What Lewis and Clark Can Teach Us About Leading People Through AI

The macro data on AI and employment keep saying the same thing: no apocalypse. Software developer employment is up. Business software spending jumped 11% in Q4 2025. The economy is absorbing this disruption the way it absorbed spreadsheets, ATMs, and the internet.

But absorbing a technology and actually using it well are different problems. MIT Media Lab’s Project NANDA found that 95% of generative AI projects fail to deliver measurable ROI. EY’s 2025 survey of 15,000 employees found only 5% use AI in ways that fundamentally transform their work. McKinsey’s December 2025 analysis identified the bottleneck — most organizations stop at “literacy” and never reach “adoption,” because adoption requires redesigning workflows, and that demands leadership courage, which most companies lack.

None of this is new. We’ve watched organizations fail to convert new capability into durable proficiency before. What’s useful is that we’ve also watched them succeed — and the pattern of success hasn’t changed in two centuries.

The last time someone had to prepare for unmapped territory

In the spring of 1803, Thomas Jefferson needed Meriwether Lewis ready for territory no American scientist had documented. Lewis was a capable Army officer. He was not a botanist, astronomer, physician, or cartographer. Jefferson had to make him competent in all four, quickly.

Jefferson didn’t hand Lewis a manual and wish him luck. He designed a system. And the architecture maps almost perfectly onto what organizations need to do with AI right now.

• Build the foundation before introducing the skill. Lewis spent two years as Jefferson’s private secretary — not studying science, but learning how Jefferson thought about exploration and the relationship between observation and decision-making. The modern translation: before rolling out AI tools, establish a shared understanding of what AI changes about how work gets done. Not a literacy course. A cognitive framework.

• Target specific gaps with specialist instruction. Jefferson contacted five leading experts in the exact disciplines Lewis lacked — Ellicott for navigation, Rush for medicine, Barton for botany, Wistar for anatomy, and Patterson for mathematics. Three months. No generic curriculum. Every hour on a specific gap. The modern translation: skills audits identifying what people actually can’t do, followed by targeted instruction for specific gaps in specific roles.
• Rehearse before the stakes are real. Lewis spent five months on the Ohio River and at Camp Dubois — applying what he’d learned in progressively challenging conditions before the expedition launched. The modern translation: structured practice environments where people integrate AI into real workflows with support, before those workflows carry full consequences.
• Perform under pressure with tapering support. Once launched, Lewis had Clark, Sacagawea, and the Corps — but no instructors. Support tapered as competence grew. The modern translation: coaching and peer networks that reduce over time. Not permanent training wheels. Not sudden abandonment either.

The results: 178 plants and 122 animals newly described in Western science. Medical care that kept all but one member alive. Celestial observations accurate within four miles. Today, many organizations trying to upskill for AI are skipping every phase except awareness. They’re handing out logins and hoping for the best.

Where the roadmap gets complicated

Brynjolfsson, Li, and Raymond studied 5,172 customer support agents and found that AI assistance boosted productivity by 14% on average, with novice workers seeing a 34% improvement. The tool compressed the learning curve on its own. For structured, repeatable tasks, modern AI tools may lower the training barrier enough that the full Jefferson sequence adds less value than we’d expect.

This is a real finding with a real boundary. Those agents worked on defined tasks with clear success metrics — exactly where intuitive tools shine. The equivalent in Lewis’s world would be a self-correcting compass. Useful for navigation. Useless for deciding whether to engage the Teton Sioux or go around them.

The second challenge is structural. Daron Acemoglu’s analysis estimates that AI will produce no more than a 0.66% increase in total factor productivity over ten years. His argument isn’t about implementation — it’s about direction. We’re using AI for automation when we should be augmenting human expertise. If the technology concentrates benefits in larger firms and educated workers, closing the knowing-doing gap won’t solve the societal problem.

Jefferson couldn’t control the continent’s geography. He could control how prepared Lewis was to navigate it. Organizations can control how they develop their people. They cannot unilaterally control the technology’s structural direction. Both matter. We focus on what leaders can actually influence.

Why brains make this harder than it looks

Amy Arnsten’s research at Yale shows that acute stress weakens the prefrontal cortex’s functions people need most for learning — working memory, cognitive flexibility, and executive control. Career uncertainty about AI triggers exactly this response. Staw, Sandelands, and Dutton’s threat rigidity theory, validated over forty years, predicts the result: narrowed information processing, centralized control, and reversion to routines.

Despite roughly 40% workplace adoption, the St. Louis Federal Reserve estimates only 0.5–3.5% of all U.S. work hours involve AI assistance. And 60% of users admit it takes longer to figure out an AI tool than to complete the task without it. Jefferson built progressive exposure — foundation, instruction, rehearsal, performance — because that’s how human beings actually learn under pressure.

The knowing-doing gap is the most actionable bottleneck in the AI workforce transition. We’ve closed gaps like this before. The pattern repeats.

• Audit gaps, not awareness. Jefferson didn’t send Lewis to “Enlightenment school.” He identified five precise gaps and matched each to the best available expert.
• Match the intervention to the task. For structured work, lighter-touch support with peer networks may be enough. For complex integration — strategic decisions, cross-functional workflows, and ambiguous problems — the full scaffolded approach remains essential.
• Build rehearsal into the timeline. Leading indicators track whether people integrate AI into complex work, not whether they completed a course. Lewis had Camp Dubois. Most organizations have a webinar and a go-live date.
• Address the threat response before demanding new learning. Psychological safety isn’t optional when stressed brains can’t learn. If people feel their jobs are at risk, no training will stick.
• Acknowledge what you can’t control. Structural forces operate alongside implementation failures. Neither lens alone is sufficient. Be honest about the limits of what capability-building can solve.

The frontier is moving. Jefferson sent Lewis into unmapped territory with the best preparation available and explicit instructions to document what he found, especially when it contradicted expectations.

That’s the job. And we’ve done it before.

This work synthesizes research from MIT Media Lab’s Project NANDA, EY’s 2025 Work Reimagined Survey, McKinsey, Brynjolfsson, Li, and Raymond’s study of generative AI in the workplace, Daron Acemoglu’s macroeconomic analysis of AI productivity, Amy Arnsten’s neuroscience of stress and prefrontal cortex function, and the Federal Reserve Bank of St. Louis — alongside the historical training architecture Thomas Jefferson designed for the Lewis and Clark Expedition.

Contact us for a complete list of works cited.