An interactive field guide to evaluative and generative AI, augmentation and automation, changing jobs, responsible deployment, product selection, and building useful AI workflows.
A useful AI application changes a task you care about. The harder questions are whether it saves time after review, who gets the benefit, and who can challenge a mistake.
The Applications layer is where the entire AMICE stack finally meets a person.
It is the meeting summary in your inbox, the fraud alert on a purchase, the coding agent opening a pull request, the image generator building a storyboard, the system ranking job candidates, and the support bot deciding whether you reach a human. Below each of those experiences may sit several models, clouds, chips, data centers, and power systems. But this is the layer where technical choices become the work people get to do and the decisions made about them.
That makes it the most personal layer of the stack. People do not usually worry that a tensor will replace their job. They worry that a manager will use a product they cannot inspect to change the work they enjoy, intensify the work they dislike, or decide whether they are hired at all. At the same time, millions of people are discovering that AI can help them write in a second language, understand unfamiliar code, make a video they could not afford to commission, analyze a spreadsheet, start a business, or spend less of the day copying information between tools.
Six boundaries worth keeping
Answers: A bounded unit of work with an observable input and output.
Does not answer: An occupation bundles many tasks, relationships, responsibilities, and tacit judgments.
The interactive product and labor-study records retain their August 13, 2026 research cutoff. The prose and further reading were reviewed on September 17, 2026; that does not redate every product disclosure. Product features and prices change constantly. Labor evidence changes more slowly, but it answers different questions: a capability study is not an adoption measure, a controlled productivity experiment is not an employment forecast, and exposure is not displacement. The evidence labels throughout this page keep those categories separate.
I start by asking which part of the work the product actually does.
Generative AI produces a new artifact: text, code, an image, video, audio, a design, a simulation, or a proposed sequence of actions. The NIST Generative AI Profile defines the category as models that emulate characteristics of input data to generate derived synthetic content.
Evaluative AI is a useful application-level label, not a universal scientific or legal category. It scores, ranks, classifies, predicts, recommends, matches, flags, or decides over an input. A spam filter evaluates a message. A demand forecast evaluates historical signals. A recommender ranks possible items. A hiring system may score an applicant. These systems can use machine learning without generating a paragraph or picture.
Many modern applications are hybrids. A research assistant can classify the question, retrieve and rank documents, generate a synthesis, evaluate its citations, and route the result through a safety check. A video tool can generate clips and use evaluators to reject unsafe or low-quality frames. A customer-service product can score intent, retrieve an account record, draft a response, and decide whether to escalate.
Application mode lab
Select a mode, then inspect real workflow shapes. Products can combine modes, so these are behaviors—not mutually exclusive company categories.
Evaluative
Design boundary: A wrong label can delay service even though no new prose was generated.
The failure modes differ. A generator can invent a plausible statement, imitate a creator, expose private context, or flood a channel with low-cost content. An evaluator can invisibly sort people, encode historical bias, make a brittle prediction look objective, or deny an opportunity at scale. A hybrid can do both.
"Does it generate?" is therefore only the first question. Follow it with:
Those questions turn an AI feature into an inspectable application.
The word application can make a system sound like one neat box. In practice, even a modest feature may contain a surprising amount of choreography.
Imagine an assistant that prepares a client proposal. It authenticates the employee, retrieves customer history from a CRM, checks permissions, converts documents into model-ready context, selects a model, generates a draft, calls a pricing tool, evaluates claims against source material, checks policy, shows the result for approval, records changes, and exports a document. A polished paragraph is only one node.
Workflow anatomy
Walk through a defensible application loop. Evaluation, generation, and action can appear together, but responsibility cannot be delegated to a diagram.
Stage 1 of 7
A person, event, or schedule starts a bounded workflow.
Ask: Who asked, and what outcome do they expect?
The surrounding workflow often determines whether the model is useful. The strongest model cannot retrieve a contract that was never digitized, recognize an exception stored only in a veteran employee's memory, or resolve an ambiguous approval policy. A cheaper model with clean context and a well-designed checkpoint may outperform a frontier model dropped into a broken process.
This is why AI integration is organizational work as much as software work. Teams need to decide which system remains authoritative, how permissions travel, what gets logged, when a request escalates, how failures recover, and who owns the result. The "last mile" is not a minor implementation detail. It is where an impressive demonstration becomes dependable work.
