Welcome to this report
How to read this page

This is a practitioner's analysis — part research, part lived experience, part numbers. You can read it straight through or jump to any section at any time using the navigation bar at the top of the screen. It stays visible as you scroll.

The Question Why AI Is Different The Model The 1% Ripple When It Fails Evidence Sources

Charts have hover tooltips. The failure scenarios are interactive — click to expand and engage. The evidence table shows what research supports, what's partial, and what comes from experience rather than citation.

In the Trenches — Michelle Deshotels

Train Every
one.

No Exceptions.

A practitioner's case for universal AI fluency — built from 25 years in the trenches, verified with real research, and backed by the numbers.

Scroll to explore
2,000
Employees modeled across three approaches
3 yrs
Time horizon with snapshots at Year 1, 2, and 3
Return difference between best and realistic approaches
1%
The minimum idea threshold that still justifies everything
$17.7M
3-year net return — universal fluency model vs $2.9M for the realistic selective approach

A question a lot of people are asking right now

A post making rounds in L&D circles made the argument that the fastest way to waste an AI budget is to train everyone on AI. The recommendation: target specific roles with measurable performance gaps. Be strategic. Be surgical. It got a lot of reactions. I disagreed — not because I don't believe in performance consulting, but because I've seen what happens when organizations decide only some people are worth the investment. So I stopped arguing from instinct and ran the numbers.

"Opportunity is missed by most people because it is dressed in overalls and looks like work."

— Thomas Edison

I have spent 25 years working inside organizations — regulated healthcare, financial services, fintech, contact centers. I have talked to the people doing the work when nobody else thought to ask them. I have seen what gets left on the table when organizations decide that only some people are worth the tool. I have also seen what happens when you give everyone access, build the infrastructure to catch what surfaces, and get out of the way.

This is not what I read in a book. This is what I have watched happen, in real organizations, under different CEOs, in multiple industries, repeatedly. The pattern is consistent enough that I decided to stop saying "I've seen this" and start showing what the numbers say when you actually model it out.

From the Trenches

I have worked inside organizations where some of the most elegant operational solutions came from people no one would have called a high-potential. They built tools that standardized outputs across entire departments. They created workarounds that turned a 20-minute process into a 3-minute one. Nobody asked them to. Nobody identified them in a Hi-Po program. They just knew the work and found a better way. When only some people get the tool, you only get some of the ideas. The rest stay locked inside people the organization decided not to ask.

This isn't Salesforce. This is the phone.

The argument for selective training makes more sense for specialized tools — a CRM platform, a financial system, a project management suite. AI is not that. AI is infrastructure — the same way the telephone was infrastructure, the same way email is infrastructure. It doesn't belong to a department. It doesn't respect a job title. Deciding which employees need to understand it is like deciding which employees need to understand email.

50–55%
of U.S. jobs will be reshaped by AI in the next 2–3 years
BCG, 2026
23pt
advantage when both employees AND leadership are trained — not just one group
Workforce study, Vision Monday 2025
$15K+
wasted per employee per year due to inefficient ways of working
Eagle Hill Consulting
<25%
of organizations actually equip frontline workers to make data-driven decisions — despite 90% saying it matters
ThoughtSpot research

There is also a risk that extends well beyond the workplace. When people use AI without understanding what it actually is — that it generates based on pattern matching, not retrieval of verified facts, that it hallucinates with complete confidence, that it reflects the biases baked into its training data — bad information doesn't stay at the desk. It travels home. It travels through social networks. It reaches people outside the workplace who still consume and share information through smartphones and social networks.

The first place most people will encounter AI in a structured, intentional way is at work. And when they go home and talk to their families — their parents, their kids, their grandparents — they carry what they learned with them. That's why I loved using cybersecurity content designed for the whole household. One program trained employees on phishing and social engineering. But the same content had activities for kids who didn't even have a phone yet, guidance for seniors about password safety, simple checks for grandparents who forward everything in their inbox. Many cybersecurity programs now treat household behavior as part of the threat surface, recognizing that the risk lives in living rooms as much as in boardrooms. AI literacy works the same way.

"Bernie Madoff's fraud didn't stay in the boardroom — it flowed through social networks to individuals, families, and charities who heard about steady returns and trusted the narrative. Misinformation from confidently misused AI spreads the same way — through social ties and trusted channels — which makes it a citizenship problem as much as a workforce-productivity problem."

