Software

On AI Wiki, software refers to the use of artificial intelligence inside end-user and enterprise software. The subject includes writing assistance, search, summarization, media generation and editing, data analysis, customer support, cybersecurity, and software development. It does not replace the broader computer-science meaning of software, and detailed coverage of programming practice belongs in Software Development.

AI features in software are not one uniform technology. Some predict the next word or classify an item. Others generate text, images, audio, video, queries, or code. Retrieval-based assistants consult documents or databases before answering. Agents can select tools and carry out a sequence of actions. A single product may combine several of these approaches, and its behavior can depend on the model, connected data, permissions, administrator settings, and deployment mode.

The rapid expansion of generative AI after the public release of ChatGPT in November 2022 changed how these capabilities were presented and packaged.^[1] Conversational assistants appeared in established applications, while newer products were designed around model interaction from the start. Product names, included features, and pricing have changed frequently, so dated launch facts should not be read as descriptions of every current plan.

Scope and terminology

AI had been part of commercial software long before chat interfaces became common. Spam filtering, optical character recognition, recommendation, predictive maintenance, fraud detection, speech recognition, and image segmentation all use statistical or machine-learning methods. Gmail's Smart Compose, introduced in 2018, used a large-scale neural language model to produce low-latency writing suggestions, for example.^[2] Salesforce introduced Einstein in 2016 as predictive and recommendation capabilities embedded across its customer-relationship management platform.^[3]

The generative wave added a natural-language control surface to many of these systems. In product descriptions, several terms are often used:

Scroll sideways for more →

These labels describe product design, not a guaranteed level of capability. "Copilot" is a branding term used by several vendors. "Agent" can mean anything from a guided workflow with a small set of approved actions to a system that selects tools and iterates with limited supervision. The relevant questions are what data the system can read, which actions it can take, how outputs are checked, and who can disable or audit it.

Many enterprise assistants use retrieval-augmented generation. The original RAG research combined a generator's learned parameters with retrieved passages from an external index, partly to make knowledge easier to update and to provide provenance.^[4] Commercial implementations may retrieve from files, messages, help-center articles, customer records, source repositories, or governed data models. Retrieval can improve relevance, but it does not guarantee that a response is complete or correct. Search quality, source quality, permissions, and generation remain separate failure points.

Historical development

AI-assisted software developed through overlapping phases rather than a single starting event. Predictive systems were integrated into established products during the 2000s and 2010s. Smart Compose demonstrated interactive neural text generation at Gmail scale in 2018.^[2] Salesforce Einstein applied prediction and recommendation to sales, service, marketing, and commerce data from 2016.^[3] Meeting transcription services and machine-learning image tools likewise preceded modern large language models.

AI programming assistance was an early public use of large generative models. GitHub Copilot entered technical preview in June 2021 and became generally available to individual developers on June 21, 2022.^[5] Its initial interface emphasized inline code completion. Later versions added chat, multi-file editing, code review, terminal interaction, and cloud coding agents.

OpenAI released ChatGPT as a research preview on November 30, 2022.^[1] Its conversational interface made general-purpose drafting, summarization, explanation, and code generation accessible without a specialized application. Established vendors soon embedded comparable interactions into existing products:

• Notion made Notion AI generally available on February 22, 2023, with writing, summarization, translation, and brainstorming tools inside pages.^[6]
• Microsoft announced Microsoft 365 Copilot on March 16, 2023, showing model-assisted work across Word, Excel, PowerPoint, Outlook, and Teams.^[7] Enterprise general availability followed on November 1, 2023.^[8]
• Adobe introduced the first Adobe Firefly beta on March 21, 2023, initially focused on image and text-effect generation and planned integration across Adobe applications.^[9]
• Google announced generative writing features for Gmail and Docs in March 2023 and subsequently expanded Gemini across Workspace applications.^[10]
• Zoom launched AI Companion in September 2023 for eligible paid accounts, with meeting summaries, in-meeting questions, recording features, and chat composition.^[11]

During 2024 and 2025, product design shifted from single prompts toward longer workflows. Salesforce announced Agentforce in September 2024 and made the first version generally available in October.^[12][13] GitHub added an agent mode in Visual Studio Code and a separate cloud coding agent in 2025.^[14][15] Notion 3.0, released in September 2025, rebuilt Notion AI around agents able to operate across pages and databases.^[16]

By mid-2026, assistants and agents coexisted with conventional controls rather than replacing them. The same vendor could offer embedded features, an interactive assistant, and a more autonomous product under different names. Zoom, for example, incorporated former AI Companion functions directly into Zoom Workplace and launched ZoomMate as a separate agentic work surface in June 2026.^[17] Product history therefore needs dates and version boundaries; a launch name does not necessarily remain the current name.