The optimistic opportunity is larger than making the current task faster. A good workflow can make expertise available at the moment it is needed, translate between specialties, give a novice a safe starting point, or make a service economical for people previously priced out. The goal should not be to sprinkle AI over every screen. It should be to remove a real constraint.
These are separate axes.
If an AI drafts an email and a person revises it, generation is augmentative. If it writes and sends thousands without review, generation is automated. If an anomaly detector highlights transactions for an accountant, evaluation is augmentative. If it freezes an account on its own, evaluation is automated.
Augmentation or automation?
Move the two axes. Automating 80% of document production is not the same as allowing a system to make an 80%-autonomous decision.
AI contributes a bounded suggestion or draft; a person performs most work and decides.
This is a design position, not a readiness score or prediction of job loss.
There is also a ladder of control between "human" and "machine":
Moving down that ladder lets one person supervise more work. It also changes the skills, time, and information a person needs to exercise real control. A reviewer asked to approve 600 decisions per hour is not meaningfully in charge. A manager who sees only the system's recommendation may anchor on it even when an override button exists. "Human in the loop" describes topology, not accountability.
Anthropic's Economic Index provides one observed-use taxonomy, although it covers Claude rather than the whole AI economy. Its January 2026 report classified directive completion and feedback-loop editing as automation, and learning, validation, and task iteration as augmentation. In its November 2025 Claude.ai sample, augmentation represented 52% of conversations and automation 45%. An earlier look at Anthropic's first-party API traffic found 77% of business uses exhibited automation patterns, which is unsurprising for software calling a model programmatically. Neither number is a universal split.
The same product may sit in different places for different people. A scheduling system augments the supervisor who receives a recommendation while automatically determining a worker's hours. An AI performance dashboard gives an executive new visibility while making an employee feel continuously watched. Always identify the operator, the subject, and the person who owns the consequences.
Headlines ask whether AI will replace accountants, designers, teachers, programmers, or lawyers. Organizations rarely make a job disappear in one clean technical step. They change a task, a handoff, a staffing ratio, an entry requirement, or the amount of work one person can supervise.
Individual tasks
Move individual tasks among human work, assistance, bounded delegation, and automation. The exercise changes a bundle; it does not forecast headcount.
Boundary: Paraphrased from O*NET task records. The list is an illustration, not a complete job description, time allocation, capability claim, or prediction that any task should be automated.
Illustrative task
A person performs and decides.
Task constraints: Requires current policy evidence and recognition of when the standard answer does not fit.
Illustrative task
A person performs and decides.
Task constraints: Resolution can affect money, access, trust, and a customer’s ability to appeal.
Illustrative task
A person performs and decides.
Task constraints: A durable record needs accuracy, appropriate retention, and a correction path.
Illustrative task
A person performs and decides.
Task constraints: Exceptions expose gaps in data, policy, and organizational ownership.
Illustrative task
A person performs and decides.
Task constraints: A real handoff needs a staffed destination, context, and service-level expectation.
A job contains activities with very different properties. Some are standardized and digitally legible. Some require physical action, local context, trust, accountability, taste, or tacit knowledge accumulated over years. Some are technically automatable but socially undesirable to automate. Some look routine but are the practice through which a new worker becomes an expert.
The ILO and NASK's refined global exposure index, released May 20, 2025, mapped nearly 30,000 occupational tasks. It estimated that occupations representing 25% of global employment have more than minimal potential GenAI exposure, rising to 34% in high-income countries. Its conclusion was transformation, not wholesale replacement: most occupations retain tasks requiring people, and implementation depends on institutions and policy.
That is not a prediction that one in four jobs will vanish. Exposure means a system may be able to assist with or perform some constituent tasks. It does not tell us whether a workplace adopts the system, whether it works in context, how demand changes when production becomes cheaper, or what organizations do with the time saved.
Task rebundling can move in several directions:
The future of a job is the result of those choices, not an exposure score alone.
I find these results encouraging, but they describe very different kinds of work.
In a preregistered experiment with 453 college-educated professionals performing bounded writing tasks, Noy and Zhang found ChatGPT reduced completion time by 40% and raised independently graded quality by 18%. In a staggered-rollout study of 5,172 support agents at one Fortune 500 business-software company, AI assistance increased resolved issues per hour by 15% overall and 30% for less-skilled and less-experienced workers. Of those agents, 89% worked outside the United States, primarily in the Philippines. Newer agents learned faster, while the most skilled agents saw little productivity gain and a small quality decline. The single-firm study did not estimate aggregate wage or employment effects.
