How to lead AI development in 2026: 10 evidence-based practices that turn adoption into ROI

Q. What is AI development in 2026?

AI development is the practice of building software that uses AI models, and building software with AI's help. In 2026, both halves have merged into one discipline.

The first half covers products. You build features on large language models (LLMs), retrieval systems and AI agents. The second half covers process. Your developers use AI coding assistants and agents to write, test and review code. According to the Stanford AI Index 2026, industry produced over 90% of notable frontier models in 2025. So most teams don't train models. Instead, they compose them.

That shift changes the core skill. Andrej Karpathy, a founding member of OpenAI and former director of AI at Tesla, calls it "Software 3.0." In his framing, prompts are programs and English is the programming language. Data from 2025 shows how fast this spread. GitHub's Octoverse 2025 reports 80% of new developers use Copilot in their first week. It also counts more than 4.3 million AI-related repositories, nearly double in under two years.

"Your prompts are now programs that program the LLM. And remarkably these prompts are written in English."

Watch: Andrej Karpathy, "Software Is Changing (Again)." His autonomy slider powers the interactive tool above.

Q. Why does 2026 change the AI development playbook?

Capability jumped faster than organizations adapted. Models now finish real tasks, but enterprise returns have stalled.

Consider the benchmarks first. According to the Stanford AI Index 2026, performance on SWE-bench Verified rose from 60% to near 100% in a single year. On OSWorld, which tests agents on real computer tasks, success leapt from 12% to roughly 66%. Still, agents fail about one in three structured attempts. Stanford calls this unevenness the "jagged frontier." Its sharpest example: a model won IMO gold, yet the top model reads analog clocks correctly only 50.1% of the time.

Meanwhile, the business picture looks flat. McKinsey's State of AI 2026 surveyed 1,719 people in 97 countries between May and June 2026. It finds 44% of organizations scaling AI across the enterprise, up from 38%. Yet only 37% attribute any EBIT impact to AI, unchanged from 2025. High performers remain stuck at about 6% of respondents. In short, the bottleneck has moved from the model to the operating model.

Watch: Stanford HAI's overview of the 2026 AI Index Report.

Q. Do AI coding tools actually make developers faster?

Probably yes in 2026, but by less than developers feel. Perception and measurement still diverge sharply.

The best evidence comes from METR's randomized trials. In its early-2025 study, experienced open-source developers took 19% longer with AI. Before starting, they had forecast a 24% speedup. That gap between belief and reality is the key finding. METR's February 2026 update tells a more hopeful story. For returning developers, it estimates an 18% speedup, with a confidence interval from −38% to +9%. New recruits showed only a 4% speedup.

However, METR warns that its new data understates the gain. Between 30% and 50% of developers withheld some tasks because they refused to do them without AI. The survey data matches this caution. According to the Stack Overflow 2025 Developer Survey, 66% of developers cite "AI solutions that are almost right, but not quite." Another 45% say debugging AI code takes longer. The lesson is simple: measure your own teams instead of trusting impressions.

Q. Which AI development stack should you build on?

Build on open standards for context, mainstream tools for orchestration, and your existing observability stack. Avoid lock-in where standards exist.

The biggest standards story is the Model Context Protocol (MCP). On 9 December 2025, the Linux Foundation formed the Agentic AI Foundation (AAIF). Its founding projects are MCP, goose and AGENTS.md. The announcement counts more than 10,000 published MCP servers. Platinum members include AWS, Anthropic, Google, Microsoft and OpenAI. Therefore, MCP is now a neutral bet, not a single-vendor one.

For the rest of the stack, follow what builders actually use. The Stack Overflow 2025 survey asked developers who build agents about their tools. Ollama (51%) and LangChain (33%) lead orchestration. Redis (43%) leads agent memory, ahead of ChromaDB (20%) and pgvector (18%). For observability, 43% use Grafana plus Prometheus. In other words, teams extend familiar DevOps tools rather than adopt new AI-native ones.

Percentages from the Stack Overflow 2025 survey (respondents who build or use agents). MCP count from the Linux Foundation, December 2025.