— Supported by UT Dallas affinity-fraud research (15) and Harvard Misinformation Review plus NPR/Reuters reporting on AI misinformation (12–13).

This is not a doomsday argument. It is not a conspiracy. It is the same logic that made us decide everyone needs to understand how email works, how to recognize a suspicious link, how not to give their password to someone who sounds official on the phone. AI is that foundational now. The question isn't whether people will encounter it. They already are. The question is whether they'll encounter it with enough understanding to know what they're dealing with.

Three companies. Same starting point. Three years.

To stop arguing from instinct and start arguing from numbers, I built a model. 2,000 employees. Conservative assumptions throughout. Three approaches to AI training. Here's what happens at 12, 24, and 36 months — and why the gap between them keeps growing.

Model Assumptions — Read This First

  • Organization size: 2,000 employees across customer-facing, knowledge worker, and administrative roles
  • Average fully-loaded labor cost: $75,000/year ($36/hour)
  • Working days per year: 235 — Annual turnover: 15% (300 people/year)
  • Time savings: Modeled conservatively at 20 min/day Year 1, scaling as fluency deepens
  • Innovation yield: Based on workflow improvements becoming documented process changes
  • These are illustrative models, not guaranteed outcomes. Conservative assumptions applied consistently across all three scenarios so the comparison is apples to apples.
Company A — Ideal
Targeted. Executed perfectly.

Hi-Po first approach. Trains identified high-potentials and selected "AI-impact roles" first. Executes exactly as planned. Everyone trained by Year 3. This is the best-case version of the selective approach — the one the argument assumes will happen.

Company A — Realistic
Good intent. No infrastructure.

Same plan as Ideal, but follows what enterprise training data actually shows. Something else becomes important at month 14–16. Momentum stalls. 48% of the organization never trained by Year 3. This is what usually happens.

Company B — Matisse
Universal baseline. Built to compound.

Full fluency push in Year 1. Champions network. AI hangouts. Prompt library. Shoutouts published org-wide. Cross-functional coordination. No performance pressure in Year 1. New hires onboard into a living system from day one.

% of Organization Actually Trained — By Year
This is where Company A Realistic tells its most honest story. Intent and infrastructure are not the same thing. Hover over any bar for the exact figure.
Co. A — Ideal
Co. A — Realistic
Co. B — Matisse
Year 1
Co. A Ideal
25%Year 1: 500 of 2,000 trained (Hi-Po + targeted roles)
Co. A Realistic
17%Year 1: ~340 actually complete — 68% of intended group
Co. B Matisse
100%Year 1: All 2,000 trained — full fluency push in 5 months
Year 2
Co. A Ideal
65%Year 2: 1,300 cumulative trained
Co. A Realistic
38%Year 2: Priority shift hits — only ~760 cumulative trained
Co. B Matisse
100%Year 2: Still 100% — fluency deepening, metrics now active
Year 3
Co. A Ideal
100%Year 3: Full 2,000 trained — as planned
Co. A Realistic
52%Year 3: Only 1,040 of 2,000 ever trained. 960 people never reached.
Co. B Matisse
100%Year 3: Full compounding. New hires onboard into a living system.

Company A Realistic ends Year 3 with 48% of the organization never trained. 960 people. Some of them had the next workflow improvement. You will never know which ones.

Net Annual Returns — All Three Approaches
Base fluency gains (time saved, error reduction, retention, new hire speed) modeled conservatively. Hover for details. Bars scaled relative to Year 3 Matisse as 100%.
Co. A — Ideal
Co. A — Realistic
Co. B — Matisse
Year 1 Net Return
Co. A Ideal
$1.03MYear 1: $1,025,900 — time savings + error reduction for 500 trained
Co. A Realistic
$446KYear 1: $445,520 — lower fluency, no reinforcement infrastructure
Co. B Matisse
$2.46MYear 1: $2,459,480 — 2,000 trained, risk reduction, retention gains begin
Year 2 Net Return
Co. A Ideal
$2.66MYear 2: $2,661,300 — 1,300 trained, innovation still isolated
Co. A Realistic
$978KYear 2: $978,280 — stalled rollout, no cross-functional spread
Co. B Matisse
$5.77MYear 2: $5,766,000 — fluency deepening, champions surfacing innovations
Year 3 Net Return
Co. A Ideal
$4.27MYear 3: $4,270,500 — everyone trained, but no infrastructure to compound it
Co. A Realistic
$1.47MYear 3: $1,474,480 — 48% never trained, ideas dying at source
Co. B Matisse
$9.49MYear 3: $9,491,500 — full compounding, process improvements embedded org-wide
Cumulative Returns Over 3 Years — The Compounding Gap
This is where the infrastructure argument becomes impossible to ignore. The gap between approaches doesn't narrow over time. It widens.
Co. A — Ideal ($7.96M total)
Co. A — Realistic ($2.90M total)
Co. B — Matisse ($17.72M total)
$0 $9M $17M Year 1 Year 2 Year 3 $17.7M $7.96M $2.9M