Three randomized field experiments at Microsoft, Accenture, and a Fortune 100 company, covering 4,867 developers, found a pooled 26.08% increase in completed tasks among developers given an AI code-completion tool. The individual experiments were noisy, adoption differed, and "tasks completed" does not capture every maintenance or quality consequence.
Then consider the counterexample. METR asked 16 experienced open-source developers to complete 246 real issues in mature repositories they knew well using early-2025 tools. With AI, they took 19% longer. The developers had expected to be faster and still felt faster afterward. METR later concluded newer tools probably helped more, but selection effects made its follow-up estimates unreliable.
Measured results
Filter studies by the mechanism they observed. Each result retains its population and study boundary; measured task effects are not universal employment forecasts.
Global workforce
Population: Nearly 30,000 tasks linked to harmonized global employment data.
mixed
25 % globally
Capability exposure, not an observed job-loss rate.
mixed
34 %
The highest-exposure category is much smaller than all exposed work.
Do not generalize past this: Potential exposure estimates do not measure adoption, actual displacement, a probability of job loss, or the net demand response. ILO concludes transformation is more likely than full replacement because most occupations retain human tasks.
These findings do not cancel one another. They describe different workers, tools, tasks, contexts, and moments in a fast-changing technology.
The best name for this is the jagged frontier. In an experiment with 758 BCG consultants, AI helped on tasks inside the tested capability frontier: participants completed 12.2% more subtasks, worked about 25% faster, and produced work rated roughly 32% higher in quality. On a deceptively plausible task outside that frontier, accuracy fell from 84.5% without AI to 70.6% with AI and to 60% for participants also given prompt guidance. The study was published in Organization Science in 2026.
The difficult part is that an application does not display a bright border around its frontier. A fluent wrong answer feels like progress until someone checks it. Strong integrations make uncertainty, evidence, exception paths, and verification visible. They do not rely on the user to detect every boundary from prose style.
There is another hopeful result. In a field experiment with 776 Procter & Gamble professionals, individuals using AI matched the performance of teams without AI. AI also helped commercial and technical specialists produce more balanced proposals. That suggests applications can make complementary expertise more accessible, not merely accelerate typing.
There is a second trap in productivity claims: AI can change the work people choose to attempt. METR's May 2026 analysis of task substitution distinguishes speed on the old task mix, speed on the new task mix, and the value of the resulting work. It is a conceptual analysis, not another randomized productivity estimate.
Imagine a team that uses AI to generate twenty prototypes it previously would not have built. Comparing their production time with twenty manually built prototypes makes the saving look enormous. The useful business question is whether those extra prototypes improved the eventual product enough to justify creating and reviewing them. Measure a completed outcome as well as time per task.
AI often gives the largest immediate productivity gains to less-experienced workers. That can democratize expertise and let more people attempt valuable work. It can also create a paradox: if organizations automate the junior tasks through which people learn, where will the next generation of experts come from?
U.S. payroll research offers an early warning, not a verdict. A November 2025 working paper using ADP records for millions of workers found employment among ages 22-25 in highly AI-exposed occupations was 16% lower relative to comparison occupations after controlling for firm-level shocks. Employment for experienced workers in the same occupations remained stable. Declines were concentrated where observed AI use looked more automating than augmenting. The authors explicitly caution that other forces may partly explain the pattern.
An Anthropic analysis using Current Population Survey data found no systematic unemployment increase for highly exposed occupations after late 2022, but tentative evidence that hiring into them slowed for workers aged 22-25. A study linking representative surveys with Danish administrative records likewise found no detectable effect on earnings or recorded hours two years after ChatGPT, even as work reorganized around content generation, AI oversight, and integration.
The careful conclusion is neither "nothing is happening" nor "entry-level work is over." Work structure can change before aggregate unemployment or wages move. Hiring is one plausible early margin. Organizations that benefit from AI should preserve apprenticeships, supervised practice, rotation through exceptions, and opportunities to build independent judgment. A person can learn with an AI; they still need chances to work without its answer already on the screen.
