still matters, and whether they’ll have time to learn. Silence leaves them to fill in those answers themselves.

IBM’s CEO study found that 83% of respondents believed AI success depends more on people’s adoption than on the technology. Microsoft found that AI users who reported stronger skills could still be held back by their organization’s rules and practices. These findings come from different surveys, but they suggest a practical question for leaders: have we made it possible for people to use these tools with confidence and care?

I’d begin by showing the team how I use AI and where I reject its output. I’d invite people to share mistakes as well as wins. Then I’d make sure their performance measures reward better outcomes, thoughtful review, and useful improvements to the process. If an employee catches a flawed AI recommendation before it reaches a customer, that’s good work. We should treat it that way.

AI is changing the role of a leader because it changes how work moves through an organization. More can be produced, tested, and proposed in less time. The leader’s job is to make sure that speed serves a clear purpose and that someone remains accountable for the result.

I don’t need to have every answer before my team uses AI. I do need to set the standard for how we question its answers, learn from what happens, and decide what comes next.

LEARN MORE >

Share The Tech Dilemma

I see five trends shaping the rest of the year. They’re connected by one question: as the pace of development rises, can companies still understand, trust, and maintain what they ship?

Share The Tech Dilemma

1. Coding agents become part of the team

AI has moved beyond suggesting the next line of code. A developer can now give an agent a task, let it inspect a codebase, and review the changes it proposes. JetBrains’ 2026 Developer Ecosystem Survey found that, during May through July, 90% of professional developers surveyed used AI coding agents at work at least weekly, and 68% used them daily.

That changes a developer’s day. More time goes toward describing the task, giving the agent useful context, checking its work, and deciding what should ship. I expect strong engineers to remain central because those decisions require an understanding of the customer, the system, and the consequences of getting it wrong.

For leaders, the question is no longer whether a team has tried coding agents. It’s whether the team has a sound way to assign work to them and review the result.

2. Code review becomes a capacity problem

When producing code gets easier, someone still has to understand the changes. That work can pile up at the review stage.

GitHub has published guidance specifically for reviewing agent-generated pull requests. Its warning is practical: an agent may follow patterns in a repository while missing the incident history, unusual customer needs, or operating constraints that the team knows from experience.

I think this will force teams to pay closer attention to the size and clarity of proposed changes. A pull request that explains its purpose, risks, and test results gives a reviewer a fair chance to make a good decision. A flood of unexplained changes puts pressure on people to approve work they haven’t fully examined.

The best measure of an agent’s value won’t be how many pull requests it opens. It’ll be how much useful work the team can review and release with confidence.

3. Software quality moves beyond the code review

A change can look good in a pull request and still cause trouble for customers. New Relic’s 2026 State of AI Coding report found that 78% of the technology leaders it surveyed reported more production incidents after AI-generated code shipped. The same report found that 86% reported more senior staff time spent fixing code. These are respondents’ reports, rather than a measure of the incident rate across all software teams, but the gap deserves attention.

I expect quality practices to follow software further into production. Teams will need to check whether a feature works under real conditions, whether people can complete the task it was built for, and whether a release creates more support work than it saves.

That’s a broader definition of quality than “the tests passed.” It asks whether the software delivers the result the business and its customers expected.

4. Security becomes part of the agent workflow

Coding agents can touch files, dependencies, tests, and development tools. Their usefulness grows with that access, and so does the need to control it.

GitHub’s 2026 guidance for agent pull requests recommends limited permissions and human approval before an agent-driven workflow takes actions that affect production. GitHub has also introduced agentic fixes for code-scanning alerts, with the proposed repair and its validation steps presented for review.

I see this as a shift in everyday development practice. Teams need to know what an agent was allowed to do, what it changed, and how the change was checked. Security can’t depend on someone noticing a problem after the agent has finished.

5. Leaders start measuring outcomes instead of output

AI makes it easier to count activity: generated lines of code, completed tasks, opened pull requests. Those numbers can tell a team what happened, but they can’t tell a leader whether the investment paid off.

DORA’s 2026 report on the ROI of AI-assisted software development argues for connecting development measures to financial outcomes. Its framework also accounts for the learning period when teams adopt new tools, which matters when judging results.

I’d ask whether a team is delivering valuable features sooner, spending less time on rework, and improving the customer experience. If code output rises while incidents and maintenance work rise with it, the productivity story is incomplete.

