A recent HBR article on AI adoption at work identifies that AI has intensified work for employees. This intensification, the article notes, takes the form of expanding tasks, more multitasking, and working beyond normal hours, all intrinsically motivated, though I'd argue there is an element of fear of being automated away among all corporate employees these days. All of this should make employers happy, but the article warns that such intensification is not sustainable and will lead to eventual burnout. The authors conclude that the solution is to slow things down by adding human-designed pauses in the workflow, which to me feels like advising that a Formula One car be driven within speed limits. This advice is not something organizations are likely to follow through, especially as they try to be ahead in the AI race.
The article points out the accurate symptoms of early adoption of AI in an enterprise but misses what is actually happening: individuals are discovering new protocols and ways of working together, while organizational standardization lags behind.
To understand this, we should first have a map for what AI adoption looks like.
The AI Capability Maturity Model
Over the last few months, the Protocols for Business Special Interest Group has been working on a Capability Maturity Model to provide legibility for AI adoption at enterprise companies. The model maps what adoption looks like at each level, from shadow adoption to utility that drives other technologies. Here are the five levels of adoption at a glance.
AI is in use but the organization doesn't know how, by whom, or with what data. Employees use personal accounts for work tasks. No policy governs what data leaves the organization. Exposure is invisible until a data leak, regulatory inquiry, or quality failure surfaces it.
Historical parallel: spreadsheets in accounting in the mid-1980s, adopted through individual initiative before IT had any policy.
The organization has granted broad AI access and signaled strategic intent. Enterprise licenses are deployed, adoption rates climb. What's missing is the governing protocol: which tasks, what review process, what escalation path when AI fails. High adoption and early productivity gains coexist with governance failures that surface when they become external-facing.
Historical parallel: corporate email in the early 1990s, mandated broadly before organizations designed how to use it.
At least one core workflow has been deliberately built around AI with a named owner, quality metrics, and a defined escalation path. Removing AI would require rebuilding the workflow. The governance question shifts from controlling use to designing protocols that handle failure, versioning, and external dependencies. Domain expertise is the limiting constraint.
Historical parallel: Git and CI/CD adoption from 2008–2015, where competitive position depended on how well the deployment workflow was designed.
AI capability has become a baseline expectation across the sector. Individual organizational advantage has been competed away. The governance challenge is now collective — how does the industry coordinate AI use, handle shared risks, and establish interoperability standards. Individual maturity is necessary but not sufficient.
Historical parallel: EDI in retail and manufacturing in the late 1980s–90s, where Walmart mandated Electronic Data Interchange for all suppliers and not adopting meant not participating in the market.
AI governs critical civilization-scale coordination systems: supply chains, financial infrastructure, public health surveillance. The governance challenge is no longer adoption or protocol design but legibility: understanding what the systems are doing well enough to intervene when they fail. No single actor controls the full system.
Historical parallel: internet protocols (TCP/IP, BGP) today, where the 2021 Facebook BGP misconfiguration took down multiple platforms globally before anyone could respond.
Who Can Best Orient To New Protocols?
I do not think the stages of AI adoption are useful to someone who is new in their field or enterprise. The CMM serves as a map for individuals who have developed tacit knowledge and domain expertise.
At the Protocols for Business Special Interest Group, individuals with specific domain expertise working with each other are often at Level 3, designing workflows for themselves to work on interesting projects and, importantly, share contexts with others. In a pre-AI workflow, collaboration meant passing discrete artifacts between people — documents, specs, pull requests, and the like. The boundaries of your contribution were defined by the artifact. But when someone spins up a prototype with AI and shares it, they're not sharing an artifact. They're sharing a context: a whole decision space, a set of assumptions baked into the prototype, an implicit workflow. The person receiving it now has to engage with that entire context, not just review a document. The artifact boundary that previously contained the scope of interaction has dissolved. The adoption of Claude Code and OpenClaw has shown that people are eager to adopt better ways to share contexts with others. (At the PfBSIG we have a GitHub project to manage multiple people working on the CMM, which you can fork or join.)
In a recent Protocolized essay, Venkatesh Rao described what he calls the F2F pattern: "factory-to-factory." The idea builds on an earlier editorial arguing that AI is shifting from destination intelligence (you go to ChatGPT) to intelligence media, where intermediate artifacts circulate between people's bespoke setups the way industrial intermediates flow between factories.
Venkat's case study is his own Claude Code infrastructure, which includes his own book manuscript factory. This factory runs on his laptop and has a couple of dozen projects within it from two decades of blog and newsletter archives. That being said, the factory itself wasn't the most interesting part. It was connecting it to another factory.