Application ethics is sometimes reduced to a model's refusal behavior. At work, the more immediate questions are about power.
Who chose the tool? Who supplied the data? Who is being measured? Who receives the productivity gain? Who absorbs an error? Who can say no? Who gets to appeal?
Ethics as product design
Describe the consequence, reversibility, data, affected parties, and appeal path. The result is a proportional control set—not a moral score or compliance certificate.
3 controls surfaced from these conditions.
Selected because: The workflow handles non-public data.
Applies to: The application connects to internal sources, can call tools, or can change an external system.
Implement: Grant only necessary records, scopes, destinations, spending, and time windows; separate read from write; log tool and parameter calls; review access regularly.
Selected because: medium-consequence output needs accountable review.
Applies to: An output influences rights, safety, livelihood, money, access, professional advice, or an external publication.
Implement: Give a named, qualified reviewer adequate time, original evidence, authority to reject, and freedom from incentives that make override ceremonial.
Selected because: Non-public data crosses an application or provider boundary.
Applies to: Sensitive data crosses a connector, vendor, model provider, region, index, log, transcript, or export boundary.
Implement: Record training use, retention, human access, subprocessors, region, embeddings, backups, deletion, exports, and connector permissions separately for each plan and feature.
Absence from this generated list is not permission. Legal duties, worker agreements, professional standards, and local context may require more.
Consider four perspectives:
One person can occupy several roles. A recruiter may enjoy faster screening while applicants face an opaque ranking. A driver may receive safer routing while continuous telematics intensifies every minute. A designer may use image generation for exploration while another creator's style is invoked without credit. An employee may save two hours drafting reports and lose them to a higher quota.
The OECD's 2022 surveys of 5,334 workers and 2,053 firms in finance and manufacturing found workers were broadly positive: around 80% of AI users said it improved performance, 63% reported greater enjoyment, and 54-55% reported better mental well-being. Yet 19% of finance workers and 14% of manufacturing workers were very or extremely worried about AI-related job loss over ten years. Employers reported both automating existing tasks and creating tasks that had not existed before.
The rollout changes the experience. The OECD reports that among finance workers, 85% of those subject to algorithmic management experienced a faster work pace, compared with 74% of other AI users. A later employer survey found nearly two-thirds of managers using algorithmic-management tools had at least one trustworthiness concern, led by unclear accountability, unexplained logic, and worker-health protection.
Ethics should therefore measure more than accuracy. It includes dignity, autonomy, privacy, accessibility, attribution, workload, learning, due process, and the distribution of gains. An application can be accurate and still create a bad workplace.
An override button is useful only when a person has authority to use it, enough time to inspect the case, information independent of the model, skill to recognize a failure, and protection from punishment for slowing the line down.
Authority ladder
Choose what the application may do. A ceremonial reviewer who lacks time, information, or permission to refuse is not a safeguard.
Selected authority
Creates a proposed work product for human review.
Moving upward describes authority, not intelligence or product maturity.
For a low-stakes, reversible task, monitoring and sampling may be appropriate. For a decision that affects employment, credit, healthcare, education, liberty, or access to essential services, the control design should become stricter. People need notice, an understandable reason, correction and appeal paths, and an accountable decision owner.
Existing rules illustrate that principle. The U.S. Equal Employment Opportunity Commission explains that federal anti-discrimination law applies to employment tests and selection procedures, including automated ones: a facially neutral procedure can still be unlawful if it disproportionately excludes a protected group and is not job-related and consistent with business necessity. New York City's Local Law 144 requires covered automated employment decision tools to have a recent bias audit, make audit information public, and provide notice to candidates or employees. The city's enforcement page explains the scope.
In the European Union, Article 50 transparency duties generally have applied since August 2, 2026, including disclosure in covered human-AI interactions and important synthetic-content cases. A limited grace period moves the Article 50(2) machine-readable marking duty for certain systems already on the market to December 2, 2026; the Commission's current guidance explains those boundaries. Following the AI Omnibus entering into force, separate Annex III high-risk rules apply from December 2, 2027. These include covered employment and worker-management systems. These are jurisdiction-specific examples, not legal advice.
The practical design rule is universal: the greater the consequence and the harder the action is to reverse, the less acceptable silent automation becomes.
The AI market is a wall of overlapping promises. A better product finder begins with a job to be done.