These trends point to a new balance for software development. Agents can help teams move faster, but speed puts more weight on review, security, quality, and leadership judgment. In 2026, I think the companies that gain the most from AI will be the ones that can show their software works where it counts: in the hands of the people who use it.

LEARN MORE …

Share The Tech Dilemma

Share The Tech Dilemma

Rather than handing employees another chatbot, Uber created what it calls “Agentic Pods,” small teams of AI engineers embedded within business functions such as finance, legal, marketing, human resources, and procurement. Their mission wasn’t to automate individual tasks. It was to observe how work actually happened, redesign entire workflows around AI, and build intelligent agents that could execute large portions of those workflows. The results were remarkable. Financial planning fell from 15 hours to 30 minutes. Financial reporting dropped from two days to 10 minutes. Marketing quality assurance shrank from two weeks to less than an hour.

For the past twenty years, digital transformation has focused on helping people work better through technology. Organizations replaced paper with software, moved applications to the cloud, digitized customer interactions, and automated repetitive processes. AI transformation changes the equation.

Instead of asking how technology can help people perform work, AI transformation asks which work should be performed by intelligent systems and which work should remain human. That distinction changes everything. It reshapes jobs, workflows, decision-making, and even the structure of the organization itself.

I’ve come to believe that successful AI transformation depends on three pillars. Ignore one of them, and the entire effort becomes unstable.

Pillar #1 … People

Companies don’t transform. People do. The biggest mistake I see is leaders treating AI as a technology project instead of a human one. They purchase enterprise licenses, deploy copilots, and expect productivity to improve automatically. It rarely does.

Employees need new skills. They need to understand not only how to use AI but how to supervise it, question it, validate it, and collaborate with it. Critical thinking becomes more valuable, not less. Judgment becomes a competitive advantage.

The luxury retailer Tapestry, the parent company of Coach and Kate Spade, offers an excellent example. Rather than limiting AI to the technology department, the company focused on changing how employees worked. It organized a companywide AI summit, created AI personas representing Gen Z shoppers to help teams understand customer preferences, introduced AI coaching for retail associates, and used AI-assisted design with 3D printing to reduce prototype development time by 75 percent. Yet executives repeatedly emphasized that the real transformation wasn’t the software. It was the culture. Human designers still made the final decisions. AI expanded creativity instead of replacing it.

That’s exactly what successful AI transformation looks like. People don’t become less important. They become more important because they’re responsible for directing increasingly capable intelligent systems.

Pillar #2 … Process

This is where AI transformation separates itself from digital transformation. Digital transformation improved existing processes. AI transformation redesigns them. Too many organizations ask how AI can automate today’s workflow. I think the better question is whether today’s workflow should exist at all.

Every business process should begin with three simple questions.

· Does this activity create value?

· Can AI perform part of the work?

· Should the process itself be redesigned before AI is introduced?

The organizations generating the greatest returns aren’t automating yesterday’s processes. They’re inventing entirely new ones.

McKinsey reinforced this conclusion in its 2026 research on enterprise AI adoption. The firm found that organizations create the greatest business value not by giving more employees access to AI, but by redesigning workflows, operating models, and organizational roles around what AI makes possible. Companies that merely automate existing processes tend to realize incremental gains. Companies that rethink how work flows through the enterprise create transformational results.

Pillar #3 … Trust

Every AI success story eventually reaches the same point. Can we trust the output? The faster AI becomes, the more important validation becomes.

Organizations need governance that establishes accountability, security, privacy, quality, and explainability. Every critical AI decision should have a human owner. Every model should be monitored. Every important output should be validated before it influences customers, finances, or operations.

I don’t view governance as bureaucracy. I view it as the foundation that allows organizations to scale AI with confidence. The software industry illustrates this perfectly.

AI can now generate thousands of lines of code in minutes. That doesn’t mean the software is secure, maintainable, compliant, or aligned with business requirements. In many development organizations, the bottleneck has already shifted. Writing code is no longer the slowest part of software development. Validating AI-generated code has become the scarce resource.

The same principle applies far beyond software engineering. The more work AI performs, the more valuable human judgment becomes.

AI Transformation Is Business Transformation

Looking back, I think we’ve been using the wrong comparison. AI transformation isn’t the next chapter of digital transformation. It’s the beginning of a new operating model.

