Enterprise Field Notes · Issue #5

The Definitions Gap

A reader called me about the 6%. We talked for an hour. Here is what he saw that I had missed.

By Ben Pickett · May 19, 2026

New here? Subscribe to Enterprise Field Notes, one new issue every week.

By Ben Pickett

I sell AI for a living. So when I tell you that the conversation most enterprises think they are having about AI is not actually the conversation they are having, take it as a confession, not a critique.

By the end of this you will have a name for the gap, the practice that closes it, and the four habits the six percent built first, in the order they have to come in.

Last week I wrote about the gap. Six percent of enterprises pull real value out of AI. Ninety-four percent do not. I argued the difference was workflow redesign. That is correct. It is also not the full answer.

The full answer sits upstream of workflow redesign. The six percent did not have better engineers. They did not have bigger budgets. They did not have newer models. They had better definitions. They were having a different conversation than the rest of the field is having, and that one upstream thing made every other thing they did cheaper. The redesign worked because the definitions were already in place. Without the definitions, the redesign is a copy of what was there before.

That is the whole piece. The rest is the evidence and the practice.

One of you wrote back and then called with the thing I cannot stop thinking about.

His name is Brian. He works at a major US telecom and is in the middle of an enterprise-scale AI push. Before he got to his analogy he told me something else. The six percent number from last week’s issue had stuck with him. “You know how some numbers just stick,” he said. That one did. So when he told me what he wanted to talk about, he was already framing it as an explanation of the six percent.

Here is the way he framed it in the call. Most of the conversations they are having about AI inside the company are Spanish and Portuguese conversations. The words sound alike. They do not mean the same thing. And nobody in the room knows they are translating.

This piece is what I owe him for that observation. The six percent did not have better engineers. They had better definitions. Once you see it, you cannot unsee it.

The Spanish-Portuguese frame is more literal than it sounds.

Spanish and Portuguese share roughly eighty-nine percent of their vocabulary. To a casual ear they sound like the same language. To a researcher counting comprehension under controlled conditions, native speakers of one understand the other at fifty to sixty percent. The number was established in a 1989 Hispania study by John B. Jensen, and follow-up work since has held the range. The overlap is the words. The gap is the meaning the words carry. And the gap is always biggest in the most common, most decision-driving words.

The study found something else worth keeping. Comprehension between the two languages is asymmetric. Portuguese speakers understand Spanish more readily than Spanish speakers understand Portuguese. Not because one language is harder. Because one set of speakers has had more exposure to the other. The ones who lean in to a second language pick it up faster. The ones who do not, do not.

That asymmetry shows up everywhere people are trying to learn something new.

Propensity to learn is the hidden variable in every AI rollout.

When Brian and I got on a call, it went long. He told me about a friend whose teenage daughter taught herself Japanese to fluency. No formal program. No tutor. No textbooks. She just leaned in and the language came.

Then we talked about vendor engineers and how they get stuck on the same workflow process. Even when Brian asks for an AI redesign, what comes back is the old workflow inside a new tool. He has had to be deliberate and forceful about making sure the redesign is actually a redesign, not just a copy.

The trade press has a name for this pattern now. SDTimes called it using extraordinary technology to preserve ordinary behavior, then calling that innovation. Most enterprise AI initiatives die quietly inside that gap, between the new system the slide deck described and the old process the team actually built underneath it.

The daughter went toward the new thing. The engineers carried the old thing into the new one.

The first group teaches itself. The second group needs the new shape drawn for them, or they will keep drawing the old one.

This is the part most enterprise AI rollouts miss. They are designed for the first group. They assume the people with curiosity and good instincts will figure it out. And the people who can, do. They become the six percent. Everybody else is either waiting for the dictionary that never gets handed out, or rebuilding the same old process inside the new tool, or both.

The six percent are not better people. They are not even necessarily smarter. They are the ones for whom neither the lack of shared definitions nor the gravitational pull of old patterns was a barrier. They figured it out anyway. Most people will not. Most people need the language laid out first and the new shape of the work drawn next, and most enterprise AI programs do neither, because the people running the programs are themselves in the lean-in group and forget that not everyone is.