His long-time publishing collaborator Jenna Dixon independently set up her own Claude Code factory for producing finished books from manuscripts. Venkat's factory takes messy raw materials and produces rough drafts. Dixon's factory takes those drafts and produces finished artifacts ready for Amazon. The handoff between them is a shared Dropbox folder plus a manuscript transmittal server she built for metadata. When Rao submitted a docx file from his factory, Dixon returned revised requirements. Her factory needed styling issues resolved before it could begin design. He told his factory to fix them and sent it back. What circulates between them is intermediate production states, specifications, and revised requirements. Industrial intermediates, not documents to be read.
Neither of them writes code. Neither has touched a line of content text. What they bring is domain knowledge. Claude provides commodity technical execution.
Venkat argues this is the invisible ninety percent of the agentic AI revolution. The public spectacle of autonomous agents draws attention, but the real capability integration happens in high-trust bilateral connections between people who know their domains. The connections look nothing like social media's friends and followers; they're closer to B2B supply chain relationships, with shared specifications, defined handoff points, and mutual understanding of quality standards. And they're emerging entirely outside organizational boundaries, on personal laptops and personal subscriptions, invisible to any enterprise that might claim these individuals as employees.
The factory model generates a problem of its own. Working across shared contexts, intermediate artifacts, and bilateral pipelines means output accumulates faster than any individual can fully evaluate. This is orientation debt: the gap between the pace at which AI produces usable work and the pace at which a person can verify that it is any good. At organizational scale this manifests as temporal divergence, the Level 3 failure mode in the maturity model: AI-accelerated output volume outrunning human review capacity. At the individual level it is simply the pressure of contexts multiplying faster than you can track them.
But orientation debt is survivable if you have something else: higher-level perception. The ability to read the shape of what's happening without getting bogged down in the details of every individual output. This is what domain expertise gives you. Not speed — comprehension at a glance.
Compare this to an experienced soccer player who has lost a step physically. Their reflexes are slower, they can't outrun the 22-year-old winger anymore, and in any raw-speed contest they lose. But they read the game better than almost anyone on the pitch. They see the dangerous pass before it's played. They position themselves where the ball is going to be, not where it is. They process the state of play at a level of abstraction that younger, faster players haven't developed yet, and that perception advantage more than compensates for the speed deficit.
This is the model for human-AI collaboration. You are never going to match the machine's processing speed, just as the aging midfielder is never going to outrun the young striker. But if you have deep domain expertise, you can read the game. You can look at an AI's output and immediately sense whether the shape is right (whether the logic holds, whether the framing fits the problem, whether something critical is missing) without laboriously checking every line. You are pattern-matching at a higher level of abstraction.
Bricoleurs as Protocol Entrepreneurs
On August 5, 1949, sixteen smokejumpers parachuted into Mann Gulch, Montana to fight what they believed was a routine fire. Within minutes, the fire exploded and cut off their escape route. Foreman Wag Dodge did something no one had ever done: he stopped running, lit a fire in the grass ahead of him, and lay down in the ashes as the main fire swept over him. He survived by creating what is known as an escape fire. Thirteen of his crew, who rejected his shouted orders and kept running uphill, did not. The crew's failure wasn't physical. They were young, fit, fast. It was a failure of sensemaking: Dodge's commands — drop your tools, stop running, lie down in a fire I just set — were so alien to anything in their experience that they couldn't process them as rational. The organizational structure that might have compelled compliance had already dissolved. Each man was alone, falling back on the most overlearned response available: run.
Karl Weick, analyzing the disaster in his 1993 paper, identifies Dodge as a bricoleur, someone who creates order out of whatever materials happen to be at hand, through intimate, embodied familiarity with their environment. Dodge's escape fire was a creative recombination of things he already knew: the fire triangle, backfire principles, ridge behavior, fuel dynamics. What made it look like genius was that he could synthesize under pressure, because his knowledge was practical and deeply handled, not theoretical. Weick's deeper point is that improvisation under crisis is what bricoleurs do every day, at higher intensity. They remain creative under pressure because working with chaotic conditions and recombining available resources is their normal mode of operation. The crisis doesn't break their sensemaking; it intensifies it.
Domain experts building nascent AI protocols for themselves and their collaborators are the Enterprise's bricoleurs. They give enterprises the grounding necessary to create protocols and standardization at the organizational level.
This reframes the enterprise AI governance challenge. The conventional approach treats it as a control problem: how do we manage what employees do with AI? Acceptable use policies, data boundaries, approved tool lists. That's necessary but insufficient, because it addresses Level 1 risks (data leakage, unauthorized use) without building Level 3 capability (designed workflows with governed handoffs).
What enterprises actually need is protocol design: the rules governing coordination behavior between agents (human and artificial) at various scales. This means seeing and empowering the handoff points where AI-produced context is already moving between people. It means defining what a governed context exchange looks like — not just which artifacts are acceptable to share, but which decision spaces and intermediate artifacts to govern. It means building the equivalent of Rao and Dixon's manuscript transmittal server, but at an organizational scale.
Ready to take the assessment?
A 2-minute diagnostic places you on one of the five levels, returns the failure modes you're most exposed to, and gives you clear direction on what do next.