Do you need to turn meetings into assigned actions, search a controlled knowledge base, draft from approved material, explain a spreadsheet, triage support, translate for a customer, prototype an interface, generate image or video variations, inspect a contract, test software, connect systems, or monitor a queue? Those are different workflows even when several vendors put a chat box in front of them.
Dated product finder
Search public product evidence by job, team, behavior, and authority. Inclusion is not endorsement; verify current pricing, security, accessibility, and contract terms before adoption.
28 reviewed products match · evidence frozen Aug 13, 2026
This is an editorial disclosure-coverage grade, not a score for product quality, safety, accuracy, or value. A means the reviewed primary documentation covers capability plus material control and data boundaries; A− leaves a narrower gap; B establishes the product but leaves important deployment, data, or control details for procurement; U means the public record was insufficient.
Abridge
Ambient clinical documentation that turns patient-clinician conversations into draft notes and structured workflow suggestions with linked evidence.
Deployment: Abridge describes EHR integrations and evidence links but not a request-level model, region, facility, or accelerator.
Training-data boundary: The public product pages do not fully disclose training, retention, and model-provider terms for every customer deployment; procurement diligence is required.
All jobs to be done: Draft clinical notes · Populate structured documentation · Link note content to encounter evidence
All integration surfaces: Mobile, desktop, and web · Epic and EHR workflows · Clinical notes and flowsheets
Categories: healthcare · ambient documentation · clinical workflow
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: draft
Adobe
Image, video, vector, editing, bulk-production, and brand-model tools across Adobe applications and APIs.
Deployment: Adobe distinguishes Firefly and selectable partner models; request-level cloud, region, facility, and accelerator routes remain undisclosed.
Training-data boundary: Adobe says Firefly is trained on licensed and public-domain material rather than user content; selectable partner models have separate terms.
All jobs to be done: Generate and edit images or video · Create variations · Apply brand controls
All integration surfaces: Firefly · Photoshop · Illustrator · Adobe Express · APIs · Partner models
Categories: image generation · video generation · design
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: draft
Canva
Conversational, editable design generation plus brand, connector, scheduling, research, sheet, and code workflows.
Deployment: Canva describes its Design Model and partner connections but not a request-level provider, region, facility, or hardware route.
Training-data boundary: Canva publishes safety controls; research-preview and beta terms, connected providers, and enterprise settings still require separate review.
All jobs to be done: Create layered designs · Adapt content across channels · Schedule recurring creative work · Apply brand context
All integration surfaces: Canva editor · Slack and Gmail · Drive and Notion · Calendar and Zoom · HubSpot and Microsoft
Categories: design · image generation · automation
Reviewed Aug 13, 2026 · exception review · maximum reviewed authority: act
OpenAI
General assistant for writing, analysis, files, cited research, images, and connected-app actions.
Deployment: OpenAI discloses a service and connector boundary, not the model, cloud, region, facility, or accelerator for each request.
Training-data boundary: Business, Enterprise, Edu, Healthcare, Teachers, and API data are excluded from training by default; personal Free, Plus, and Pro workspaces have a separate opt-out control.
All jobs to be done: Draft and transform content · Analyze files · Research with citations · Use connected tools
All integration surfaces: Web, mobile, and desktop · Uploaded files · Apps and connectors · Deep Research
Categories: assistant · writing · research · analysis
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: act
Anthropic
Terminal, IDE, and web coding agent that reads repositories and can edit, run, and test software.
Deployment: Anthropic documents local and cloud surfaces plus sandbox controls; request-level cloud, region, and accelerator remain undisclosed.
Training-data boundary: Commercial data is not used for training by default and has documented retention choices; consumer choices differ, and local transcripts are stored in plaintext by default.
All jobs to be done: Navigate codebases · Implement and test changes · Use development tools
All integration surfaces: Terminal · IDE · Web · MCP servers · Git providers
Categories: coding · agent
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: act
Anysphere
AI code editor with interactive and asynchronous background agents that edit and run repositories.
Deployment: Background agents clone a repository to an isolated AWS Ubuntu virtual machine with internet access; exact model and hardware routing can vary.
Training-data boundary: Privacy mode is supported, but code is retained as needed to run background tasks; changing the mode does not retroactively change an existing run.