Digital transformation digitized work. AI transformation redistributes work between people and intelligent systems.

Digital transformation focused on efficiency. AI transformation focuses on redesign.

Digital transformation changed the tools employees used. AI transformation changes the role employees play.

That’s why I believe every successful AI transformation rests on three pillars. Develop people who know how to work alongside intelligent systems. Redesign processes instead of automating outdated ones. Build the trust that allows AI to operate at enterprise scale. Get those three pillars right, and AI stops being another technology investment.

It becomes the way the business creates value.

LEARN MORE …

Share

Share The Tech Dilemma

1. Wipro redeployed capacity equivalent to 20,000 employees

Wipro reports that its AI initiatives created productivity gains equal to the capacity of approximately 20,000 employees. The workers weren’t eliminated. Many were reassigned to other responsibilities. Wipro has also trained more than 100,000 employees in advanced AI skills. This may be the largest publicly reported workforce-productivity result of 2026, although the figures come from Wipro itself.

2. AI became a genuine revenue engine for Oracle

Oracle secured more than $30 billion in new AI cloud contracts during its first fiscal quarter. Its contracted revenue backlog reached $664 billion, and quarterly revenue increased 30 percent. This is strong evidence that enterprise demand for AI infrastructure has moved beyond experimentation. Oracle’s heavy capital spending and negative cash flow make it an impressive commercial success, though not yet an unqualified financial success.

3. Small businesses converted AI into productivity and sales

The Federal Reserve Banks’ 2026 employer-firm survey produced one of the strongest independent pictures of practical AI value:

• 71 percent of AI-using businesses reported increased productivity.

• 39 percent reported improved product or service quality.

• 31 percent reported higher sales.

This may be more important than any single corporate case because it demonstrates value across thousands of ordinary businesses. Writing, marketing, analysis and personal productivity produced the clearest gains.

4. Adobe successfully monetized AI inside existing products

Adobe reported that annual recurring revenue from its AI-first products more than doubled year over year. Its third-quarter revenue reached $6.76 billion, exceeding analysts’ expectations. This shows that customers will pay for AI when it is embedded inside familiar creative workflows. Competitive pressure from Canva and Figma remains significant. Reuters

5. AI made prenatal ultrasound faster and more accessible

The FDA cleared Butterfly Network’s AI system for estimating gestational age from ultrasound images. The system was trained on more than 21 million images and can produce an estimate in under two minutes without traditional fetal measurements or specialist image interpretation. Its greatest potential is in rural and underserved settings where trained sonographers are scarce. Unlike many health-AI announcements, this system has passed a real regulatory threshold. Reuters

6. Technical workers reported substantial increases in work value

METR surveyed 349 technical workers and found a median self-reported increase of roughly 1.4 to 2 times in the value of their work when using early-2026 AI tools. METR warns that self-reported gains can be exaggerated, so this is encouraging evidence rather than a definitive productivity measurement.

The larger conclusion is that 2026’s strongest AI successes haven’t come from autonomous agents replacing entire organizations. They have come from narrow, well-defined applications: assisting employees, accelerating specific workflows, improving access to expertise and adding AI to products people already use.

The Wipro case is probably the year’s largest productivity story. Butterfly Network is one of the strongest human-impact stories. The Federal Reserve findings provide the best evidence that AI’s benefits extend beyond a handful of technology giants.

LEARN MORE …

Leave a comment

Leave a comment

AI has changed how software is produced. Developers can now generate features, tests, integrations, and documentation in a fraction of the time once required. Coding agents can complete entire work items with limited human involvement. That speed has created an uncomfortable possibility. A team can produce more software, close more tickets, and release more frequently while the quality of its product declines.

I believe software quality now has to begin with a different question: Does the system produce the right result for the business and the people who depend on it?

Working Code Isn’t the Same as Quality Software

AI-generated code can look remarkably polished. It may follow naming conventions, include comments, pass automated tests, and survive a code review. None of those signals proves that it handles the right business rules or responds correctly when real-world conditions become messy.

