The surveys of the eight domains reveal patterns that transcend the specific applications. Across consulting, legal, insurance, finance, real estate, marketing, and operations, AI shows up in specific ways that cluster into a small number of collaboration patterns. Understanding these patterns as a general framework is useful both for interpreting the surveys and for thinking about where AI might apply in specific work the reader does that the surveys did not explicitly cover.
A useful three-level framework distinguishes the collaboration patterns by how much autonomy the AI system has and how much direct human involvement each action requires. The three levels are assisted work, augmented work, and autonomous systems. They represent a progression rather than a strict sequence, and professional practice involves AI at all three levels simultaneously across different tasks.
Level One: Assisted Work
At the first level, AI functions as a helper inside existing tasks without fundamentally changing how the work is organised or who is responsible for it. A consultant using a large language model to draft a memo is doing assisted work. A lawyer using a legal research tool to find relevant cases is doing assisted work. An analyst using a coding assistant to write analytical code is doing assisted work. A real estate agent using a tool to generate a first draft of a listing description is doing assisted work. An operations manager using AI to draft a vendor communication is doing assisted work. The human performs the task and uses AI to accelerate specific steps within it.
Assisted work is characterised by the human remaining in full control of what gets produced, the AI making suggestions or providing drafts that the human evaluates and edits, and the human taking responsibility for the final output. The AI reduces friction and accelerates specific steps without taking on judgment about what the output should be or whether it meets the relevant standard. The professional is still doing the professional work. The AI is helping them do it faster.
The pattern across the surveys is that assisted work dominates the current professional use of AI tools. Most day-to-day AI uses as a consultant, lawyer, agent, or marketer does fall into this category. The barrier to adoption is low (the tool produces a draft, the professional accepts or rejects it, the work proceeds), the risk is bounded (a bad AI suggestion gets caught when the professional reviews it), and the value is immediate (the work gets done faster). The conditions under which assisted work produces substantive value include clear human ownership of the task, AI outputs that are evaluated rather than adopted without review, and feedback loops that allow the human to correct or redirect AI output when it goes wrong. These conditions are usually straightforward to maintain. The result is that most professionals can absorb assisted AI use into existing work patterns with relatively little friction, and most professionals discover that some tasks are well-suited to AI assistance and others are less well-suited, developing a working repertoire over time.
Level Two: Augmented Work
At the second level, the collaboration becomes more substantial. The AI takes on meaningful parts of the work rather than just accelerating specific steps, and the human concentrates on higher-level judgment while the AI handles substantial analytical or productive work. A radiologist who reviews AI-flagged regions on a scan is doing augmented work. A lawyer whose firm uses AI to produce initial contract reviews for human validation is doing augmented work. An insurance underwriter who reviews AI-generated risk assessments is doing augmented work. A CRE analyst who reviews AI-produced lease abstracts on a 47-tenant portfolio is doing augmented work. A credit officer who reviews AI-generated credit recommendations before approving loans is doing augmented work. The AI produces a substantive work product that the human then validates, refines, and takes responsibility for.
Augmented work changes the shape of the professional's job. The time spent on mechanical production decreases. The time spent on judgment, exception-handling, and quality control increases. The professional's value concentrates in the areas where their training and experience produce insight that the AI cannot reliably replicate. For many professionals, this shift is substantively positive (the work becomes more interesting and higher-leverage) and organisationally challenging (adapting to a work pattern where the AI produces the first version of most outputs requires developing different habits and skills).
The conditions under which augmented work produces substantive value are more demanding than those for assisted work. The AI system needs to be reliable enough that most of its outputs are acceptable with modest human revision. The human needs to maintain sufficient underlying skill to catch AI errors rather than becoming dependent on AI output that they cannot validate. The work processes need structured review and approval steps that keep the human in substantive ownership of the output rather than rubber-stamping whatever the AI produces. These conditions require deliberate design, and the organisations that have deployed augmented work successfully have typically invested substantially in the processes and skills that make it work. The organisations that have deployed augmented work badly typically fail on one of these conditions. The AI system is not reliable enough, and professionals spend as much time fixing errors as they would have spent producing the work themselves. The humans become dependent on AI output and lose the underlying skill to catch errors, producing a workforce that cannot function when the AI is unavailable and that cannot catch the errors that do appear. The review processes become perfunctory, and the AI output passes through unexamined, producing failures that reach clients and stakeholders. The design of the work processes matters as much as the capability of the technology, and the design is where most of the difficulty sits.
Level Three: Autonomous Systems
At the third level, AI systems operate with substantial independence within carefully defined limits. Humans set the goals, rules, and oversight structures, and the system executes its work without requiring explicit human approval for each action. Algorithmic trading systems operating under defined risk parameters are autonomous. Automated claims processing systems that pay low-dollar routine claims without human review are autonomous. Anti-money-laundering systems that flag suspicious transactions for review are autonomous (the flagging happens automatically; the review is human). Fraud detection systems that decline transactions in real time are autonomous.
Autonomous systems require the most deliberate design and the most substantive governance. The scope of their autonomy needs to be defined precisely. The rules that bound their behaviour need to be explicit. The monitoring that catches anomalies needs to be continuous. The escalation paths that bring humans in when the system reaches its boundaries need to be clear. The accountability structure needs to specify who is responsible for the system's actions, because accountability does not transfer to the system itself even when the system is acting without direct human approval.
Autonomous systems work best in domains where the rules of acceptable behaviour are clear, the consequences of individual actions are bounded, and the patterns the system learns from historical data remain stable over time. They work less well in domains with substantive edge cases, rapid regime change, or high-consequence individual decisions where human judgment is required for defensibility. Most professional domains use autonomous systems for specific narrowly defined tasks within broader work that remains assisted or augmented, and this pattern reflects both the technology's current capabilities and the governance requirements of professional practice.
Patterns That Apply Across Levels
Three observations hold across all three levels of collaboration and across all eight domains the surveys covered.
The first is that the human's role shifts toward judgment, exception-handling, and responsibility as AI takes on more of the mechanical work. This is consistent across domains and across levels. The professional work moves from the production of output to the oversight of output production rather than disappearing, and the skills that matter most become those that the AI cannot reliably replicate, including understanding specific client context, making judgment calls on edge cases, accepting accountability for outcomes, and directing the AI's work toward the right questions.
The second is that the technology works best when it is treated as a tool rather than as a substitute for the professional. Tools that are substantively useful get adopted because they help the professional do better work faster. Tools that are presented as replacements for professional judgment tend to be rejected either by the professionals who would use them or by the clients, regulators, and stakeholders who hold the professionals accountable. The deployments that succeed position AI as augmenting professional capability, and the framing matters both for adoption and for the governance around how the tools actually get used.
The third is that the underlying capability of the technology has stabilised enough that specific products matter less than underlying capabilities. A professional who understands what large language models can do, what machine learning contributes, where deep learning fits, and how the four-layer stack composes into real systems can evaluate any specific product against the general capability and form accurate expectations. Specific products will continue to evolve and to come and go. The capability pattern has been stable for several years and is likely to remain so. Professionals who build their practice around understanding the capability rather than around familiarity with specific products are building a foundation that will age well.