Find your weakest productivity vector in 5 minutes with the free assessment, and if you want to work on it with a group, Productive, Not Busy runs live on Maven.
Your process documents describe one company. The chasing, the side spreadsheet, and the approval that sits for four days describe a different one. AI in the workplace reads the second company, because that is the one leaving a record, and then it copies what it finds at speed.
David Dean spent 20 years in enterprise technology, 10 of them leading AI, automation, and data programmes across healthcare, insurance, hospitality, consulting, and regulated industries. He built the governance frameworks and ran the modernisation projects, and on one enterprise service he cut processing time by more than ninety percent. In January 2026, he published An Inbox Between Us: AI and the Reality of Modern Work. He calls himself a business AI realist, and he is blunt about what a rollout actually exposes.
Three things change for you by the end of this piece. You will be able to name the undocumented behaviours your own role depends on, which turns a tool decision from a guess into a choice. You will have one question that flags a doomed AI project before the budget is spent. And you will have a five-minute habit that makes invisible work readable by the people who buy the software.
If a new tool is about to land on your team and nobody’s asked what your day actually looks like, keep reading.
AI in the workplace reads behaviour, not process documents
David’s argument starts with a distinction that sounds administrative and turns out to be the whole story. Every organisation runs on two contracts. One of them is written down. The other one is what actually happens.
“You have the official contract of the business, which is your job description, your SOPs, your policies,” he says. Then there is the other one, “the unwritten contract, which is how the business actually runs behind the scenes.” Nobody signs the second contract. Everybody works inside it.
He points to a line sitting at the bottom of almost every job description ever written. Other duties apply. Four words, no detail, and David reads them as the honest part of the document. “I really actually hired you to deal with the ambiguous aspect of all of those duties,” is how he puts the subtext. The listed responsibilities are the part a system can see. The ambiguity is the part a person absorbs.
What does absorbing it look like in practice? It looks like a chain of small human moves that never appear on an org chart. “John has to go talk to Susie, Susie has to go talk to Steve.” It looks like a spreadsheet that exists outside every official system, maintained by one person, for reasons nobody above them has asked about. David has asked. When he does, the answer rarely sounds strategic. “Why did you create the spreadsheet?” he asks, and the answer “fundamentally comes down to one category to survive.”
That is the raw material AI gets pointed at. Not the tidy process map in the transformation deck. The behaviour.
Why a workaround outlives the reason it was created
If undocumented work is so widespread, why has nobody cleaned it up? David’s answer is that the cleanup never happens because the original reason disappears while the behaviour stays.
Deviations from process, he says, come down to two drivers. One is control. The other is survival. Someone wants a handle on an outcome they are accountable for, or someone wants to get through the week. Either way they build a small private system. Then time passes. “That deviation gets lost in time,” he says, and what replaces the reason is a sentence every operator has heard: “This is how it’s always been done.”
Ask the question and you find the answer is gone. The person who built the workaround has moved on, or the system it was working around has been replaced, or the policy that made it necessary was rewritten three reorganisations ago. The behaviour carried on anyway, because it worked, and because nobody was rewarded for examining it.
There is a second pattern David has noticed, and it explains a lot of the gap between the boardroom and the floor. The questions he gets about AI depend entirely on where the person asking sits. Executive-level questions are about strategy, competitive position and risk appetite. Ground-level questions are about whether the job survives the quarter. Same technology, two conversations, and the people designing the rollout usually only hear one of them.
If you want to improve team productivity in a real sense, that gap is the work. The process document is already written. The behaviour is not.
The 79 percent nobody is designing the rollout for
Here is the number that gives this episode its title. “Seventy-nine percent of the workforce is what I call AI survivalists,” David says.
The word matters. A survivalist is not a resister and not an evangelist. “They are not against AI, you know, they’re not for AI. They’re just people doing their jobs.” They turn up, the tool arrives, they work out the minimum viable relationship with it, and they carry on. No manifesto, no sabotage, no enthusiasm worth putting in a case study.
