The 31% Problem: How Systems Thinking (Not Just AI) Cut Our Development Cycle Time

Every vendor slide deck I've seen in the last twelve to eighteen months promises 80% efficiency gains. Some go higher. One platform claimed we could "eliminate the bottleneck of human content creation entirely." And I'll be honest, part of me wanted to believe it. Who wouldn't want to 5x their team's output overnight? My former VP gave part of my team Microsoft Copilot Premium, and seriously asked two weeks later if we had 10x'd their output yet.

But here's the number our team actually hit after a full year of redesigning our content development system: 31%.

Not 80%. Not "10x faster." Thirty-one percent. And the part that took me the longest to admit? The AI tools were maybe 30% of that equation. The other 70% was systems thinking, planning, accountability, trust, and a whole lot of uncomfortable conversations about ego. Why aren't more L&D leaders talking about the non-AI parts of their efficiency gains?

---

TL;DR

• My global L&D team achieved a 31% reduction in development cycle time after redesigning our entire content pipeline as a system, not just adding AI tools. Measured across 637 eLearning modules over 12 months.
• The biggest gains came from systems thinking: planning first, cadenced workshops, shared dashboards, accountability tools, ego agreements, and trust conversations. AI accelerated specific tasks, but without the system redesign, it would have accelerated chaos.
• Josh Bersin's February 2026 research calls AI "poised to reinvent the $400 billion corporate training market," but only 25% of L&D teams factor AI into workflows routinely (LinkedIn Workplace Learning Report, 2025). The teams getting results are redesigning workflows, not just deploying tools.

---

📊 How We Actually Measured This

I keep seeing a version of this question in L&D forums: "67% of IDs report 'moderate-to-significant time savings' with AI, but what does that actually mean in hours?" It's a fair question. And I think the L&D field has a credibility problem when it comes to AI claims, and it starts with how we (don't) measure things.

Here's what we tracked. Before integrating any AI tools, our average content module took roughly 34 working days from initial brief to publication in the LMS. That included needs analysis, SME interviews, storyboarding, first draft, review cycles, multimedia production, QA, and deployment.

After 12 months of system redesign (including AI integration), that average dropped to approximately 23 working days. That's the 31%.

But here's what I want to be transparent about: was every single variable controlled? No. Were there other process improvements happening simultaneously? Yes, we also compressed our content approval cycle from 6 weeks to 2 weeks during that period. Could some of the gains be attributed to the team simply getting better at their jobs? Absolutely.

So is 31% a clean, isolated number? I'd be lying if I said yes. But it's the best measurement we had across 637 modules built by a 26-person team spanning three continents, and it's a lot more honest than a vendor demo. What would it look like if more of us shared our actual numbers rather than vendor projections?

Someone recently asked me: "What metrics should I use to prove AI is actually helping my team? Completion rates and satisfaction surveys don't tell the real story." They're right. We tracked cycle time (brief to publish), rework rate (how often modules were sent back for revision), and SME review turnaround time. Those three together told a much more honest story than any single satisfaction metric. If we're serious about proving impact, we need to measure the pipeline, not just the output.

---

🧭 The System, Not the Tool

Here's the thing I wish I'd understood from day one: we didn't get 31% because we added AI tools. We got 31% because we redesigned our entire content development pipeline as an interconnected system. AI was one piece. If I'm being honest, it wasn't even the most important piece.

I hear some version of this constantly: "Our content development process was never designed for the pace AI enables. How do I redesign the pipeline without breaking everything?" That question captures exactly where most teams are stuck. They've bolted AI onto a process that was designed for a different era, and they're wondering why it feels like they're running a modern engine on a dirt road.

Cornerstone's 2026 research on learning workflow transformation makes this point clearly: organizations replacing traditional L&D models (where training is separated from daily work) with integrated systems are seeing the biggest gains. It's the integration that matters, not any single tool.

McKinsey's research on AI high performers found that only about 6% of organizations are seeing 5%+ EBIT impact from AI, and those organizations share a common trait: they redesign workflows rather than just adding tools. Are we redesigning, or are we just adding?

I started thinking about our content pipeline the way a systems engineer might think about a production line. Every stage is connected. A bottleneck in SME review doesn't just slow review, it cascades into storyboarding delays, pushes back multimedia production, and compresses QA into a rushed final sprint. How many of us have treated each stage of content development as if it exists in isolation?