Integration architecture

An AI feature inside a mature application typically sits above several existing software layers. The model rarely acts alone.

Context assembly

The application first constructs context for a request. Context may include the current document, selected text, the active spreadsheet, meeting transcript, source-code files, customer record, access-controlled messages, or retrieved search results. The system may also include instructions that define tone, output format, safety constraints, and available tools.

The quality of this step determines what the model can know. A model asked to summarize a meeting without a complete transcript cannot recover missing speech. An analytics assistant needs table descriptions, relationships, metrics, and business definitions to translate a question into a useful query. A coding agent must locate the relevant files and understand repository conventions before editing.

Model and inference layer

The model may be supplied by the software vendor, a third-party provider, or a mixture selected by task. Products also change models without changing their main interface. For that reason, a software feature should not be assumed to have the fixed behavior of one model release.

Inference can occur in the cloud, on the device, or in a hybrid arrangement. Apple Intelligence, announced in June 2024, was designed to run many requests on supported Apple devices and to send more complex requests to Private Cloud Compute.^[18] Apple's published security design says the device verifies server software before sending a request, the service does not retain request data, and outside researchers can inspect released server software.^[19] This is one architecture, not a general property of every product described as on-device or private.

Retrieval and semantic layers

Enterprise products often place retrieval between the user and the model. A support bot may search approved help-center articles. A collaboration assistant may search only messages and files the user can access. A business-intelligence assistant may query a semantic model that defines measures and relationships. The result is then passed to a generator for explanation or presentation.

Grounding should be visible where practical. Source links let a user distinguish retrieved evidence from generated prose. Generated SQL or formulas should be inspectable. Data products may expose the query and result table rather than only a narrative. These mechanisms aid review, but they do not remove the need to test whether the source was appropriate and whether the model represented it faithfully.

Tools and actions

Agents add a tool layer. Tools may search, edit a file, execute a command, create a ticket, update a customer record, schedule a meeting, run a query, or interact with a website. Safe design depends on limiting each tool to the permissions needed for the task and separating read operations from consequential writes.

Tool execution also creates a new distinction between an inaccurate answer and an incorrect action. A mistaken draft can be rejected. A mistaken payment, deleted record, sent message, or production deployment can be harder to reverse. Approval prompts, previews, transaction limits, isolated environments, audit logs, and rollback mechanisms are therefore product functions, not cosmetic safeguards.

Data, permissions, and administration

Consumer and enterprise editions of the same product can have different data terms and controls. Organizations evaluating an AI feature commonly need to determine:

• whether prompts and outputs are retained, and for how long;
• whether customer content is used to train or improve models;
• where inference and storage occur;
• which model providers or subprocessors receive data;
• whether existing document, channel, record, and repository permissions are enforced;
• whether administrators can enable features by group, restrict models or connectors, and view audit events;
• whether generated actions require approval;
• how users can delete history, revoke a connector, or report an incorrect result.

The permission model is especially important for retrieval. Slack's current documentation says its AI answers include only content available to the requesting user, including private channels that user belongs to and direct messages involving that user.^[20] It also allows administrators to manage access to AI features. This is the intended product policy; organizations still need to configure identity, sharing, retention, and connectors correctly.

NIST distinguishes several AI-specific or intensified risks from traditional software risks, including dependence on data quality, model and concept drift, opacity, difficult-to-predict failure modes, and challenges in testing statistical behavior.^[21] Its voluntary AI Risk Management Framework organizes work around Govern, Map, Measure, and Manage. The 2024 Generative AI Profile applies those functions to risks such as confabulation, privacy, information integrity, security, and intellectual property.^[22] These frameworks do not certify a product. They provide a structure for deciding what should be measured, who owns a risk, and what happens when a system fails.

Productivity and collaboration software

Productivity suites made AI visible inside familiar documents, spreadsheets, presentations, email, calendars, and team chat. The integration pattern is usually contextual: the assistant can work with the item already open rather than requiring the user to copy material into a separate chatbot.

Scroll sideways for more →

Microsoft introduced Microsoft 365 Copilot as an assistant across application and Microsoft Graph context in March 2023.^[7] General availability for enterprise customers began in November 2023.^[8] The suite has since added agents and specialized experiences, so the 2023 demonstrations are best understood as launch history rather than a complete current feature list.