Research published in March 2026 examined 304,362 verified AI-authored commits across 6,275 GitHub repositories. The researchers identified 484,606 issues introduced by AI-generated changes. Code smells accounted for 89.1 percent of them, and 24.2 percent of the tracked issues were still present in the latest version of the repository. More than 15 percent of commits produced by every AI assistant studied introduced at least one issue. The findings suggest that AI can create maintenance costs that remain long after the initial burst of productivity has passed.

A January 2026 study reached a related conclusion after examining AI-generated build code. Researchers found 364 maintainability and security issues, including missing error handling and hardcoded paths or URLs. More than 61 percent of the agent-generated pull requests were approved and merged with little human intervention. The code entered the system, but approval didn’t establish that the code was dependable.

This is where our traditional quality signals begin to weaken. A green test suite may show that the software behaved as expected under the conditions we anticipated. It can’t tell us whether our expectations were complete. A successful deployment tells us that the code reached production. It doesn’t tell us whether the software helped a customer finish the work they came to do.

Quality Has Become a Business Outcome

In the age of AI, I define software quality across five connected dimensions: correctness, security, maintainability, business performance, and trust. If one dimension fails, the product’s apparent success can conceal a growing liability.

Correctness means more than producing a technically valid response. The result has to reflect the company’s rules, the user’s circumstances, and the intent behind the request. Security now includes the behavior of models, agents, tools, data connections, and third-party services. Maintainability requires code that engineers can understand and change, even when no person wrote the first version.

Business performance asks whether the software completes valuable work. Trust measures whether customers and employees are willing to rely on it. A system that receives thousands of interactions but forces users to confirm every answer isn’t delivering quality. It’s transferring validation work to the customer.

A March 2026 synthesis of 24 empirical studies found that the quality of AI-generated code depends on prompt design, task specifications, developer expertise, the AI system, and the way people interact with it. Results varied across correctness, security, maintainability, and complexity. The researchers’ conclusion matters because it moves quality away from the model alone. Quality emerges from the full human and technical system surrounding the model.

Human review can’t be assumed either. A May 2026 analysis of AI-generated pull requests found that most received no review. When reviews occurred, AI agents performed much of the work. Human involvement frequently appeared as instructions to another agent instead of an independent evaluation of the change. A review count may therefore create the appearance of oversight without proving that a person examined the code.

This doesn’t mean companies should slow AI adoption. It means validation has to advance at the same pace as generation. We need tests tied to real requirements, review standards that identify who or what performed the review, production monitoring connected to customer outcomes, and clear ownership when the system fails.

I no longer think quality can be inspected into software at the end of development. It has to be defined before the AI begins, measured while the system operates, and challenged whenever the business conditions change.

AI can help us create more software than ever. The companies that win won’t be those that generate the most code. They’ll be the ones that can prove their software deserves to be trusted.

LEARN MORE

Leave a comment

Leave a comment

That’s the kind of AI story every business leader wants to hear. The work was defined. The improvement was measured. The value was tied to an actual business process. Blue Pearl didn’t claim that AI had transformed the entire company. It applied AI to a specific problem and calculated what changed. While this was a vendor-reported result, it provides a useful example of what measurable AI value looks like.

When executives ask me about the ROI of AI, my answer is that there isn’t one universal number. AI doesn’t generate a return because a company bought licenses, launched a chatbot, or gave employees access to a collection of tools. It generates a return when it improves a measurable business outcome.

When AI Activity Is Mistaken for Value

PwC’s 2026 Global CEO Survey, based on responses from 4,454 CEOs across 95 countries and territories, found that only 12 percent said AI had produced both cost and revenue benefits. Fifty-six percent reported no significant financial benefit. Companies that had embedded AI extensively across products, services, demand generation, and decision-making were far more likely to report returns. Those with strong AI foundations were three times more likely to see meaningful financial benefits.

To me, that doesn’t mean AI is failing. It means many companies are measuring deployment instead of value.

They count users, prompts, licenses, agents, or generated documents. Those numbers tell us whether people are touching the technology. They don’t tell us whether costs declined, revenue increased, customers stayed longer, errors fell, or work reached completion faster.

Productivity is especially easy to misunderstand. Workday’s 2026 research, conducted with 3,200 employees using AI at companies with more than $100 million in annual revenue, found that 85 percent saved between one and seven hours per week. That sounds like a significant return. However, nearly 40 percent of those savings were lost to correcting errors, rewriting material, and checking AI-generated output. Only 14 percent of employees consistently experienced clear positive results.