All jobs to be done: Edit code · Run tests and commands · Delegate repository tasks
All integration surfaces: Desktop editor · GitHub repositories · Background-agent virtual machines
Categories: coding · agent
Reviewed Aug 13, 2026 · exception review · maximum reviewed authority: act
Elicit
Literature-review specialist for semantic search, screening, extraction, and evidence-linked synthesis.
Deployment: The product describes its review workflow and evaluations but does not disclose a request-level model, cloud, or hardware route.
Training-data boundary: The cited product evaluations do not establish a universal retention or training policy for every plan; verify contract terms.
All jobs to be done: Find studies · Screen papers · Extract structured evidence · Draft a review
All integration surfaces: Academic-paper corpus · Uploaded PDFs · Project exports
Categories: research · literature review · evaluation
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: recommend
Intercom
Customer-support agent that retrieves configured knowledge, generates answers, disambiguates, and hands conversations to people.
Deployment: Intercom discloses regional processing boundaries at product level, not a model, facility, or accelerator for each conversation.
Training-data boundary: The cited FAQ describes processing and regional differences but is not a substitute for plan-specific retention, subprocessor, and training terms.
All jobs to be done: Answer support questions · Resolve routine requests · Escalate exceptions
All integration surfaces: Intercom Messenger · Email · WhatsApp · SMS · Social channels · Workflows
Categories: support · customer service · agent
Reviewed Aug 13, 2026 · exception review · maximum reviewed authority: act
Multimodal assistant with research planning and optional grounding in Google Search, files, Workspace, and notebooks.
Deployment: Google does not expose a request-level model, region, data-center, or TPU trace in Gemini Apps.
Training-data boundary: Google says Workspace customer data is not used to train or improve models outside Workspace without permission; consumer Gemini policies differ.
All jobs to be done: Research a topic · Synthesize connected sources · Draft and analyze content
All integration surfaces: Web and mobile · Google Search · Files · Gmail and Drive when connected · Gemini Notebook
Categories: assistant · research · workspace
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: draft
GitHub / Microsoft
Coding assistant spanning completion, chat, code review, repository work, and cloud agents that can prepare changes.
Deployment: Model-provider hosting routes vary; repository-to-cloud-agent processing is documented, but per-request model, region, and hardware are not.
Training-data boundary: Business and Enterprise interaction data is not used for model training; individual-plan interaction data can be used unless the user opts out.
All jobs to be done: Explain and generate code · Review changes · Research a repository · Prepare a pull request
All integration surfaces: GitHub · VS Code · Visual Studio · JetBrains · Xcode · CLI
Categories: coding · code review · agent
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: act
Gong
Conversation analysis, trackers, coaching, deal signals, and forecast inputs for revenue teams.
Deployment: Gong publishes security and governance controls but not a request-level model, region, facility, or hardware route.
Training-data boundary: Gong says customer data is not used to train generative models and describes retention, redaction, and bring-your-own-key controls.
All jobs to be done: Transcribe customer interactions · Identify deal signals · Coach representatives · Support forecasts
All integration surfaces: Calls · Email · CRM · Revenue workflows
Categories: sales · analytics · meetings
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: recommend
Superhuman
Cross-application writing evaluation, drafting, and proactive assistance.
Deployment: Public documentation describes provider processing but not a request-level model, region, facility, or hardware route.
Training-data boundary: Business and Education training is off by default; provider processing and consumer controls differ by plan.
All jobs to be done: Check grammar, tone, and brand style · Draft and rewrite text · Use cross-app context
All integration surfaces: Browser and desktop writing surfaces · Email and documents · Connected business applications
Categories: writing · communication · cross-app assistant
Reviewed Aug 13, 2026 · pre action review · maximum reviewed authority: draft
Filter first by fit:
Popularity is useful evidence that a product exists and has distribution. It is not evidence that it is safe for personnel decisions, compatible with your data policy, or productive in your workflow. Prices, model providers, integrations, and retention terms can change, so the finder preserves dated sources and avoids pretending one ranking applies to everyone.
The most useful first application is often boring: a high-volume, reversible task with a clear quality standard and an obvious owner. A reliable reconciliation assistant may create more value than a general autonomous "digital employee." Small, measurable wins teach an organization how to attempt larger ones.
Once a task is selected, draw the current process. Include delays, exceptions, handoffs, and unofficial workarounds. Then decide which part AI should inform, suggest, draft, approve, act on, or monitor.