Google's packaging changed materially after the first Duet AI and Gemini add-ons. In January 2025, Google said it was including Gemini capabilities in Workspace Business and Enterprise plans and ending separate Gemini Business and Enterprise add-ons for those customers.^[23] Premium or experimental capabilities can still have separate availability. A historical add-on price should therefore not be presented as the current universal way to buy Workspace AI.

Notion's path shows the transition from an assistant to an agent. The February 2023 release centered on generating and transforming content within a page.^[6] Notion 3.0 in September 2025 described a personal agent able to work across pages and databases and announced planned custom agents that could run through schedules or triggers, subject to instructions and permissions.^[16] These are different interaction models under the continuing Notion AI name.

Slack similarly expanded from thread summaries and search to a larger set of plan-dependent features. Its documentation lists summaries and huddle notes across current paid plans, while search answers, recaps, workflow generation, file summaries, and enterprise search depend on plan.^[20] Search answers provide citations to the underlying messages or files, which makes retrieval visible to the user.

Apple began releasing Apple Intelligence features in October 2024 after announcing them in June.^[18][24] The design combines device models, Private Cloud Compute for some requests, and explicit user permission before a request or attached material is sent to ChatGPT through the system integration.^[18] Feature availability has continued to vary by device, operating-system version, language, and country. Claims about an announced personalized Siri should be separated from features Apple has actually released.

Creative software

Creative applications combine long-standing machine-learning tools, such as subject selection and noise removal, with newer generation and transformation models. An AI label can therefore cover very different operations.

Adobe launched Firefly in beta in March 2023 with image and text-effect models and plans to embed them across Creative Cloud, Document Cloud, Experience Cloud, and Adobe Express.^[9] Photoshop added Generative Fill in public beta in May 2023, placing prompt-based additions, removals, and extensions inside its layer-based editing workflow.^[25] Premiere Pro's Generative Extend moved from beta to general availability in April 2025, with support for extending video and audio clips and for 4K and vertical video.^[26] It should not be confused with the separate Firefly web application's text-to-video generation.

Adobe says its Firefly foundation models are trained on licensed and public-domain content rather than on customer content.^[27] Model versions and partner models can have different training disclosures, so the statement should be attached to Adobe's specified models rather than generalized to every model accessible through an Adobe interface.

Canva AI brought several workflows together in Magic Studio in October 2023. Magic Design creates draft layouts, Magic Media generates images or short video, Magic Grab separates editable subjects, Magic Expand extends an image beyond its frame, and Magic Write generates or transforms text.^[28] Canva also described administrator controls and opt-in rather than default use of customer content for training its models in that launch material.

DaVinci Resolve illustrates the other side of the category: many AI functions analyze existing media instead of generating a new scene. Blackmagic Design announced Resolve 20 in April 2025 with AI IntelliScript for assembling a timeline from a script, Animated Subtitles, Multicam SmartSwitch, Audio Assistant, Dialogue Matcher, Music Editor, Magic Mask 2, and other analysis and editing tools.^[29] Resolve 20 did not introduce the text-to-image generator previously claimed on this page.

ByteDance's CapCut combines conventional editing with AI functions for generated media, captions, speech, and avatars.^[65] Feature availability and names vary by region and plan. This article therefore treats CapCut as an example rather than maintaining a volatile inventory of every tool.

Content provenance is a related software function. The Coalition for Content Provenance and Authenticity publishes the C2PA specification for attaching signed, tamper-evident provenance records called Content Credentials to digital assets.^[30] A credential can record assertions about origin and editing history. It does not determine whether the depicted event is true, and missing credentials do not prove that media is synthetic.

Meetings and communication

Meeting assistants combine speech recognition, speaker attribution, summarization, retrieval, and workflow integration. Their output depends on audio quality, transcript completeness, supported language, participant consent, and access to the recording.

Zoom introduced AI Companion in September 2023 with meeting summaries, in-meeting questions, smart recording, and chat composition for eligible paid accounts.^[11] In June 2026, Zoom said these functions had been incorporated into Zoom Workplace under descriptive names, while ZoomMate became a separate product for search, content creation, and cross-application actions based on meeting and enterprise context.^[17] Calling every current Zoom AI function "AI Companion" would therefore be outdated.

Microsoft Teams recap brings together recordings, transcripts, files, notes, summaries, and follow-up tasks. Microsoft states that intelligent recap uses the transcript, attendance data, and PowerPoint Live data, and warns that generated content can be inaccurate, incomplete, or inappropriate.^[31] Some features require Teams Premium or a Microsoft 365 Copilot license, and transcription is required for AI notes and summaries.