And then there's the harder question underneath all of this: "How do I move from 'learning operations' as a concept to an actual operating model?" For us, the answer started with mapping the full system before touching any tools. Not a theoretical exercise. A literal wall-sized flowchart showing every handoff, every approval, every place where work sat waiting.

That shift in perspective, seeing the whole pipeline as one system, changed everything. And it had to happen before we touched a single AI tool.

---

📋 Planning First. Always.

This might sound obvious, but I need to say it because I made the mistake of skipping it: the plan has to come first. Always. Without a solid plan, AI tools just accelerate chaos.

When we first started experimenting with Claude Opus 4 (the model version we started with, later upgrading to Claude Opus 4.1 for more complex work), I was so excited about speed that I let a few projects skip the planning stage. The thinking was, "We can draft faster now, so let's just start drafting." That was a mistake I'm still a little embarrassed about.

What happened? We generated first drafts in 30 minutes, only to find they took three days to untangle because nobody had aligned on learning objectives, audience context, or success criteria upfront. We were moving fast in the wrong direction. Has anyone else fallen into this trap?

After that, we made planning non-negotiable. Every project started with a documented plan that included learning objectives mapped to Bloom's taxonomy, audience profiles, SME availability commitments, quality criteria agreed upon by all parties, and a timeline with accountability checkpoints. No plan, no AI tools. Period.

The Seertech 2026 L&D trends report puts it well: organizations moving "from strategy to proof" in 2026 are those that start with strategic alignment, not with tool selection. How often are we choosing tools before we've defined the problem?

---

🤝 Ego Gets Checked at the Door

This is the uncomfortable part. And I'll admit, I didn't handle it well at first.

Before any work began on a project, we needed an upfront agreement from all parties — IDs, SMEs, reviewers, and project owners — that ego doesn't drive decisions. We needed to agree on what quality looks like before work begins. Not after the first draft. Not during the review cycle. Before.

Why? When AI generates a first draft in 30 minutes, it shifts the team's emotional dynamics. The ID who spent years mastering storyboarding might feel threatened. The SME might reflexively dismiss AI-generated content. The reviewer might over-correct to justify their role. I saw all of these reactions on our team, and I should have anticipated them. One question I've been sitting with: "Some of my senior IDs see AI as a threat and quietly refuse to use it. How do I have that conversation?" I don't think there's a clean answer. But I've learned that the conversation has to start with acknowledging the threat is real, not dismissing it.

Harvard Business Review's February 2026 article by Amy Edmondson and Jayshree Seth nails this: "The same AI tools that promise to enhance productivity can create predictable patterns of team dysfunction that mirror classic organizational behavior problems." They recommend treating AI adoption as a team development effort, not just a tech upgrade. I wish I'd read that article before we started.

So we built what I call "ego agreements" into our project kickoffs. Everyone in the room agrees: we're optimizing for learner outcomes, not for anyone's sense of ownership over a particular stage. The best idea wins, whether it came from an ID, an SME, Claude, or an intern. Is that easy? No. Do people still get defensive sometimes? Yes. But having the agreement in place gives us something to point back to. What would change in our teams if we had these conversations openly?

---

🔧 Where AI Saved Us the Most Time

Not all tasks benefited equally. That's probably the most important thing I learned. When we broke down the workflow, three areas stood out.

First drafts. This was the biggest win. Using Claude 4 (later upgraded to Claude 4.5), our IDs could generate a rough first draft of a module in 30-45 minutes instead of 3-4 hours. But here's the catch: that draft still needed 2-3 hours of revision by someone who understood the content. We weren't saving 80%. We were saving maybe 40-50% on this one stage. And honestly? I initially overestimated even that number before we started tracking it properly.

Assessment generation. Quiz questions, knowledge checks, scenario stems. AI is consistently strong at this when given clear learning objectives and Bloom's taxonomy levels. Our team reported cutting assessment development time roughly in half. Were the AI-generated questions perfect? No. Did they give us a running start that eliminated the "blank page" problem? Every time.

Formatting and structure. Converting SME brain dumps into structured storyboards, generating consistent formatting across modules, and creating SCORM-compliant metadata. These repetitive, low-creativity tasks were exactly where AI shines. Nobody misses spending 45 minutes reformatting a storyboard template.

The Synthesia AI in L&D Report 2026 (surveying 421 professionals across 20,000+ data points) found similar patterns: 84% reported faster production as AI's top benefit, with content and quiz drafting (60%) and voice generation (63%) as the most common applications. Does that match what other teams are seeing?