OpenAI’s 2025 enterprise report found the top five percent of users send six times more AI messages than the median employee. EY’s Work Reimagined survey found eighty-eight percent of workers use AI at work, but only five percent use it in transformative ways. The six-percent shape at the firm level shows up inside the firm too. A small cohort runs ahead, and the rest stand still until someone draws them a new shape.

If your AI rollout is reaching the same fifteen percent of your workforce that always volunteers for new things, you are not running an AI strategy. You are running a self-selection mechanism. The six percent of companies that pulled real value out of AI built for the other eighty-five percent.

What the six percent actually fixed.

Here is the real argument of this piece. Definitional alignment is the upstream input every other AI initiative depends on without naming. Platform selection. Training. Talent. Workflow redesign. Even the data quality work. They are all silently translating before they decide anything. Most programs are paying that translation cost in every downstream decision and have no line item for it.

RAND’s 2024 study of AI program failure ranked five root causes. The number one cause was business stakeholders misunderstanding or miscommunicating the problem. RAND said it causes more AI projects to fail than any other factor. Data quality is the most-cited competing upstream factor, and it is real, but it sits downstream of this one. You cannot have data quality when the teams using the data cannot agree on what the data represents. The same is true of talent strategy, vendor selection, change management, and use-case prioritization. Every one of them is silently translating the word AI through someone’s private dictionary before any decision gets made.

Gartner has been tracking what they call agentwashing in 2025, the drift of the word agent across vendors, products, and customer expectations. The market is now confused about whether agentic AI is a real category, a marketing term, or a fork of the same word. The same thing is happening at industry scale that happens in your conference room. Three meanings, one word, no shared dictionary. The thing you are buying does not match the thing you think you are buying.

The definitions slide.

Before AI strategy landed on my plate, I was running an automation and efficiency program at a Fortune 500 global retailer. That is where I learned to walk in with a definitions slide first.

Different VPs owned different parts of the business. Each had built up their own working vocabulary for the same handful of words. Automation meant one thing to ops, another to finance, a third to engineering. Efficiency was worse. Every cross-functional meeting started with the same first thirty minutes of trying to figure out whether we were arguing or agreeing. Most of the time we were doing both, about different versions of the same thing.

So I started walking in with a slide. For this conversation, here is what we mean by these words.

The first time I did it I felt ridiculous. I was presenting to senior leaders who had been in their fields for thirty years. They did not need a glossary.

Ten minutes in I found out they did.

Before the slide, getting eight senior leaders to come away from a meeting with the same understanding of the content and the decisions was difficult. Sometimes impossible. After the slide, the same meeting was efficient. People came away aligned because they had spent the first five minutes agreeing what the words meant. The meetings were also less painful, which mattered more than I expected. People stopped dreading the cross-functional review. They started bringing harder questions, because the easy ones were resolved on the first slide.

When AI strategy showed up on my plate, I brought the same habit.

AI made the problem acute, not new. New terminology lands every six months. Old terminology gets reused for new things. Each VP’s scope still shapes the meaning they attach to the word, and now the word might have shifted under everyone’s feet since the last meeting. The senior leaders in those meetings were some of the best technologists I have ever worked with. The definitions had drifted because the field had moved faster than the conversations could keep up with. Every six months the same word meant something new at the frontier, and not everyone read the frontier papers. So we worked from the meaning each of us had picked up the last time we encountered the term. We all had different last times.

The first AI strategy meeting looked exactly like the automation meetings used to look. The SVP of Ops meant one thing by agent. He meant a piece of software that took an action on his behalf inside his enterprise systems. The CIO meant another. He meant a Slack bot that summarized threads. The data science lead meant a third. He meant a model with a particular kind of memory and tool access. We were all using the same word. We were not all talking about the same animal. And we were about to commit budget to a roadmap none of us had actually agreed on.