Nearly every AI communication plan is written for the other twenty one percent. The champions get a pilot programme and a Slack channel. The sceptics get a change management workshop. The survivalists get an email. And they are the group whose behaviour determines whether anything sticks, because they are the ones holding the unwritten contract together.
David is clear about what fails with that group. “Cramming something down people’s throats is not going to allow for real change to happen.” He frames the alternative as a relationship rather than a deployment. “In any relationship, it takes time and effort,” he says, and the trust has to move in both directions. “People have to trust AI to a certain degree and AI has to trust people.” He does think the long run is symbiotic. He also thinks we are early, because “it’s still too chaotic.”
What AI copies when you scale it
This is the part of the conversation that should worry anyone with a 2027 automation roadmap.
“AI is is here to mimic human behavior and industrialize human behavior,” David says. Read that twice, because the second verb is where the risk lives. Mimicry is a feature. Industrialisation is a multiplier. And a multiplier applied to a broken process does not fix the process.
“You’re going to industrialize, you know, dysfunction exponentially utilizing AI with this,” he says. That is his explanation for a large share of failed AI projects. The pilot works, because a pilot runs in a controlled slice of reality with a motivated team. Then it scales into the real behaviour of the organisation, and what scales with it is the ambiguity, the workarounds and the undocumented chasing that the pilot never saw.
Santi puts the pressure that drives this into plain terms during the episode: short termism, and an environment where output sits above quality. Five articles by Monday. Reports reconciled by Tuesday. In that climate a tool that produces volume looks like an obvious win, and nobody is measured on whether the volume was worth producing.
David’s counterweight is a line about where authority has to stay. Risk decisions belong to executive leadership, with a human in the loop and real training behind the change. Not because the technology is incapable, but because of a limit he states flatly: “the AI, no matter even when we get to AGI, is never going to understand lived experience and lived consequence. That is a human trait.”
He has applied that test to himself, out loud, about writing. “By letting AI write, I’m losing something. I’m losing my writing ability. I can still read, I can still assess, but I’m not typing anything anymore. I’m not constructing, I’m just assessing. Is that okay with me?” The question is the method. You can answer it either way for your own work. The failure is never asking it.
The four behaviours that keep work moving
Midway through the episode Santi names the behaviours that sit underneath the undocumented work: interpreting requests, clarifying intent, prioritising, and resolving ambiguity.
Look at that list next to a job description and notice how little overlap there is. None of those four appear as responsibilities. All four happen every day. They are what somebody does between receiving a message and producing an output, and they are invisible precisely because they are fast and unremarkable to the person doing them.
Which is why David reframes what a tool is for. “AI is here is to extend and augment, you know, enhance or replace behaviors,” he says. Behaviours, rather than roles. A tool that takes over interpreting requests inside a team with clear intent will help. The same tool dropped into a team where intent arrives vague and gets repaired informally by three people will produce confident output based on a misread, at volume.
His view of what AI understands is worth keeping as a sanity check on any vendor demo. “AI learns the map, not the terrain.” The process document is the map. The four behaviours are the terrain.
Start a work diary before you choose a tool
The practical instruction in this episode takes about five minutes and costs nothing. It was also David’s answer when Santi asked what people should put on the calendar before Friday: time for self discovery, and a work diary.
The method is deliberately low friction. “Open up your mobile phone, open up a notes app, hit the dictate button button, and just unload. Tell me your work story,” he says. Talk through what actually happened in your day. The decision you made with incomplete information. The message you rewrote because the first version would have been misread. The call you made to unblock something that was nobody’s assigned task.
“You need to start creating digital diaries on how you behave out there,” he says. And then it scales. Individual, team, department, division, company. Each level produces a readable record of behaviour that previously existed only in people’s heads.
Two reasons to do this even if your organisation has no AI plans at all. It is the fastest honest audit of your own role you can run, and the answers tend to surprise people who have been doing the job for years. It is also the only version of your work that a tool decision can be made against. Without it, a buyer is choosing software for the company in the process document.
If you want a structured starting point for the same question, the free productivity assessment takes 16 questions and five minutes and tells you which vector is costing you ground. The free tools and templates library covers the follow-up work.