This is why I don’t consider time saved to be ROI by itself. If an employee saves five hours with AI but spends two hours checking its work, the gross benefit isn’t five hours. It’s three. If the remaining time disappears into more meetings, unnecessary output, or a larger workload, the company may never convert that capacity into financial value.

Turning AI Into a Measurable Return

The basic calculation hasn’t changed. What has changed is what belongs in that calculation. The investment side should include software, computing, integration, training, process redesign, security, governance, maintenance, and human review. The benefit side should include validated labor savings, reduced operating expenses, additional gross profit, fewer costly errors, faster cycle times, improved retention, and lower risk exposure.

Google Cloud’s 2026 survey of more than 2,400 executives found that 84 percent were seeing increasing financial returns from AI, while 26 percent reported that returns were accelerating year over year. The leaders shared three characteristics: clear ownership, AI embedded into core workflows, and continuing employee training. Nearly half had integrated AI into major business processes, revenue streams, or new business models.

That finding reinforces my view that the greatest returns won’t come from giving everyone an AI assistant and hoping productivity improves. They’ll come from identifying high-volume, expensive, error-prone workflows and redesigning them around measurable outcomes.

I’d begin with one question: What business result should change?

From there, I’d establish the current cost, speed, accuracy, and outcome rate. Then I’d introduce AI, measure the same factors, subtract every associated cost, and validate that the improvement is real. Usage can support the analysis, but it can’t substitute for it.

The ROI of AI can be substantial. It can also be negative, invisible, or imaginary. The difference isn’t determined by the model alone. It’s determined by whether the company connects AI to completed work, financial performance, and business outcomes that someone is accountable for delivering.

LEARN MORE

Leave a comment

Leave a comment

The first thing I’ve noticed is that AI is no longer an experiment. It has become part of the operating environment. According to McKinsey’s 2026 global AI survey, nearly nine out of ten respondents said their organizations regularly used AI in at least one business function. Forty-four percent said AI was scaling across the enterprise, up from 38 percent one year earlier.

Those numbers tell us that adoption has accelerated. They don’t tell us whether companies are getting the results they expected. McKinsey found that only 37 percent of respondents could connect AI to an improvement in earnings before interest and taxes, essentially unchanged from the previous year. Companies are using more AI, spending more on AI, and introducing it into more parts of the business. Many still can’t show that it’s producing measurable financial value. That’s the central tension defining AI in 2026.

Software Development Became the Proving Ground

Software development has emerged as one of the most visible testing grounds for AI. The JetBrains Developer Ecosystem Survey 2026, which included more than 15,000 professional developers, found that 90 percent were using AI coding agents at work at least weekly. Sixty-eight percent were using them every day.

AI tools can now generate features, explain unfamiliar code, write tests, find defects, prepare documentation, and complete some development assignments with limited human involvement. The role of the developer is shifting from writing every line to directing, reviewing, and correcting work produced by machines.

That doesn’t mean the productivity question has been settled. METR reported in February 2026 that wider adoption of coding agents had made their productivity experiments harder to conduct. Some developers didn’t want to participate if they might be required to work without AI. METR’s preliminary results suggested that AI had begun accelerating some development work, but the organization said selection effects made the size of that improvement difficult to measure reliably.

I think that distinction matters. AI can increase output without increasing value. A development team may produce more code, close more tickets, and complete more pull requests while creating additional review work, security exposure, technical debt, and production failures.

The Validation Problem Is Growing

The New Relic 2026 State of AI Coding Report, based on research conducted by Hanover Research among 200 U.S. technology leaders, found that 67 percent said AI generated or significantly refactored between 51 and 75 percent of their organizations’ weekly code output. That’s a remarkable volume of machine-produced code entering corporate systems.

The challenge is that AI can generate code faster than most organizations can validate it. It can produce convincing explanations, well-structured documentation, and clean-looking code that still contains flawed assumptions. The problem isn’t always visible during a demonstration or code review. It may appear only when the system encounters real customers, incomplete data, unusual workflows, or production-scale demand.

This is why I believe the next phase of AI won’t be defined by generation alone. It will be defined by validation. Companies will need stronger testing, observability, security reviews, human oversight, and business-level performance measures. The question won’t be how much AI produced. It will be whether the result worked, remained secure, and delivered the intended business outcome.