Cumulative net returns. Matisse's Year 1 return alone ($2.46M) exceeds Company A Realistic's entire 3-year total ($2.90M). The gap is not about which company worked harder. It's about which one built infrastructure.

Full 3-Year Comparison
Green = strong. Amber = partial. Red = weak or absent. Dollar figures are illustrative models — conservative assumptions applied equally to all three.
Metric Co. A — Ideal Co. A — Realistic Co. B — Matisse
% Trained by Year 125%17%100%
% Trained by Year 265%38%100%
% Trained by Year 3100%52%100%
Total 3-yr delivery cost$717,600$655,000$1,408,520
Net Year 1$1,025,900$445,520$2,459,480
Net Year 2$2,661,300$978,280$5,766,000
Net Year 3$4,270,500$1,474,480$9,491,500
3-year total net return$7,957,700$2,898,280$17,716,980
3-year ROI11×4.4×12.6×
Innovation infrastructureNoNoYes
Risk reduction (hallucination/bias)PartialMinimalFull
New hire integrationDelayedInconsistentDay one
Knowledge asset createdNoNoYes

This is what separates "we trained everyone" from "we built a system."

The training itself is assumed. What most organizations skip is everything around it. These eight elements are what turned training into a system — each one feeding the next.

None of this runs without someone whose only job is to make it run. Not a committee. Not a champion with a day job. One person — sitting inside the organization, not outside it — whose explicit responsibility is cross-functional coordination, visibility, and continuity. That is the difference between infrastructure and a good idea that fades.

🛡️
Year 1 — No Performance Pressure

The condition that makes everything else possible. Fluency first. No metrics tied to AI use in Year 1. Psychological safety intact. People explore, make small mistakes, build real understanding. Accountability comes in Year 2 — after the foundation exists.

🔥
AI Hangouts

The first point of contact. Open, recurring, themed sessions — email drafting, transcript tactics, power prompts. Anyone could attend. No registration required. This is where fluency starts building visibly and communally.

📚
Prompt Library

What the hangouts produce. The best thinking gets captured here. Living document. Crowd-sourced. Not generic internet downloads — built in-house, reviewed by champions, specific to this organization's systems and language.

📣
AI Shoutouts

People submit what they did with AI, what department, what task, how much time it saved. Published org-wide. Visible. Celebrated. The prompt library stops feeling like templates — it starts feeling like proof of what's possible.

Recognition and Incentives

What keeps the shoutouts coming. Small rewards, public acknowledgment, people seeing their name and their work surfaced to the whole organization. Psychological reward drives more sharing than monetary value. People contribute because they're seen and thanked.

🏆
Champions Network

The human infrastructure that validates, curates, connects, and carries everything across departments. Every team has one. They meet cross-functionally. Ideas don't die in the team they were born in — they travel.

📊
Analytics Layer

Champions now have data. Time saved by task, adoption patterns by team, trends in poor prompts feeding coaching updates. The system starts informing itself. Champions bring specific examples to team meetings — not generic encouragement.

🆕
New Hire Integration

The proof point. A new hire walks in and everything above already exists. They don't onboard into a blank slate. They onboard into a living system — prompt library, shoutouts, hangouts, champions. Speed to fluency is measurably faster from day one.

What if only 20 people have the idea?

The objection I hear most often: not everyone will contribute something meaningful. Correct. That is not the point. The 1% framing is not a pessimistic estimate — it is a deliberate floor. I want to show you what happens even at the absolute minimum, before I show you what actually happens when the infrastructure is there to catch more.