After that meeting I added agent, copilot, model, fine-tuning, RAG, orchestration, and governance to the slide. The shape did not change. Just the vocabulary did. Every time I started, I felt ridiculous. Every time, by the end, somebody thanked me for it. Usually quietly, on the way out.

Workflow redesign is the proximal cause. Definitions are the precondition.

McKinsey’s 2025 State of AI report identified fundamental workflow redesign as the single highest correlated factor with EBIT impact from AI. McKinsey’s own framing was blunt. The most common AI use cases focus on task acceleration, not workflow transformation. Companies are layering tools onto existing workflows instead of redesigning. The six percent of organizations that captured significant value from AI were 3.6 times more likely to pursue transformational change. Only twenty-one percent of organizations using generative AI have redesigned even some of their workflows. The other seventy-nine percent are layering AI onto processes built before AI existed.

The framing has critics. Some argue survivorship bias inflates the gap, and that heavy AI spend correlates with outperformance rather than causing it. The directional finding that redesigners outperform is corroborated by BCG and MIT Sloan analyses, but the exact six-percent number is a snapshot of a moving target. The argument of this piece is about the shape of the gap, not the precise decimal.

Boston Consulting Group frames the same point in resource terms. BCG’s now-widely-cited 10–20–70 rule for AI transformation: ten percent of the work is algorithms, twenty percent is technology and data, seventy percent is people and processes. Most enterprises invert the ratio in their spending and their attention. They put the money on the ten percent and wonder why the seventy percent is not delivering.

The numbers line up. The lever is workflow redesign. The lever is not getting pulled.

But that is the proximal answer. The harder question is what makes workflow redesign possible in the first place.

You cannot redesign a workflow when the people in the room do not share a working definition of what they are redesigning around. You cannot train people on a tool whose use cases they cannot describe. You cannot get adoption without training. You cannot justify the redesign without adoption. The chain breaks at the first link.

Drifting definitions. Confused training. No task awareness. Workers bypass the tool. No adoption. No business case for redesign. AI bolted on. Incremental gains at best. The ninety-four percent.

The six percent are the six percent because they fixed the first link.

Klarna is the case study most people get backwards.

Klarna did the redesign. In late 2024 they replaced roughly seven hundred customer service agents with an AI agent. Their workforce dropped from five thousand five hundred to about three thousand four hundred. In its first month the AI handled 2.3 million conversations. Resolution time fell from eleven minutes to two minutes. Eighty-two percent faster responses. Twenty-five percent fewer repeat inquiries. Ten million dollars in annual savings. On every dashboard it looked like the perfect AI workflow redesign. The CEO said publicly the AI was doing the work of seven hundred people.

Six months later customer satisfaction collapsed.

The CEO publicly acknowledged the company had pushed too hard. Klarna started hiring humans back. The new model is hybrid. AI handles about two thirds of inquiries. Humans handle the third where nuance, emotion, and multi-step resolution actually matter.

Here is what I take from the Klarna story. They did not fail at workflow redesign. They failed at definitions. The leadership team did not share a precise definition of what the AI could actually handle. Volume yes. Complexity no. They optimized the boundary in the wrong place because nobody had named what was on each side of the boundary. The redesign was real. The vocabulary that should have driven the redesign was not.

If the same leadership team had started with a definitions slide, the question “what counts as a case AI can handle” would have been on the table from day one. The boundary would have been drawn explicitly. The reversal would not have been necessary. Klarna saved ten million dollars in year one and spent a much larger amount in year two repairing customer trust and rebuilding the human side of customer service. The math on definitional clarity, after the fact, is very good.

This is not a new problem. Software has been here before.

In 2003, Eric Evans published Domain-Driven Design: Tackling Complexity in the Heart of Software. The book introduced a concept Evans called Ubiquitous Language. The argument was that the most expensive failures in software projects come from translation between two vocabularies. Business stakeholders speak the language of their domain. Engineers speak the language of code. The cost of translating between the two is so high that important concepts get lost on the way, and the software ends up doing the wrong thing accurately. Evans’s prescription was simple. Build one shared, rigorous language for the project. Use it in conversations, in documents, in code. Refuse to translate.