Q. When do AI agents actually pay off?

Agents pay off when a decision is needed inside a redesigned workflow. They fail when teams bolt them onto old processes for hype.

Gartner is blunt about the risk. It predicts over 40% of agentic AI projects will be canceled by the end of 2027. The causes are escalating costs, unclear business value and inadequate risk controls. Gartner also flags "agent washing," the rebranding of chatbots and RPA as agents. It estimates only about 130 of the thousands of agentic AI vendors are real, according to its June 2025 forecast.

Yet adoption keeps climbing among large firms. McKinsey's 2026 survey shows 40% of large organizations now scale AI agents, up from 27%. Smaller organizations stayed flat at 22%. Coding agents also reshape budgets. Notably, 32% of respondents skipped buying at least one software product because they could build it with agentic coding tools. Gartner's rule of thumb is useful here. Use agents when decisions are needed, automation for routine workflows and assistants for simple retrieval.

Categories follow Gartner's guidance (Anushree Verma, June 2025). Examples are illustrative.

Q. What are the 10 evidence-based AI development practices?

Ten practices separate AI high performers from everyone else. Each one maps to a finding from 2025–2026 research.

1. Redesign workflows, don't insert toolsNearly three-quarters of McKinsey's high performers fundamentally redesign workflows, up from 55%. Only one-quarter of others do.
2. Set autonomy per taskUse the autonomy slider. Low autonomy for unfamiliar code; higher autonomy only where tests can catch errors.
3. Invest in your internal platformDORA 2025 finds 90% of organizations use at least one platform. Platform quality correlates with unlocking AI value.
4. Fortify safety nets before scalingDORA finds AI raises throughput but hurts delivery stability without strong testing and fast feedback.
5. Give models your contextConnect internal docs, APIs and codebases via MCP servers and AGENTS.md files.
6. Build evals before featuresWrite task-level evaluations first. 87% of developers worry about agent accuracy (Stack Overflow 2025).
7. Budget tokens like cloud spendAbout 20% of organizations say AI operating costs already constrain use (McKinsey 2026).
8. Keep a human in the loop where trust matters75% of developers would still ask a person when they don't trust AI's answer.
9. Manage agent risk explicitlyHigh performers actively mitigate unauthorized actions and exploited vulnerabilities. Stanford counts 362 AI incidents, up from 233.
10. Define impact metrics up frontHigh performers are twice as likely to have defined processes to measure AI impact (McKinsey 2026).

"AI doesn't fix a team; it amplifies what's already there."

Q. What does Klarna's AI rollout teach developers?

Klarna shows that AI can absorb huge volume fast. It also shows that optimizing only for cost erodes quality.

The launch metrics (February 2024)

Klarna launched an OpenAI-powered assistant and reported results after one month. According to Klarna's press release, it handled 2.3 million conversations, two-thirds of all service chats. That equaled the work of 700 full-time agents. Resolution time fell from 11 minutes to under 2. Repeat inquiries dropped 25%. It ran in 23 markets and more than 35 languages. Klarna estimated a $40 million profit improvement for 2024.

The correction (May 2025)

Then the story turned. In a Bloomberg interview reported by CX Dive, CEO Sebastian Siemiatkowski admitted a quality problem. Klarna began recruiting human agents for a flexible, Uber-style pilot. The AI still handles two-thirds of inquiries, and response times improved 82% since launch. However, customers now always get a path to a human.

"As cost unfortunately seems to have been a too predominant evaluation factor when organizing this, what you end up having is lower quality."

Developer takeaway: build an escalation path into every agent from day one. Measure quality signals, like repeat contacts and satisfaction, alongside cost.

Q. What do industry experts recommend?

Experts converge on three points: cut through hype, design for partial autonomy, and keep humans available.