The Ripple Effect — In Business Terms
CO. A REALISTIC CO. A IDEAL CO. B MATISSE Idea surfaces. Ripple stops here. Travels one layer. Then stops. Becomes process. Crosses the lake.

In business, a ripple is what happens when one person's idea doesn't stay with that person. It gets caught, validated, documented, and embedded into how the work gets done — until everyone is operating on a better process, whether they know where it came from or not. The size of the ripple is determined by the infrastructure that catches it.

This Is Not Theoretical — Kruze Consulting

At Kruze Consulting — a fully remote accounting and finance firm of 170 people — I rolled out an AI program built exactly like the Matisse model. We captured 20+ documented AI shoutouts in just the first couple of months. The shoutout tracker started from the beginning of the rollout, during busy season, when people were still ramping up and getting used to the idea. The prompt generation form produced dozens of real, usable prompts within weeks — which champions then verified and pushed into the library for everyone to access. And this was the floor. We were just getting started. Once people see their work published, once they see someone else's idea and recognize they have one too, it compounds. That's not speculation. That's what happened. The 1% is never going to stay at 1%.

What 20 Ideas Are Worth — When They Become Process
These ideas don't ripple by choice. They become the way the work gets done. Process change affects everyone downstream whether they know where it came from or not.
10 Workflow-Level Ideas
People affected per idea~50
Time saved / person / day20 min (conservative)
Value per idea / year$142,000
10 ideas / year total$1,420,000
10 Dept/Function-Level Ideas
People affected per idea~200
Time saved / person / day20 min (conservative)
Value per idea / year$568,000
10 ideas / year total$5,680,000
Same 20 ideas. Three different fates. Hover for detail.
The quality of the ideas is identical across all three organizations. What differs is what happens next.
Year 1 — Value Captured from the 20 Ideas
Co. B Matisse
Becomes process org-wide$7,100,000 — full ripple captured. Ideas become training. Everyone benefits.
$7,100,000
Co. A Ideal
 $710,000 — 10% captured. Idea travels one layer. No infrastructure to carry it further.
$710,000
Co. A Realistic
 $213,000 — 3% captured. Great point. Then the big customer called. Idea gone.
$213,000
How the ripple compounds — or decays — over 3 years
$0 $10M $28M Year 1 Year 2 Year 3 $28.3M $2.2M $594K

The same 20 ideas. Company A Realistic's ripple doesn't just stall — it decays in Year 3 as the program loses momentum and people move on. Company B Matisse compounds because the ideas are baked into training, onboarding, and standard operating procedure.

Company A Realistic captures 3% of the ripple value from the same ideas. The idea surfaces in a meeting. Someone says "great point." Then the big customer calls. That's where good ideas go to die — not because the idea was bad, but because there was no one whose job it was to catch it.

The 1% is never only 1%. But even if it were — even at the absolute floor — the number says the same thing. Train everyone. Build the infrastructure to catch what surfaces. Get out of the way.

— Michelle Deshotels

The three failure scenarios — and what they actually tell us

There are scenarios where training everyone on AI produces bad outcomes. I'm going to name them honestly. Each one has an interactive question inside — take a guess before you read the answer. Then I'll show you why none of them are arguments for training fewer people.

Failure Scenario 01
You train everyone but the training is wrong.
This is a design failure — not an argument against training everyone.
+

You teach people AI is a search engine. You don't teach hallucination — that AI generates responses based on pattern matching, not retrieval of verified facts. You don't teach bias — that early AI systems were trained on internet data full of human assumptions and prejudices, including some deeply wrong ones. You don't teach prompt construction. You create 2,000 people who are confidently wrong. That is worse than 2,000 people who are appropriately skeptical because they don't understand the tool. Bad AI literacy is more dangerous than no AI literacy because it comes with false confidence and no brakes.