Workflow builder
Choose a workflow and assign every step. The blueprint stays in this page session; it does not send prompts, call products, save private text, or claim the workflow is ready.
Goal: Reduce repetitive handling while preserving accurate policy application, escalation, and customer recourse. · Team: Customer support
New customer message and permitted account metadata → Category, urgency, and queue
Required controls: Representative routing evaluation · Priority override · Ageing and missed-SLA monitor
Classified request and current knowledge base → Relevant policy passages with source links
Customer request and approved evidence → Proposed response with basis and uncertainty
Verified eligibility and proposed response → Refund, cancellation, account change, or escalation
Continue into evaluation; this is a promising boundary, not a production approval.
Resolve a customer-support request: Classify and route — delegate; Retrieve approved policy — assist; Draft an answer — assist; Apply a remedy — keep
A sound first deployment usually has:
Run it in shadow mode when possible: let the application produce a recommendation without changing the real outcome, then compare it with actual decisions. Pilot with people who perform the work, not only managers or an innovation team. They know which apparent shortcuts create rework downstream and which edge cases are actually the job.
The ILO's global case studies on social dialogue show worker representatives influencing employment and skills, algorithmic management, and conditions in AI supply chains. An OECD experiment in three German manufacturing firms found consultation could preserve simulated productivity gains while improving perceived job quality. Worker voice is not merely an ethical courtesy; it is a source of system knowledge.
A model benchmark cannot tell you whether a workflow saved time, improved a decision, reduced access, raised rework, or made employees miserable.
Evaluation plan
Mark checks included in the evaluation plan. A checked box means planned—not passed—and every criterion still needs representative data, owners, and thresholds.
0/8 checks included in the plan
Start with a baseline. Measure the existing process before launch, including its errors and frustrations. Then assess at least five dimensions:
Define the unit carefully. "Documents generated" rewards volume. "Tickets closed" can reward premature closure. "Lines of code" says little about working software. Measure the value delivered and the work displaced elsewhere.
Treat high-stakes evaluators differently from creative generators. A marketing ideation tool can be judged through usefulness, novelty, brand fit, and editing time. A hiring score needs job-related validity, subgroup analysis, documentation, an appeal path, and monitoring after deployment. A support agent needs both resolution quality and a safe handoff. A workflow agent needs action-level permissions and rollback tests.
NIST's voluntary AI Risk Management Framework offers a useful rhythm: Govern, Map, Measure, Manage. Its core guidance calls for defined human roles, representative evaluation, production monitoring, documented limitations, third-party risk management, and feedback from affected people. Evaluation is not a launch gate passed once. A model update, new data source, policy change, or broader user population can change the system.
GenAI spread quickly because it arrived through devices and software people already had. A nationally representative U.S. study published in January 2026 found that by late 2024, 45% of adults aged 18-64 had used GenAI. Among employed respondents, 27% used it for work in the previous week. Yet respondents' estimated time savings equaled only 1.4% of total work hours.
Business adoption shows the same combination of breadth and limited depth. U.S. Census data for November 2025 through January 2026 found 18% of firms used AI in a business function, or 32% when weighted by employment. Among adopters, 57% used it in three or fewer functions. Larger and knowledge-intensive firms were far ahead.
Adoption roadmap
Move from discovery to shadowing, assistance, bounded automation, and scale only when the current stage meets its exit criteria.
Purpose
Choose a bounded problem worth improving before choosing a product.
A responsible adoption path can be ambitious without pretending to transform everything at once:
The World Economic Forum's 2025 employer survey found 77% planned to upskill workers in response to AI disruption, 47% expected to transition affected staff into other roles, and 41% expected some workforce reduction where AI could replicate tasks. These are employer plans, not measured outcomes. The different plans reflect leaders' choices about how to use the technology.
The best outcome is not a company with the highest percentage of "AI users." It is one where people can identify the right tool, understand its limits, improve the workflow, retain accountable judgment, and share in the value created.
The Applications layer is full of claims that deserve a second question.
When a company says AI increased productivity, what was the baseline, who was studied, what was counted, and what happened to quality and workload? When it says a role was eliminated by AI, did the system perform the work, or did executives reorganize during a broader cost reduction? When a vendor says "human review," how many cases reach each reviewer, what evidence do they see, and how often are overrides accepted?