Google Meet's "Take notes for me" records discussion points and action items in a Google Doc and associates the notes with the Calendar event.^[32] The workflow is useful because the output remains editable and shareable, but generated notes still need review for names, decisions, and assigned actions.

Independent products include Otter.ai, Fireflies.ai, Granola, Krisp, and Read. Otter announced Meeting GenAI in February 2024, including chat across a user's meeting history rather than only one transcript.^[33] Products differ in whether a bot joins the call, whether audio is captured locally, which conferencing services are supported, and where recordings are stored. Those differences can matter more to an organization than the summary format.

Browsers and operating systems

Browsers and operating systems expose AI to especially broad context. A browser assistant can read a page or several tabs. A browser agent may click controls or fill forms in authenticated sessions. An operating-system feature may work across applications and local files. The resulting utility comes with a larger permission and security boundary.

Google began rolling Gemini in Chrome out to English-language Mac and Windows users in the United States in September 2025. The assistant was designed to work across tabs and integrate with services such as Docs and Calendar.^[34] Earlier Chrome AI features, such as writing help and tab organization, should not be conflated with this later multi-tab assistant.

OpenAI launched ChatGPT Atlas for macOS on October 21, 2025.^[35] Atlas included optional browser memories and a preview agent mode for selected paid plans. The launch post said Windows, iOS, and Android versions were forthcoming, so those platforms should not be described as part of the initial release. OpenAI also warned that an agent operating in logged-in sites could expose data or take unintended actions and that safeguards would not stop every attack.^[35]

Opera's Aria and Brave's Leo were earlier examples of assistants embedded in conventional browsers.^[66][67] Their supported models, quotas, and plans have changed since launch, which makes historical launch pricing a poor description of the current products. AI-first browser projects have also changed direction quickly; an enduring category description should focus on capabilities and permissions rather than a ranking of vendors.

Microsoft announced Copilot+ PCs and Recall in May 2024, but did not ship Recall with the initial devices. After security and privacy criticism, Microsoft changed the design and conducted an Insider preview. Recall became available as an opt-in preview through the April 2025 Windows update on eligible Copilot+ PCs.^[36] Current documentation says snapshots are processed and stored locally, encrypted, protected through Windows Hello, and filterable or deletable by the user.^[37] Managed devices have separate administrator policies, and Recall is removed by default unless an organization allows it.^[38]

This chronology matters. The initial announcement, the pre-release security concerns, and the redesigned 2025 preview describe different versions. It is inaccurate either to present the original design as the current architecture or to erase the concerns that led to the redesign.

Analytics and business intelligence

AI in analytics often translates natural-language questions into queries, formulas, visualizations, or narrative summaries. Reliable use requires more than a language model. Systems need an understandable data model, documented measures, permissions, and a way to inspect generated queries and results.

Power BI Copilot supports report creation, analysis, DAX assistance, and narrative functions. Microsoft's current documentation requires a supported paid Fabric capacity of F2 or higher or Power BI Premium capacity of P1 or higher, together with administrator enablement and supported regional deployment.^[39] Some experiences are generally available while others remain in preview. Product status should therefore be checked at the feature level.

Tableau's product is now called Tableau Agent, not Einstein Copilot for Tableau. It uses a conversational interface to explore data, create visualizations, and create or explain calculations. Current documentation lists support in Tableau Desktop from version 2025.1, Tableau Server from version 2025.3, and configured Tableau Cloud environments, with limitations by role, data source, language, and feature status.^[40]

Looker's Conversational Analytics uses Gemini to interpret questions and answer from Looker data.^[41] Its semantic layer can supply governed definitions and relationships, but Google explicitly warns in documentation that generated answers can sound plausible and still be wrong. Analysts should inspect data and calculations before using an answer for a consequential decision.

Databricks renamed Genie Spaces to Genie Agents in July 2026. A Genie Agent is a domain-specific natural-language interface that uses curated Unity Catalog datasets, sample queries, expressions, and instructions to produce SQL, result tables, and visualizations.^[42] Databricks distinguishes it from Genie One, a broader business interface, and Genie Code for technical work. Using the current names avoids treating these products as one undifferentiated "Genie" feature.

Snowflake Cortex Analyst similarly converts natural-language questions to SQL against a semantic model. Snowflake's documentation emphasizes role-based access and recommends semantic views or model files that describe business concepts and relationships.^[43] Across these products, the model can lower the barrier to querying data, but it cannot repair an ambiguous metric or poor source data by itself.

CRM, sales, marketing, and customer service

Customer-relationship software had predictive AI before generative chat. Salesforce Einstein's 2016 launch covered lead scoring, recommendations, forecasting, case classification, and other embedded predictions.^[3] This history is distinct from Einstein Copilot and Agentforce.