---

🚫 Where AI Made Things Worse

This is the part vendors don't put on their slides. I'll admit, it took us a few painful months to figure out.

Dr. Philippa Hardman's research on the "AI Illusion in L&D" captures something I've been struggling to articulate. She found that IDs feel more confident and faster with AI, but the actual output quality and volume don't always reflect that confidence. It's the "illusion of impact," and it's uncomfortable to confront. One question from her research stopped me cold: "We adopted AI tools last year, and we're still producing roughly 12 projects annually. Where are the efficiency gains everyone promised?" If that sounds familiar, you're not alone.

Nuanced SME content. We tried using AI to draft technical product training for a complex automotive DMS platform. The results were confidently wrong. Not slightly off. Confidently, authoritatively, dangerously wrong. The kind of wrong that would have damaged our credibility with the engineering team we were training. How many L&D teams have shipped AI-generated content without catching errors like these?

Cultural sensitivity. With a team across three continents building content for global audiences, we quickly learned that AI's cultural awareness is, at best, shallow. Idioms, examples, humor, and even the appropriate level of formality in certain markets. These required human judgment that AI consistently missed. I should have anticipated this one. Looking back, it was obvious, but I was so focused on speed metrics that I didn't think about what speed might cost us in cultural nuance.

Complex scenario design. Branching scenarios with multiple decision points, emotional nuance, and realistic consequences? AI produced generic, predictable scenarios that felt like they were written by someone who'd read about the job but never done it. Our experienced IDs could spot this immediately. Could a less experienced team have missed it?

Here's the question that keeps coming up in forums: "AI is great for first drafts, but my team is spending almost as much time fixing AI output as they would writing from scratch. Is that normal?" Based on our experience, yes, for complex content. The fix isn't better AI. The fix is knowing which tasks to hand to AI and which to keep human. We learned that the hard way.

The METR developer productivity study from mid-2025 found something that resonated deeply with our experience: developers using AI tools actually took 19% longer on complex tasks, but believed AI had sped them up by 20%. That perception gap is real, and it's dangerous. Are we measuring our actual output, or how productive we feel?

---

🔄 The Workflow Changes Nobody Talks About

Here's what I keep coming back to: the tools are maybe 30% of the equation. The workflow redesign is 70%. And the redesign had seven components that mattered just as much as any AI tool we deployed.

Cadenced Workshops, Not Ad-Hoc Meetings

What changed for us was shifting from reactive meetings to a predictable cadence. Instead of scheduling time for when things went wrong, we moved to weekly 90-minute workshops where IDs brought work in process, SMEs provided live feedback, and reviewers flagged issues before they compounded. The keyword is "cadence." Not "as needed." Not "when there's a problem." Every week, same time, same structure. Would this work for every team? I'm not sure. But for ours, the rhythm mattered more than any single agenda item.

This alone probably saved us more time than any AI tool. Why? Because problems got caught in days instead of weeks. An ID heading in the wrong direction got redirected in the next workshop, not after completing a full draft. What if our biggest efficiency gains come from rhythm, not technology?

Accountability Tools (The Tool Doesn't Matter)

I want to be clear about something: the accountability tool doesn't matter nearly as much as the commitment to using it. We started with Excel spreadsheets. Moved to Jira, though in retrospect, Asana or a Microsoft Planner plan would have been a better fit. Why is a tale for another day. Some teams preferred Monday.com or Asana. In our experience, the tool mattered far less than the commitment to using it. What matters is that every task has an owner, a deadline, and visibility.

This connects to a question I see constantly, often tied to Amy Edmondson's research on psychological safety: "How do I create accountability without it feeling punitive? RACI charts exist, but nobody follows them." We struggled with this, too. What worked for us was making accountability visible (dashboards, not check-ins) and framing it around the work, not the person. When a task is overdue, the dashboard shows it as overdue. It doesn't say someone failed. That distinction matters more than I initially realized.

Research from Gallup consistently shows that teams with clearly defined roles are 53% more efficient and 27% more effective than those with ambiguous responsibilities. Our experience confirmed this completely. When every module had a clear owner at every stage, and that ownership was visible to the whole team, things moved.

Dashboards Shared to All Stakeholders

This one made some people uncomfortable, and that's exactly why it worked. We built simple dashboards that show each project's status: on track, at risk, or blocked. And we shared them with everyone. Not just the L&D team. Stakeholders. SMEs. Leadership.