Reporter mode
A dated, vendor-neutral research snapshot—not a ranking or buying recommendation. Features, plans, integrations, model routes, retention, and controls change; verify the linked primary source and evaluate the actual workflow before deployment. Research frozen Aug 13, 2026.
12 reporting leads
product
Which model, router, tool, cloud, region, and subprocessor handled each product mode, and when did that mapping change?
A product brand can route among providers and retain different data in search, generation, agent, and connector paths.
product
For each plan and feature, what is processed, retained, viewed by people, sent to subprocessors, stored as embeddings, used for training, exported, and deleted?
“No training” answers only one of several distinct data-lifecycle questions.
product
Which exact connector and tool scopes does the application inherit, and can a shared agent reveal data outside the invoking person’s own access?
AI can amplify a preexisting access-control mistake or introduce an independent agent identity with broader permissions.
governance
At the consequential moment, can a named person see the source evidence, reject the output, stop the action, and repair the result?
A ceremonial reviewer is not a safeguard when they lack time, information, authority, or a recovery path.
governance
What representative cases, subgroup slices, error costs, expert standards, and stop thresholds were defined before the pilot?
A polished demo can hide a jagged capability frontier and average away the cases where error is most harmful.
governance
Is the same product used to draft information, recommend an outcome, or make a decision about employment, education, credit, healthcare, law, or access?
Intended purpose and authority—not the product logo—determine consequence, professional responsibility, and potential legal obligations.
work
Which tasks disappeared, expanded, moved to another role, became monitored, or were newly created—and what happened to hiring, hours, pay, pace, and learning?
Task exposure, productivity, adoption, and employment outcomes are different evidence records.
work
If AI takes the routine work, how do new workers still acquire context, judgment, relationships, and the evidence needed for promotion?
Near-term output gains can weaken the apprenticeship pipeline or concentrate opportunity among already-experienced workers.
work
Does the claimed gain include prompting, review, correction, waiting, incidents, training, integration, displaced work, and quality—not just generation time?
AI helped in several controlled tasks and slowed experienced developers in another real-work setting.
adoption
How did performance, permissions, staffing, model routing, and error change between the demo, shadow run, assisted pilot, and production rollout?
A license or benchmark is not evidence that a production workflow remains safe or valuable.
adoption
Who receives saved time or increased output, and who absorbs review, exception handling, surveillance, correction, or lost opportunity?
The same application can augment one worker while increasing pace, monitoring, or burden for another.
adoption
What do the kill-switch tests, permission revocations, rollbacks, state reconciliation, customer corrections, appeals, and incident postmortems show?
An agent with permission to act can turn a model error into an external event. Recovery tests show whether people can undo the damage.
79 dated sources
Other useful leads live between organizational layers:
The evidence ledger behind this page preserves a basis-tagged date for every source and, for the labor studies visualized here, the study population and boundary. Unknowns remain unknown. A forecast should look like a forecast; a self-report should not be colored like a randomized result. Readers can use those distinctions to decide which questions the evidence can answer.
Start with one task whose result you can check. Record how long the whole job takes, including review and rework. Ask the people doing it what improved and what became harder. Expand when those answers justify it.
I help organizations do that through practical AI integration: mapping a workflow, choosing an appropriate model route, and measuring the result. Contact me about a workflow you want to improve.
Put the application layer to work
I help companies and institutions find practical AI opportunities, choose the right application and model boundary, integrate with existing work, and build the evaluation and controls needed to improve it over time.
The useful first conversation is not “Which AI should we buy?” It is “Which work should improve, for whom, and how will we know?”
Opening the link places these non-sensitive choices in the contact-form URL; no contact message is submitted until you review and send it.
What would make the next step useful?
Use the workflow builder above and this handoff will include your non-sensitive allocation choices.
I want to know whether the people using the application can do better work and still control consequential decisions. The time spent checking and correcting a draft belongs in that assessment.
These reading notes were checked on September 17, 2026. The interactive source ledgers retain their separately dated records.
METR: why its developer experiment changed. Read alongside the early-2025 slowdown result. It explains why later participants and tasks are not directly comparable.
METR: task substitution and uplift. A framework for distinguishing time saved from value created when people change their task mix.
NIST: AI Risk Management Framework. Voluntary guidance for evaluating and managing deployed systems. A process guide, not a product certification.