Agents Are Moving Faster Than Governance

AI agents represent the next major shift. These systems don’t just answer questions. They can plan tasks, use tools, modify files, communicate with other systems, and take actions on behalf of an organization.

Governance hasn’t kept pace. Deloitte’s 2026 State of AI in the Enterprise report found that only one in five companies had a mature governance model for autonomous AI agents. Deloitte also found that just 34 percent of organizations were using AI to reimagine the business rather than applying it primarily to efficiency and productivity.

That’s where I believe we are in 2026. AI has moved from novelty to infrastructure, but management practices are still catching up. We’ve become better at giving machines work. We haven’t become equally good at defining accountability, measuring outcomes, and recognizing when those machines are wrong.

Where are we going? Toward more autonomous agents, smaller AI-enabled teams, faster software production, and new business models built around outcomes rather than labor. The companies that succeed won’t be the ones that adopt the most AI. They’ll be the ones that know what to delegate, what to measure, what to verify, and when a human still needs to make the final decision.

LEARN MORE

Leave a comment

The company moved forward with layoffs and organizational changes, but internal evidence showed that the technology hadn’t produced the expected productivity. Technical disruptions increased, employees pushed back, and Meta canceled a planned second round of workforce reductions. By July, Mark Zuckerberg acknowledged that agent development had progressed more slowly than expected. A Reuters’ investigation showed what happened when expectations for AI agents moved faster than their demonstrated performance.

Leave a comment

I didn’t see Meta’s experience as an argument against AI agents. I saw it as evidence that companies had treated agents as technology deployments when they should’ve treated them as a new category of worker. Organizations gave agents assignments, system access, data, and decision-making authority. They didn’t always give them defined responsibilities, measurable standards, ongoing supervision, or structured performance reviews.

Deployment Isn’t the Finish Line

Traditional software either executes its programmed logic or produces an identifiable error. AI agents operate differently. They interpret objectives, select tools, make intermediate decisions, and adjust their actions as conditions change. Two executions of the same assignment can produce different results.

That variability makes uptime, speed, and token usage inadequate measures of success. An agent could remain available, respond quickly, and complete thousands of actions while still damaging the business. It might close support cases without resolving customers’ problems, generate sales leads that never converted, or produce code that increased the review burden on engineers.

The 2026 APEX-Agents benchmark evaluated agents across 480 realistic tasks in investment banking, management consulting, and corporate law. The agents worked across documents, spreadsheets, email, calendars, and other workplace tools. When the benchmark was released, even the strongest systems completed fewer than one-quarter of the professional assignments successfully on their first attempt.

A separate 2026 study, Towards a Science of AI Agent Reliability, evaluated 14 models using 12 measures across consistency, robustness, predictability, and safety. The researchers found that capability improvements had produced only modest gains in reliability. A model could perform well on average and still fail unpredictably across repeated runs.

I believe that distinction matters. Capability tells me what an agent could do under favorable conditions. Performance management tells me what it did consistently inside my business.

Every Agent Needs a Performance Contract

I would have started each deployment with a performance contract. That doesn’t mean an employment agreement. It means a precise definition of the agent’s job, the results it was expected to produce, the decisions it could make, and the conditions that required human intervention.

A customer-service agent, for instance, shouldn’t have been evaluated by the number of conversations it handled. I would’ve measured verified resolution, repeat contacts, escalation quality, customer satisfaction, policy compliance, and the cost per completed resolution. Those measures would’ve revealed whether the agent created value or just activity.

I also would have evaluated the path the agent took. An agent might reach the correct outcome after accessing unnecessary data, calling the wrong tools, repeating expensive actions, or bypassing an approval. The result could look successful while the process introduced risk. Performance management had to examine both the outcome and the conduct that produced it.

Agents Need Reviews and Consequences

Human performance management doesn’t end with an annual rating, and agent performance management shouldn’t end with a pre-deployment test. I would have reviewed agents continuously because models, data, tools, customer behavior, and business rules kept changing.

Those reviews would have compared actual results with defined thresholds. They would have identified recurring failure patterns, tested performance after system changes, and examined cases in which humans corrected or reversed an agent’s work. The findings could then guide better instructions, narrower permissions, additional training data, revised workflows, or a different model.