Software has spent twenty years learning this lesson. Most agile teams now treat ubiquitous language as table stakes. The vocabulary of the product, the team, and the codebase are aligned by deliberate effort. When the words drift, the work goes wrong. The discipline is to keep pulling them back into alignment.

Engineers building AI systems have begun to apply Ubiquitous Language to LLM and agent design. Eric Evans himself made the connection in a 2024 InfoQ keynote, and a small set of practitioners have written about it since. What has not yet shown up at scale is the same discipline applied one level up, in the room where business stakeholders, product owners, and senior leaders argue about AI strategy before the first line of code gets written. That is where most enterprise AI conversations still drift. That is where the word agent means three different things to three people who are about to commit budget. That is where the six percent quietly built the habit first.

The six percent figured this out by accident, or by hire, or by exposure. The rest of the field can figure it out on purpose.

Healthcare has been here too.

Software is not the only place this lesson has been learned the hard way. Healthcare has spent the last twenty years and tens of billions of dollars on the same problem at much larger scale. The persistent number one barrier to value capture from electronic health record investments is not the software, the workflow, or the user experience. It is semantic interoperability. Terminology standardization between providers, payers, and vendors. Two decades of national investment have not been enough to overcome a vocabulary drift problem that nobody named upstream. SAGE’s 2025 scoping review concluded the same definitional gaps identified in the early 2000s are still defeating EHR ROI in 2025.

AI is now standing where EHR stood in 2005. Enterprise AI rollouts in 2026 look exactly like EHR rollouts in 2006. Vendor selection. Big budget. Workflow integration. Training. Change management. And a vocabulary mess underneath that nobody is naming. We do not have twenty years to figure this out.

The regulatory clock is also already running.

Article 4 of the EU AI Act came into force on February 2, 2025. The text is short. Providers and deployers of AI systems must ensure a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of the systems, taking into account their training, the context the AI is used in, and the people the AI is used on.

Fines under Article 4 are not yet a high-profile enforcement priority. No safe harbor either. As of right now, if your enterprise deploys AI involving EU users and you do not have a documented AI literacy plan, you are already non-compliant under EU law. The Commission has been explicit that this is not a future obligation. It is in force. It started fifteen months ago.

AI literacy is no longer a slide in next year’s training plan. It is a regulatory obligation, and one with no clear ceiling on how broadly “literacy” can be interpreted by an enforcement body when something goes wrong. The first enforcement actions, when they come, will likely involve companies whose AI failed in some way and whose staff could not credibly demonstrate that they had been trained to operate the system responsibly. The cheapest insurance against that is the same definitions slide. Documented. Refreshed. Embedded into onboarding.

The compliance reading is one more reason to do the thing the operational reading already demands.

What to do.

The six percent built four habits in this order. Not in parallel. Not opportunistically. In order, because each link is the foundation for the next.

First, definitions.

Every conversation about AI starts with a working glossary. Print it. Pin it. Refresh it. Make it role-specific. The product team’s definition of agent is the same as the security team’s definition of agent, or you fix it before the next meeting. The glossary is owned by someone. It is reviewed quarterly. It evolves with the field. When a new term shows up in a vendor pitch, somebody adds it to the glossary, or pushes back and says we are not using that word. The definitions slide is not bureaucracy. It is the bar everyone has to clear before the next conversation can be productive.

Second, training for task awareness, not tool features.

The ninety-four percent train on tools. The six percent train on tasks. When do you reach for AI. When do you not. How do you know if the answer is right. What do you do when it is wrong. Effective AI training is task-focused, shows rather than tells, and gives people templates they can modify and reuse. Tool-feature training is what makes people request more training a year later and still not know what to do. The training that lands is the training that lets a worker close a real ticket the day after the session.