Why? Because transparency drives accountability in a way that status meetings never can. When an SME can see that their delayed review is blocking three downstream tasks, the conversation changes. We don't have to chase them. The dashboard does it for us. Analytics dashboards that reveal skill gaps, performance trends, and bottlenecks early are a core 2026 performance management trend (Synergita, 2026), and our experience shows they work just as well for content development pipelines.

And this addresses another pain point I hear constantly: "My stakeholders stopped looking at our L&D reports because they don't find actionable insights." The dashboards we built weren't reporting tools. They were decision tools. When a VP can see three projects blocked by the same SME bottleneck, that's actionable. When they see a 42-slide deck about completions, that's not. Are there risks to this kind of transparency? Absolutely. It can feel exposing. But we found that the teams most resistant to dashboard visibility were often the teams with the most hidden bottlenecks. What are we afraid of when we resist transparency?

Trust Conversations

This might be the most important thing we did, and it's the one I almost skipped because it felt "soft." We held explicit conversations about trust. Trust in each other. Trust in the process. Trust in the AI tools.

HBR's February 2026 research found that AI integration can erode interpersonal trust and coordination if leaders don't proactively address it. We experienced this firsthand. When we introduced AI-generated first drafts, some SMEs stopped trusting the content altogether. They reviewed AI drafts more skeptically than human drafts (which is actually appropriate), but they also started second-guessing the IDs who used the tools. That's a trust problem, not a tool problem.

So we put trust on the agenda. Literally. We asked: "Do we trust each other's judgment? Do we trust this process? What would need to change for trust to increase?" Those conversations were awkward. I'm still not great at facilitating them. But they surfaced issues that would have festered for months. Timothy R. Clark's four stages of psychological safety (inclusion, learner, contributor, challenger) gave us a useful framework. Are our teams at the "challenger" stage, where people feel safe enough to push back on bad ideas, including ours?

ATD's 2025 conference sessions on leadership development emphasized psychological safety as a foundational element for effective learning teams, not just for learners, but for the teams building the learning. How often do we apply to ourselves what we teach to others?

AI-Specific Roles in the Review Cycle

Someone had to be responsible for fact-checking AI-generated content against SME source material. That's a new task that didn't exist before. Does every team account for this added step when calculating their efficiency gains?

Prompt Libraries Built Over Months

Standardized prompts for common content types, tested and refined over months. This took significant upfront investment. Our IDs easily spent 40-60 hours each learning effective prompting before we saw consistent results. I had to keep reminding my team and myself that we were doing what Brene Brown calls "embracing the suck." I underestimated this learning curve badly. I expected the team to be proficient in a week, not a month. Josh Cavalier's ATD AI certification program calls this kind of structured approach an "L&D AI Ecosystem," and he's right that it needs to be systematic rather than ad hoc.

And this connects to a mistake I see teams making everywhere: "We gave our team AI tools but no strategic direction. Now we have IDs using AI tools for specialized tasks it wasn't built for." We almost made the same mistake. Without standardized prompts and clear guidance on which tools to use for which tasks, our team defaulted to whatever they'd heard about on LinkedIn. That's not a strategy. That's chaos with a subscription fee.

Changed Quality Gates

AI-generated first drafts need different review criteria than human-generated first drafts. They're grammatically flawless but conceptually suspect. Our SME reviewers had to learn to read differently, looking for plausible-sounding inaccuracies rather than typos and formatting issues.

Consistent Evaluation Throughout

Here's the last piece, and it ties the whole system together: we didn't wait until launch to evaluate. We built evaluation checkpoints into every stage of the pipeline.

Not just "did the module pass QA?" but "is this still aligned with the learning objectives we defined in planning?" at every stage. Were the assessments actually measuring what Bloom's taxonomy says they should? Did the scenarios reflect real job tasks? Was the content culturally appropriate for all three regions?

This will slow things down, and I expected that. But consistent evaluation throughout the process actually saved us time by catching drift early. A module that drifts from its objectives in the storyboard stage is a 30-minute fix. A module that drifts and doesn't get caught until QA is a two-week rework. Which one costs more?

The Training Industry 2026 trends report describes this as "outcomes-led learning," where evaluation isn't a final gate but a continuous alignment check. That's exactly what we experienced.

---

📉 Why 31% Is the Honest Number (and 80% Isn't)

The vendor claims of 80%+ efficiency aren't entirely fabricated. They're just measuring the wrong thing.