Agents need consequences. A high-performing agent could receive broader authority. An inconsistent one might require more approvals. An agent that violated safety or compliance limits might be suspended until the failure was understood. Autonomy should be earned and retained through demonstrated performance.

I expect AI agents to become an important part of how companies operate. That makes management more important, not less. The organizations that succeed won’t be the ones that deploy the most agents. They’ll be the ones that know which agents perform, which ones fail, why they fail, and when human judgment still produces the better result.

LEARN MORE

Leave a comment

Instead, AI has created a requirements crisis. The better AI becomes at producing software, the more important it becomes for people to define exactly what that software is supposed to do. Code is becoming easier to generate, while clarity is becoming harder to find.

Leave a comment

The need for clarity challenges how many companies approach AI-assisted development. They invest in coding tools because implementation appears to be the largest source of time and cost. Once AI accelerates implementation, however, the bottleneck doesn’t disappear. It moves into domain understanding, requirements, architecture, validation, integration, and maintenance.

When a human developer receives an ambiguous requirement, that person may stop, ask questions, challenge assumptions, or recognize that two business rules contradict each other. An AI agent can ask questions as well, but it can also fill the gaps with plausible assumptions and continue working. And that’s a problem. Because AI operates at machine speed, a small misunderstanding can negatively impact thousands of lines of code before anyone notices.

That’s where speed becomes dangerous. If a team misunderstands a requirement in a traditional development process, the error may appear in one feature developed over several weeks. With AI agents working across multiple tasks, the same misunderstanding can spread through source code, tests, documentation, database structures, and integrations in a matter of hours.

Requirements Debt

I think of the accumulated result as a requirements debt. Technical debt describes compromises embedded in software. Requirements debt describes ambiguity, missing context, conflicting rules, and untested assumptions embedded in the instructions used to create it. Once those weaknesses enter an AI-driven workflow, each new generation cycle can carry them forward.

There’s also the risk that an AI-generated system can satisfy the written specification and still fail to solve the real business problem. If the requirement says that a loan application should be approved when a particular score exceeds a threshold, the agent can implement that instruction with precision. It won’t know that a compliance rule requires a second review, that a data field has become unreliable, or that the policy applies only in certain states unless someone includes those conditions.

That’s why software quality is no longer limited to whether the code matches the specification. We also have to ask whether the specification matches the business, the customer, the law, and the operating environment. AI can help verify the first question. Human judgment remains essential for answering the others.

Requirements Are Part of the Production System

Companies have traditionally treated requirements as documents that developers need before coding begins. In an AI-driven environment, requirements become part of the production system. They need owners, version control, traceability, acceptance criteria, business-rule validation, and continuous revision. If an organization can’t explain a requirement in a form that people and machines can test, it isn’t ready to delegate the work to an autonomous agent.

This doesn’t mean every requirement needs to become a massive technical document. More documentation doesn’t guarantee more clarity. Teams need specifications that identify the intended outcome, relevant constraints, exceptions, data sources, security rules, acceptance criteria, and the person responsible for resolving conflicts. The goal isn’t to describe every keystroke. It’s to remove the uncertainty that can change the result.

The Need to Understand Intent

AI is also increasing the value of people who understand the business and can translate that knowledge into explicit decisions. Product managers, business analysts, architects, quality engineers, security professionals, and experienced developers aren’t becoming less important. Their work is moving closer to the center of software creation because they define the boundaries within which AI operates.

This changes the skills companies need from their engineering teams. Writing code still matters, but understanding intent matters more. The most valuable professionals increasingly know how to identify missing information, expose conflicting assumptions, define measurable outcomes, and determine what evidence will prove that a system works. They don’t just tell the AI what to build. They create the conditions under which its output can be trusted.

The future of software development doesn’t depend only on how much code AI can produce. It depends on whether people can define the right problem, specify the right outcome, and validate that what the machine builds is what the business needs.

AI has made code abundant. But clarity continues to remain scarce.

LEARN MORE

Leave a comment

Leave a comment

Instead of asking AI to generate a class or fix a bug, I gave it an objective. The system decomposed the work into dozens of smaller tasks, assigned those tasks to specialized AI agents, reviewed their output, requested revisions when necessary, and assembled the finished product. The code no longer came from a single large language model. It came from an organized team of AI specialists coordinated by what the industry now calls an Agent Manager.