Task awareness also means designing for the two propensities. The lean-in workers need permission to experiment. The other workers need a step-by-step playbook. Most enterprise training picks one and serves the other badly. The six percent serve both deliberately.

Third, reduce friction in the current workflow before you redesign.

Most enterprise users will not engage with AI in a separate app, a separate portal, or a separate set of credentials. Bring AI to where their work already happens. The first lift comes from showing up in Slack, Teams, the inbox, the ticketing system, the CRM. Once people are using AI inside the work they already do, you have permission and data to redesign the work itself. The WalkMe global study found workers lose fifty-one workdays a year to technology friction, up from thirty-six the previous year. Fifty-four percent of workers bypassed AI tools in the past thirty days to do the task manually instead. The friction is real. The bypass is the symptom. Fix it where the work happens.

Fourth, redesign.

This is the work the next issue will be about. You cannot redesign what you have not defined, you have not trained for, and you have not gotten adoption on. Once you have, the redesign is the lever that produces the gains the six percent capture. The teams that fundamentally redesign workflows are the ones McKinsey identifies as high performers. The teams that try to redesign without the first three steps tend to look like Klarna’s first quarter and Klarna’s third quarter.

Your next AI meeting. The thing you do.

Pick one cross-functional AI conversation on your calendar this week. Before anyone presents anything, write one word on the whiteboard. Agent. Ask every senior leader in the room to write down what they think it means. Compare the answers.

What you find is the most common-sounding word in your AI strategy vocabulary meaning three different things to the three most senior people in the room. None of them knew it. The decisions the program has made for the last three months were sitting on top of that misalignment the whole time.

Fix that meeting. Then fix the next one. The four habits flow downstream once this one is done.

Closing.

The six percent are not magic. They are not particularly more advanced technically. The thing that separated them from everybody else was a discipline most of the industry still treats as optional. They agreed on what the words meant. Then they trained people on tasks instead of features. Then they reduced friction where the work already happened. Then they redesigned.

If you are running an AI program inside an enterprise this year, the cheapest, fastest, most-overlooked thing you can do is the first link.

The next issue is the workflow rebuild. The team. The process. The before and the after. If you have a story you would tell, reply.

Ben Pickett is Co-Founder and Chief Operating Officer of Swa and former Global Director of Site Reliability Engineering at Nike. Swa delivers enterprise AI inside Slack, Teams, WhatsApp, Zoom, SMS, and more channels coming, integrated into the systems behind them through user-scoped auth. One interface across many models. Zero data retention by default. Flat-rate pricing. Deployment in under three minutes. Swa Solo, our first version for individuals and small teams to try Swa before bringing it to the enterprise, ships soon. swa-ai.com

Header artwork generated with Swa.