Quick check: What's the most dangerous outcome of bad AI training?
Close — but no. If people don't use it, you lose the return. That's expensive. But you can fix it. The dangerous outcome is the next one.
Exactly right. Confident and wrong is the most dangerous state. It spreads. It informs decisions. It gets shared with people who have no reason to question it. That's why the training has to be correct before it's universal.
Budget is a real concern — but wasted budget can be recovered. Confidently wrong information that's already spread through an organization is much harder to walk back.
This is a training design failure. Fix the design. Don't restrict the audience.
Failure Scenario 02
You train everyone but build no infrastructure to catch what surfaces.
This is an infrastructure failure — not an argument against training everyone.
+

The ideas come up and go nowhere. A champion who gets pulled back to their real job the moment the big customer calls. A prompt library that nobody owns. A committee that meets monthly to talk about AI ideas — but it's everybody's side job, not anyone's actual responsibility. When it's everybody's job, it's nobody's job. That is where good ideas go to die, every single time, in every industry, in every organization I have ever worked in or worked with. The training wasn't the problem. What surrounded it — or didn't — was the problem.

How much of the 1% ripple value does an org without infrastructure actually capture?
Feels reasonable — but no. Without someone whose job is to catch and carry ideas, most of the value evaporates long before it reaches 50%.
Closer — but still too optimistic. "Great point" in a meeting is not infrastructure. Ideas need a home or they die.
Right. The model shows Company A Realistic captures approximately 3% of the ripple value from the same ideas. The idea surfaces. Someone says "great point." Then something else happens. That's it.
Build the infrastructure before you train. Champions, prompt libraries, shoutout systems, cross-functional visibility — these are not nice-to-haves. They are the mechanism that turns training into return.
Failure Scenario 03
You train everyone and then punish people for using it wrong.
This is a change management failure — not an argument against training everyone.
+

Someone uses AI, makes a mistake, gets called out publicly. Now nobody uses it. You have trained 2,000 people to be afraid of the tool. Research on AI adoption links poor communication, exclusion from decisions, and weak psychological safety directly to worse adoption outcomes. This is exactly why a Year 1 with no performance metrics is not a soft decision — it is the structurally correct one. People need room to explore, make small mistakes in low-stakes environments, and build genuine fluency before you hold them accountable for outcomes. Fluency first. Accountability after. The sequence is not optional.

When should performance metrics tied to AI use be introduced?
This feels right — but the research and the lived experience say otherwise. Accountability before fluency creates fear, not adoption. People stop using it publicly and start hiding their mistakes instead of learning from them.
Correct. Build the fluency foundation first — ideally a full year with no performance pressure. Then introduce accountability once people have the foundation to meet it. The sequence is the design.
Voluntary-only creates an uneven organization — some people adopt, most don't, and the ones who don't get left behind. Eventually AI fluency becomes a capability gap that costs real money. It needs to be required. Just not punished while people are still learning.
Fluency first. Accountability after. That's not softness — that's the correct sequence for sustainable behavior change.
The Pattern Across All Three
It is easy to blame training. Most of this has nothing to do with training.
  • Bad design fails. Wrong content, wrong framing, no hallucination awareness, no prompt foundation — this produces confident and wrong, which is worse than nothing.
  • No infrastructure fails. Training without champions, without a prompt library, without someone whose job it is to catch what surfaces — good ideas die in the meeting they were born in.
  • No psychological safety fails. Punish people for learning and they stop learning in public. Adoption goes underground and you lose the visibility you need to course-correct.

Training everyone doesn't fail. The conditions around it fail. These are solvable problems with known solutions. None of them are reasons to limit who gets access to the tool.

The dental analogy nobody wants to hear.

Your tooth starts to ache. It isn't going away. The sooner you address it, the better off you are — and the cheaper it will be. The longer you wait, the more expensive the solution becomes. This is not a metaphor. This is how delayed intervention works in every context. AI training is not different.

🦷
Act Now
A filling
Build the infrastructure. Train everyone. Do it right the first time. Manageable cost. Compounding return. The gap between what you spend and what you gain keeps widening in your favor.
😬
Wait 6–12 months
A root canal
Partial rollout. Catch-up training. Half the org on different versions of the same tool. No infrastructure. Higher cost to fix what should have been built from the start.
😰
Year 2–3
An extraction
The AI initiative gets shelved. Another one starts. You spend the budget twice. The gains never compound. People stop believing these things work. That last part is the most expensive.
💸
After the Extraction
The implant
Now you need to replace what you extracted. Higher cost. Longer timeline. Less trust. And the 960 people who were never trained? Still there. Still untrained. Still paying the gap forward.
Organizations Have Seen This Play Out

This pattern is not unique to AI. It repeats with every enterprise software rollout, every new process initiative, every change management effort that gets real traction and then loses its champion. Organizations invest in building something that works — training, infrastructure, accountability — and then something shifts. A budget cycle. A leadership change. A "more urgent" priority. The investment stops. The gains stop. And then months later, someone notices the gap and the conversation starts over from scratch, usually at a higher cost and with lower goodwill from the people being asked to go through it again. You don't get a refund on the momentum you lost.