• Anushree Verma, Senior Director Analyst, Gartner: "Most agentic AI projects right now are early stage experiments or proof of concepts that are mostly driven by hype and are often misapplied." (Gartner, June 2025)
• Jim Zemlin, Executive Director, Linux Foundation: "We are seeing AI enter a new phase, as conversational systems shift to autonomous agents that can work together." (Linux Foundation, December 2025)
• Julie Geller, Principal Research Director, Info-Tech Research Group: "The key takeaway is that AI should augment human agents — not replace them." (CX Dive, May 2025)
• Andrej Karpathy, founding member of OpenAI: recommends products with an "autonomy slider," where "you are in charge" of how much autonomy you give up per task. (YC Library)

Q. How do you keep AI development compliant in 2026?

Track the EU AI Act's revised dates, label AI-generated content, and document risk controls now. Several obligations already apply.

The EU's Digital Omnibus agreement moved several deadlines. According to Gibson Dunn's analysis, Article 50 transparency obligations largely kept their 2 August 2026 date. Systems placed on the market before then get a grace period for watermarking until 2 December 2026. Stand-alone high-risk systems under Annex III now face obligations from 2 December 2027. AI embedded in regulated products under Annex I moves to 2 August 2028.

Governance matters beyond Europe too. The Stanford AI Index reports documented AI incidents rose to 362, up from 233 in 2024. It also finds responsible AI benchmark reporting remains "spotty." Global context matters for teams shipping worldwide. For example, the U.S. reports the lowest trust in its own government to regulate AI, at 31%. As a result, internal controls must carry more weight than external rules.

This is not legal advice. Confirm obligations with counsel and the EU AI Act Service Desk timeline.

Q. How do you measure ROI on AI development?

Measure outcomes at the workflow level, against a pre-AI baseline. Usage and self-reported speed are not ROI.

The gap between individuals and companies explains why. McKinsey finds 80% of respondents say AI improved their personal productivity. Yet only 37% see enterprise EBIT impact. Similarly, DORA reports more than 80% of professionals feel more productive. Still, AI adoption correlates with lower software delivery stability. Therefore, track a balanced scorecard. Pair speed metrics with quality and cost metrics.

Investment trends raise the stakes. McKinsey reports 28% of organizations spend over 10% of their ICT budget on AI. Further, 60% expect to increase AI investment next year. Workforce expectations also need checking against reality. Last year, 32% expected AI-driven headcount declines, but only 14% reported them. So model ROI on throughput and quality, not on assumed job cuts.

Q. What is the 90-day AI development implementation plan?

Run six steps over 90 days: baseline, guardrails, context, evals, one redesigned workflow, then a scale-or-cancel decision.

1. Days 1–14Capture a baselineRecord lead time, change failure rate, review load and support metrics. Without this, METR-style perception bias will mislead you.
2. Days 8–21Publish policy and guardrailsName approved tools, data rules and review gates. DORA lists "clarify and socialize your AI policies" as step one.
3. Days 15–35Build the context layerStand up MCP servers for docs, tickets and APIs. Add AGENTS.md files to key repositories.
4. Days 22–50Ship evals and safety netsCreate 30–50 task evals per use case. Raise test coverage on code that agents will touch.
5. Days 36–75Redesign one high-value workflowChoose one workflow with a clear owner and metric. Redesign it end to end, including a human escalation path.
6. Days 76–90Measure, then scale, fix or cancelCompare against baseline on all five scorecard dimensions. Cancel early if value is unclear, as Gartner advises.

Q. What comes next for AI development in 2026–2027?

Expect agents inside most enterprise software, a shake-out of weak agent projects, and rising pressure on cost and governance.

• Agents become a default feature. Gartner predicts up to 40% of enterprise apps will include task-specific agents by the end of 2026, up from under 5% in 2025 (Gartner, August 2025).
• Autonomous decisions grow. Gartner expects at least 15% of day-to-day work decisions to be made by agentic AI by 2028, up from 0% in 2024.
• A cancellation wave arrives. Over 40% of agentic projects face cancellation by end-2027. Teams with evals and metrics will survive it.
• Build-vs-buy shifts further. With 32% already skipping software purchases, expect more internal tooling built by coding agents.
• The open ecosystem broadens. Stanford reports open-source contributions from outside the U.S. and Europe now outpace Europe on GitHub.
• The revenue prize grows. In its best case, Gartner sees agentic AI driving about 30% of enterprise application software revenue by 2035, over $450 billion.