References

1. Article 4: AI Literacy. EU Artificial Intelligence Act. Entered into application 2 February 2025. Requires providers and deployers of AI systems to ensure a sufficient level of AI literacy among their staff and other persons dealing with the operation and use of AI systems.
2. The state of AI in 2025: Agents, innovation, and transformation. McKinsey & Company, November 2025. Approximately 6% of organizations are AI high performers (5%+ EBIT impact from AI). AI high performers are 3.6x more likely to target transformational change versus efficiency. Only 21% of organizations using generative AI have redesigned even some workflows. Fundamental workflow redesign ranks highest in correlation with EBIT impact from AI.
3. How AI is reshaping workflows and redefining jobs. MIT Sloan Ideas Made to Matter, April 22 2026. Article distinguishes between organizations that treat AI as a plug-in tool (incremental improvements) and those that rethink how work is structured (unlocks full potential). Quoting researcher Shahidi: “It’s not about how I’m going to introduce AI in my existing workflow, it’s about how I can redesign my workflow in such a way that is more AI-friendly.”
4. Scaling AI Requires New Processes, Not Just New Tools. Boston Consulting Group, 2026. BCG’s 10–20–70 rule for AI transformation: 10% of effort to algorithms, 20% to technology and data, 70% to people and processes. Around 70% of AI implementation challenges stem from people and process issues.
5. Enterprises Lose 51 Workdays Per Employee to Technology Friction Annually Despite Record AI Investment. WalkMe Global Study, 2026. Survey of 3,750 respondents (1,700 senior leaders + 2,050 office and hybrid workers) at enterprises with 1,000+ employees across 14 countries. Workers lose 51 workdays per year to technology friction, up 42% from 36 days in 2025. 54% of workers bypassed AI tools and completed tasks manually at least once in the past 30 days.
6. How Klarna’s AI assistant redefined customer support at scale for 85 million active users. LangChain Customer Case Study, 2024–2025, with subsequent reporting on the reversal in 2025. Klarna replaced approximately 700 customer service roles with an AI agent, cutting workforce from roughly 5,500 to 3,400. Initial results: 2.3M conversations handled in month one, resolution time fell from 11 minutes to 2 minutes, 82% faster response times, $10M in annual savings. Six months later customer satisfaction collapsed and Klarna publicly reversed course. New model is hybrid: AI handles routine volume; humans handle complexity, emotion, and multi-step resolution.
7. Domain-Driven Design: Tackling Complexity in the Heart of Software. Eric Evans, Addison-Wesley, 2003. Introduced the concept of Ubiquitous Language: the practice of building a single, rigorous, shared vocabulary across business stakeholders, designers, and engineers as a precondition for software that does the right thing. See also Martin Fowler’s summary.
8. On the Mutual Intelligibility of Spanish and Portuguese. John B. Jensen, Hispania Vol. 72, №4 (1989). Native Spanish-speaking and Portuguese-speaking college students comprehended each other’s languages at 50% to 60% in controlled testing, despite high overall lexical similarity. Comprehension was asymmetric, with Portuguese speakers understanding Spanish more readily than the reverse.
9. Stop Paving the Cowpath: Why Agentic-First Is the Only Way to Build for the Enterprise. SDTimes, 2025. Names the failure mode of using AI to preserve existing legacy workflows, framing scripted workflows with intelligence layered on top as the pattern through which most enterprise AI initiatives quietly die.
10. The State of Enterprise AI 2025. OpenAI, 2025. Frontier workers (top five percent) send six times more AI messages than the median employee. Multi-model usage skews to a small power-user cohort, who account for a disproportionate share of total platform value.
11. 2025 Work Reimagined Survey. Ernst & Young, November 2025. 88% of employees use AI at work; only 5% use it in transformative ways. The shape that appears at the firm level (six percent as AI high performers) appears again inside the firm at the worker level.
12. The Productivity Disconnect. The Enterprise Field Notes Issue #4, May 14 2026.
13. The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed. RAND Corporation, 2024 (Ryseff, De Bruhl, Newberry). Five-cause taxonomy of AI failure. “Business stakeholders misunderstanding or miscommunicating the problem” is ranked the number one root cause and explicitly stated to cause more AI projects to fail than any other factor.
14. A Scoping Review of Electronic Health Records Interoperability Levels, Expectations, Approaches, and Problems. El-Yafouri & Klieb, SAGE / Health Informatics Journal, 2025. The persistent number one barrier to value capture from EHR investments remains terminology standardization. The same definitional gaps identified in the early 2000s are still defeating EHR ROI two decades later.
15. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027. Gartner, June 25 2025. Documents the drift of the term ‘AI agent’ across vendors, products, and customer expectations as evidence of vocabulary breakdown at industry scale; introduces “agentwashing” as the rebranding of existing products (AI assistants, RPA, chatbots) as agents without substantial agentic capability.

About the author

I'm Ben. I write Enterprise Field Notes, and by day I'm COO at Swa, after years running reliability, data protection, and database operations at Nike. The lesson that keeps proving itself: anything you cannot run without, and cannot walk away from, is a risk you have not priced yet. What is yours?

Header artwork generated with Swa.

Read more of Ben's Enterprise Field Notes at benpickett.com.

© 2026 Ben Pickett · Enterprise Field Notes