What held up, what needed adjusting, and what comes from experience.

I started with instincts built over 25 years. Then I ran the claims through independent research before publishing anything. Here's exactly what the evidence supports, what's partial, and where I'm speaking from experience rather than citation. I'd rather be honest about the distinction than cite something that doesn't hold.

Claim Research says What the evidence actually shows
Training everyone outperforms selective training for AI adoption Supported Organizations training both employees and leadership show a 23-point advantage in high-value AI outcomes. Broader transformation participation correlates with better results. Best evidence-based framing: universal baseline plus role-specific depth.
AI will reshape most jobs across most industries Partial BCG: 50–55% of U.S. jobs reshaped in 2–3 years. Microsoft research shows high AI applicability across diverse occupations. "Every job equally" is overstated — impact intensity varies by role. The direction and the breadth are real.
Underutilization of enterprise technology is measurably costly Supported Organizations waste $15,000+ per employee per year on inefficient ways of working. Underutilized licenses, inactive users, and poor governance create direct and hidden costs. Well supported across multiple sources.
Frontline workers are rarely included in decisions about their own work Supported Only 23% of frontline workers believe senior leaders understand their day-to-day. 90% of organizations say frontline decisions matter; fewer than 1 in 4 actually equip frontline workers to make them. 85% of executives believe their orgs are bad at diagnosing problems.
Enterprise training often fails to reach its full intended audience Partial No specific failure percentage is sourced to a clean primary study — and this report does not claim one. What is supported: only 2% of employees are directly involved in a typical enterprise transformation. Involving 7% or more doubles positive outcomes. The direction is real. The specific figure is not.
Excluding people from AI training creates fairness and trust problems Supported 2025 Nature research links poor communication, exclusion from decisions, and weak psychological safety with worse adoption outcomes. MIT Sloan identifies this as one of the most common AI strategy failures.
AI misinformation spreads through social networks, not just organizations Supported Harvard misinformation research and NPR/Reuters AI reporting document how AI-generated misinformation spreads through social channels. Research links overconfidence in the medium (polished AI output) with sharing behavior. Citizenship framing is increasingly supported by policy analysts and researchers.

What Comes From Experience, Not Citation

Several claims in this report come from 25 years of working inside organizations — not from published studies. I have named them clearly throughout. They are not anecdotes used to replace data. They are examples of patterns the research also identifies. I have lived them across multiple industries, multiple organizations, and multiple leadership environments. The experience and the evidence point the same direction.

The cost that has no line item

What the research cannot yet fully capture: the specific dollar value of the idea that never surfaced because the person who had it was never given the tool, never asked, never in the room. That line item doesn't exist. It never will. But that doesn't mean the cost isn't real.

Michelle
Deshotels
Organizational Performance Architect

San Diego, CA
25 years across regulated healthcare, financial services, fintech, and contact center operations

michelledeshotels.com →

I am not a thought leader who reads about organizations. I am a practitioner who has worked inside them — in QA, in operations, in L&D, in AI enablement, across industries that don't forgive mistakes. I have built training programs from scratch. I have implemented enterprise software. I have talked to the people doing the work when nobody else thought to ask them.

This report is not a framework. It is not a methodology with a catchy acronym. It is what I have seen, what the research supports, and what the numbers show when you run them honestly — without assuming the conclusion before you start. I went looking for data that would prove me wrong. The data pointed the same direction I was already facing.

I built and ran the Matisse model in a real organization. 100% AI curriculum completion. 50% voluntary adoption. 70–80%+ monthly LMS engagement sustained over time. I know what it takes to build something like that. I know what the infrastructure requires. And I know that none of it happens by accident.

If you want to talk about what it actually takes to build AI fluency across an organization — not the version that sounds good in a proposal, but the version that holds up — I am not hard to find.

Where the cited data comes from

All sources verified independently through Perplexity research prior to publication. Claims that could not be verified against primary sources are clearly marked in the Evidence section as coming from experience rather than citation.