Q. What should you do this week?

Start with measurement and one workflow. Everything else compounds from there.

1. Today: pick one workflow with a named owner and a dollar metric.
2. This week: capture a two-week baseline and publish a one-page AI policy.
3. Within 30 days: connect one internal system through an MCP server and write your first 30 evals.
4. Within 90 days: run the go/no-go review against your five-part scorecard.

Q. Frequently asked questions

Q. Which resources, schema and links should you use?

Resource list

• Stanford HAI AI Index 2026: benchmarks, investment and incident data.
• McKinsey State of AI 2026: enterprise adoption and high-performer practices.
• DORA State of AI-assisted Software Development 2025: the AI Capabilities Model.
• Agentic AI Foundation: MCP, goose and AGENTS.md.
• METR late-2025 productivity dataset: raw data for your own analysis.
• Stack Overflow 2025 AI survey: tool and trust benchmarks.

Schema markup instructions

1. Paste the Article, FAQPage and HowTo JSON-LD blocks into the page <head>. This page already includes all three.
2. Author details are filled in. Replace the remaining [BRACKETED] fields with your organization, logo and canonical URL.
3. Keep FAQ answers in schema identical to the visible answers on the page.
4. Validate with Google's Rich Results Test and the Schema.org validator before publishing.
5. Review/rating schema does not apply here. Use it only if you publish genuine product reviews.

{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "How to Lead AI Development in 2026: 10 Evidence-Based Practices That Turn Adoption Into Measurable ROI",
  "datePublished": "2026-10-06",
  "dateModified": "2026-10-06",
  "author": {"@type": "Person", "name": "Marco Ballesteros",
    "jobTitle": "Senior Project Manager | SEO, Automation & Growth Strategy",
    "url": "https://www.linkedin.com/in/marcoaballesteros/",
    "sameAs": ["https://www.linkedin.com/in/marcoaballesteros/"]},
  "publisher": {"@type": "Organization", "name": "[YOUR ORGANIZATION]",
    "logo": {"@type": "ImageObject", "url": "[LOGO URL]"}},
  "mainEntityOfPage": "[CANONICAL URL]"
}

{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "Why do so many AI agent projects fail?",
    "acceptedAnswer": {"@type": "Answer",
      "text": "Gartner predicts over 40% of agentic AI projects will be canceled by the end of 2027. It cites escalating costs, unclear business value and weak risk controls."}
  }]
}

{
  "@context": "https://schema.org",
  "@type": "HowTo",
  "name": "How to roll out AI development practices in 90 days",
  "totalTime": "P90D",
  "step": [
    {"@type": "HowToStep", "position": 1, "name": "Baseline (days 1-14)",
     "text": "Measure cycle time, change failure rate and review load before changing tools."}
  ]
}

Internal linking suggestions

Target URLs are suggested slugs. Map them to real pages on your site.

Recommended images and charts (with alt text)

Q. Sources

1. Stanford HAI. The 2026 AI Index Report. 2026.
2. McKinsey & Company. The State of AI: Global Survey 2026. Fielded 4 May–8 June 2026; 1,719 respondents.
3. METR. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity. 10 July 2025.
4. METR. Wider adoption of AI has made it more difficult to measure task-level productivity. 24 February 2026.
5. Stack Overflow. 2025 Developer Survey: AI. 2025.
6. Google Cloud / DORA. Announcing the 2025 DORA Report. 2025.
7. GitHub. Octoverse 2025. October 2025.
8. Gartner. Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. 25 June 2025.
9. Gartner. 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026. 26 August 2025.
10. Linux Foundation. Formation of the Agentic AI Foundation. 9 December 2025.
11. Gibson Dunn. EU AI Act Omnibus Agreement. 2026.
12. Klarna. Klarna AI assistant handles two-thirds of customer service chats. 27 February 2024.
13. CX Dive. Klarna changes its AI tune and again recruits humans. 9 May 2025.
14. Y Combinator. Andrej Karpathy: Software Is Changing (Again). 2025.