One question that predicts a failed AI project
In the quickfire round Santi asks for the first sign an AI project is going to fail. David does not reach for technology, budget or data quality. He reaches for a conversation that did or did not happen.
“Did you have any conversation with the frontline people that are gonna be directly affected on the outcome of that AI project?” he asks. Then the consequence, in order: “You never talk to who is actually going to use the solution. And if you don’t understand how they behave, they’re not gonna adopt it and there’s gonna be no success. And you just threw all that money away.”
Adoption is downstream of behaviour. Behaviour is downstream of someone having asked. The spend is downstream of all of it, which is why the failure is only visible at the end, by which point it reads as a technology problem.
The rest of the quickfire is worth the listen on its own. The meeting David would delete is the everyday touch base with no real purpose, and he prices it: “You realize this meeting’s costing the the company twenty four thousand dollars in payroll to have this meeting.” The slide he finds least honest is the one claiming “that we’re a family.” And the AI use he considers close to worthless in a business context is the low-consequence fun. “Socially that’s great, but when it comes to a business model and take action, it’s not gonna provide a lot of opportunity value.”
Where the book title came from
The story behind An Inbox Between Us involves a frontline worker in tears. “I feel like I’m training my replacement,” they told David. “I’m not going to have a job a year from now.”
His response was a question that reframes the whole AI conversation. “Who’s going to write the message for them?” Somebody has to speak for the behaviour that nobody wrote down, and the people best placed to describe it are the ones currently assuming it will be used against them.
That is also where the inbox comes in. David calls it the behavioural record, and he means it literally. It sits in “your emails, your chat messages, and meeting transcripts, comments, etc.” Every clarification, every escalation, every four-day approval and every follow-up nudge is logged there. The organisation already has the data about how it really operates. It has simply never read it as evidence.
And his sequencing on the future of work is the same either way. “The only way to really get to what if,” he says, “is you need to understand the what is.”
Frequently asked questions about AI in the workplace
What is an AI survivalist?
David Dean’s term for the 79 percent of workers who are neither advocates nor opponents of AI. In his words, “they’re just people doing their jobs.” They adopt whatever the tool requires of them and nothing more, which makes them the group that decides whether a rollout holds.
Why do AI projects fail even when the technology works?
Because the technology is applied to the documented version of the work rather than the real one. AI mimics and industrialises behaviour, so scaling it across an undocumented process multiplies the workarounds instead of removing them. The earliest warning sign is that nobody spoke to the people who will use the tool.
Is AI going to replace jobs?
David’s framing is that jobs are labels and the real unit is behaviour. A tool can extend, augment, enhance or replace specific behaviours, and which ones it can take over depends on whether the surrounding intent is clear. That is a question about how a given team operates rather than a single answer for every role.
How do I document the undocumented work on my team?
Start with a dictated work diary. Open the notes app, press dictate, and describe what actually happened in your day, including the decisions you made with incomplete information. Do it for a week, then compare it with your job description. Scale the same habit from individual to team once people see the point.
What should leaders own in an AI rollout?
Risk. David puts risk decisions with executive leadership, backed by a human in the loop and real training, on the grounds that AI will never understand lived experience and lived consequence. He also argues that companies are “not properly capitalizing on their existing assets as people,” which is a leadership problem rather than a tooling one.
What to do with this before Friday
Santi closes the episode on two points. The first is that we are discovering each other, people and tools both, and that discovery takes time. The second is sharper: companies need to double down on human capital, and probably not in the way they did it over the last thirty or forty years.
So pick the smallest version of that. Dictate one work diary entry tonight about a decision you made today with incomplete information. Then take the free productivity test to see which vector is actually holding your output back, and if you want to fix it alongside other operators rather than alone, Productive, Not Busy runs live on Maven. Weekly thinking on the same questions goes out in the productivity newsletter, and there is more on the frameworks at santiagotacoronte.com.
Full episode with David Dean on the podcast. His book, An Inbox Between Us: AI and the Reality of Modern Work, is out now, with the audiobook narrated by Zachary Locklear, and he writes The AI Mirror newsletter on LinkedIn.