Deciding How Work Gets Done

An AI agent performs a task. An Agent Manager decides which agent should perform that task, determines the order in which the work should occur, evaluates the quality of each result, and redirects work whenever something goes wrong. In many ways, it resembles an experienced engineering manager who coordinates a team of specialists rather than writing every line of code personally.

This wasn’t an isolated trend. Gartner predicted that by 2026, 40 percent of enterprise applications would incorporate task-specific AI agents, compared with fewer than 5 percent only a year earlier. At nearly the same time, Deloitte’s 2026 Global Software Industry Outlook concluded that autonomous AI agents were becoming central to software engineering, changing how organizations build products, manage engineering teams, and allocate technology investments.

Those reports pointed toward the same conclusion. As organizations deployed more AI agents, someone had to coordinate them. That responsibility increasingly belonged to the Agent Manager.

I found the implications fascinating because they represented a complete shift in the software development lifecycle. One agent gathered requirements. Another designed the architecture. A coding agent generated the implementation. Separate agents produced automated tests, scanned for security vulnerabilities, reviewed compliance requirements, updated documentation, and prepared deployment pipelines. The Agent Manager supervised every step. It tracked dependencies, evaluated progress, requested corrections, and determined when work was ready for human review.

The Role of Human Expertise

Instead of spending most of the day writing syntax, developers increasingly define objectives, establish constraints, review outcomes, and exercise judgment. As AI made software production dramatically faster, human expertise became concentrated where it always created the greatest value -- deciding whether the software actually solved the right business problem.

In July, Microsoft Research published The Rise of AI Coding Agents, an analysis of enterprise adoption across tens of thousands of software engineers using autonomous coding systems. The researchers found that engineers working with agentic coding tools merged approximately 24 percent more pull requests than expected. The paper also reached an important conclusion that received less attention. Organizations that achieved the greatest improvements weren’t necessarily using better AI models. They were better at integrating AI agents into structured engineering workflows and management practices.

The same month, researchers published Human Oversight in Agentic Software Development on arXiv. The study analyzed more than 25,000 AI-generated pull requests across 2,361 public GitHub repositories. The authors discovered that successful projects almost always relied on continuous human oversight throughout the development process. Fully autonomous software development remained rare. Human managers reviewed decisions, resolved conflicts between agents, validated outputs, and intervened whenever quality declined.

The Increasing Importance of Software Validation

AI didn’t eliminate software quality as a competitive advantage. It increased its importance. When generating code becomes inexpensive, validating that code becomes dramatically more valuable. The Agent Manager exists not to maximize code generation but to maximize confidence in the software that reached production.

Researchers from several universities published Lessons from a Twelve-Week Agentic Software Engineering Project in July 2026 after documenting a production software initiative built largely through frontier AI coding agents. Rather than celebrating autonomous development, the authors concluded that organizations needed stronger governance, repeatable engineering controls, and continuous supervision as AI systems became more capable. Greater autonomy increased the importance of management rather than reducing it.

For decades, software managers assigned work to developers. Tomorrow’s engineering leaders will increasingly assign work to intelligent agents. Their responsibilities will extend beyond sprint planning and code reviews. They’ll monitor agent performance, optimize task allocation, evaluate reasoning quality, manage operating costs, enforce security policies, and determine where human judgment should remain part of the workflow.

AI Governance and the Future

The engineering dashboard of the future won’t focus exclusively on developer productivity. It will also measure how effectively the Agent Manager coordinated specialized agents, how frequently human intervention became necessary, how reliably AI-generated software passed validation, and how efficiently computational resources were consumed.

That shift also explains why AI governance has become one of the fastest-growing areas of enterprise investment. Deloitte’s 2026 Global Software Industry Outlook identified governance, operational controls, and responsible AI oversight as critical factors separating successful AI deployments from expensive pilot projects that never reached production. The technology itself was only part of the equation. Sustainable value depended on disciplined management.

Software development has always been an exercise in coordination. The people involved are changing, but the need for leadership isn’t. Yesterday, engineering managers coordinated teams of developers. Tomorrow, they’ll coordinate teams of AI agents. The technology will continue advancing at an extraordinary pace, yet the organizations that thrive won’t be the ones generating the most code. They’ll be the ones managing autonomous who manage autonomous intelligence better than everyone else.

LEARN MORE