If I measure only the time it takes to generate a first draft and compare "human writes from scratch" to "AI generates a draft in 60 seconds," then sure, that isolated task might show a 90%+ improvement. But that task was never the bottleneck. What percentage of our total development time was actually spent on first-draft generation before AI?

Here's what makes the inflated numbers dangerous. Leadership heard from AI vendors that training content can be "easily generated by anyone." I've had to push back on this in budget conversations, and it's uncomfortable because it can sound like you're protecting your job. But the reality is: generating a first draft was never the hard part. SME alignment, review cycles, cultural adaptation, assessment validity, scenario design, that's where the time goes. And AI doesn't meaningfully compress any of those stages yet.

ATD's 2025 State of the Industry data shows the average cost per learning hour reached $165, a 34% increase from 2023. Learning hours per employee dropped from 17.4 to 13.7. Organizations are being asked to do more with less. That pressure makes inflated efficiency claims dangerously attractive. How many of us have cited an 80% figure in a budget presentation because it's what leadership wanted to hear?

Josh Bersin's February 2026 research describes AI as "poised to reinvent the $400 billion corporate training market." I believe that. But reinvention and efficiency are different things. Reinvention means changing what we build and how learners experience it. Efficiency means doing the same thing faster. Are we chasing the right goal?

The LinkedIn Workplace Learning Report 2025 found that while 71% of L&D professionals are experimenting with AI, only 25% have integrated it into routine workflows. That gap between experimentation and integration is where the honest numbers live. And in that gap, 31% starts looking pretty good. What if the teams that cross that gap are the ones investing in systems thinking, not just tool adoption?

---

🧠 What This All Means

Here is what I don't have a clean answer for: what's the right expectation for AI efficiency in L&D? And is "efficiency" even the right metric? I think this is going to vary widely by the tools being used, the team's willingness to learn and adopt and the systems you put in place.

Someone recently asked me: "My CFO wants cost savings data, my CHRO wants retention impact, my VP wants throughput numbers. How do I build one dashboard?" One dashboard isn't the answer. The answer is knowing which story to tell to which audience. The CFO gets cycle time reduction and cost-per-module trends. The CHRO gets learner satisfaction and ramp-time data. The VP gets throughput. Same underlying system, different lenses. If we try to build one report that satisfies everyone, we end up satisfying no one.

Here's what the data suggests as of March 2026. McKinsey reports that organizations investing in AI upskilling see 30%+ productivity improvements. Our 31% aligns with that. The Synthesia report shows that 84% of L&D teams report faster production, but most spend less than 5% of their budgets on AI tools. The METR study found that on complex knowledge work, AI can actually slow people down. And HBR's February 2026 research warns that AI adoption without attention to psychological safety creates predictable team dysfunction.

So maybe the answer is: 20-35% cycle time reduction is realistic for teams that invest in the full system: planning, ego agreements, cadenced workshops, accountability tools, transparent dashboards, trust conversations, consistent evaluation, and (yes) AI tools in the right places. Anything above 50% probably means we're either measuring a single isolated task or we've cut corners on quality that we haven't noticed yet.

LSA Global's research found that strategic clarity accounts for 31% of the difference between high and low-performing teams. I find it notable that our efficiency gain matches that number almost exactly. Maybe what we really built wasn't an "AI-powered content pipeline." Maybe we built strategic clarity. And the AI tools were just one expression of a team that finally knew what it was doing and why.

I've been designing what I'm calling the L&D AI Operating System, a playbook that connects all of these pieces (efficiency audits, delegation, workflow redesign, measurement) into a single practice. The 31% story is one chapter of that larger system. More on that soon.

I could be wrong about this. I'm still figuring out where the ceiling is. What I do know is that our 31% was hard-won, required significant process change, and came with tradeoffs we're still navigating. Would a team starting today, with better models and more established practices, do better? Probably. But I'd bet they'd land somewhere between 25% and 45%, not 80%. I'm going to bet the ones who reach 45% invest more in systems thinking than in tool selection.

What would change in our industry if we all agreed to share real numbers instead of aspirational ones?

---

🎯 The One Thing to Do This Week

Before touching an AI tool, map out the full content development pipeline as a single system, every stage, every handoff, every decision point, and identify the three biggest bottlenecks. At least two of them are human processes (SME availability, review cycles, approval chains), not technology gaps. That map is the starting point for real efficiency, whether or not AI is part of the solution.

---

What efficiency gains are you actually seeing, and how much of it comes from tools vs. systems? I'd love to compare notes, find me on LinkedIn.

-- Eian