Salesforce AI announced Agentforce on September 12, 2024 as configurable agents for service, sales, marketing, and commerce.^[12] The initial product became generally available on October 29.^[13] Agentforce can connect model reasoning to Salesforce data, Flow, Apex, APIs, and other actions. Its output and authority depend on how an organization configures data access, topics, instructions, and tools.

HubSpot launched Breeze in September 2024 as an umbrella for embedded AI, including a copilot, role-oriented agents, and Breeze Intelligence for enrichment.^[44] Microsoft Dynamics 365 Customer Service also embeds assistants and agents for tasks such as retrieving case information, summarizing context, drafting responses, and updating records.^[68] Conversation-intelligence and marketing products use models to summarize calls, draft follow-ups, extract fields, and prepare content.

Customer support is a natural retrieval use case because a company can ground answers in its own help center. Intercom announced Fin in March 2023 as a GPT-4-based support bot that answers from approved support content, links to source articles, and hands unresolved questions to human staff.^[45] Intercom acknowledged at launch that hallucination, latency, and cost remained limitations. Zendesk announced Zendesk AI in May 2023 with bots, agent assistance, intent detection, and triage, combining proprietary models with large language models.^[46]

Salesforce Agentforce, Intercom Fin, Zendesk AI, Microsoft Dynamics, and other services also support agents that can carry out approved steps in a service workflow. Marketing claims about resolution rates are not directly comparable unless the products are evaluated on the same queries, knowledge bases, escalation policy, and definition of resolution. Revenue and valuation figures do not establish product accuracy and are omitted here.

The best public evidence for workplace effects is narrower than a category-wide claim. A peer-reviewed study of 5,172 customer-support agents found that access to one generative assistant increased issues resolved per hour by 15 percent on average, with larger gains for less experienced and lower-skilled workers.^[47] The authors caution that the result concerns one tool, one firm, and one occupation. It does not show that every support agent or every support product will achieve the same result.

Security software

Security products use language models to summarize incidents, translate natural-language questions into searches, explain scripts, generate reports, and recommend response steps. The model is normally layered over telemetry, threat intelligence, case records, and existing detection systems.

Microsoft announced general availability of Copilot for Security for April 1, 2024. The product integrates with Microsoft security services and third-party sources and supports promptbooks and knowledge-base connections.^[48] Microsoft reported results from its own randomized study, but those figures should be attributed to Microsoft and not generalized to all security operations.

CrowdStrike introduced Charlotte AI in May 2023 as a natural-language interface over the Falcon platform and its security data.^[49] Other vendors, including SentinelOne, Palo Alto Networks, Darktrace, Snyk, and GitHub, offer model-assisted investigation, detection explanation, or code remediation. These products do not replace the underlying sensors, access controls, or human incident authority.

AI also creates security risks for the software that embeds it. The 2025 OWASP Top 10 for LLM Applications includes prompt injection, sensitive-information disclosure, supply-chain risk, data and model poisoning, improper output handling, and excessive agency.^[50] An assistant with tools should be treated as both a user interface and a security principal: it needs least privilege, input and output controls, logging, and human approval for high-impact actions.

Developer tools

Developer tools are discussed more fully in App Development and Web Development. This section distinguishes several current interaction models rather than ranking products by transient revenue or seat counts.

Inline completion and chat. GitHub Copilot began with generated code suggestions and later added chat and multi-file editing.^[5] JetBrains AI Assistant, Amazon Q Developer, and other IDE integrations use a similar combination of code context, completion, explanation, and transformation.

Local or IDE agents. Cursor provides an agent that can search a codebase, edit multiple files, run terminal commands, and review diffs, with separate read-only and configurable modes.^[51] Cline is an Apache-2.0 open-source agent available in IDE and terminal forms; it can read and write files, run commands, use a browser, and connect to different model providers, with approval controls in interactive use.^[52] Aider is a terminal pair programmer designed to edit a local Git repository.^[53]

Cloud coding agents. Devin, introduced by Cognition in March 2024, combined a shell, editor, and browser in a sandboxed environment and reported progress to the user.^[54] GitHub's cloud coding agent, introduced in public preview in May 2025, works in a GitHub Actions environment, makes changes, runs available tests or linters, and opens a pull request for review.^[15] Cursor also offers background agents that clone a repository into isolated remote environments.^[55]

Agent-oriented IDEs. Windsurf was built by Codeium, later renamed Windsurf. It was not acquired by OpenAI. Cognition signed an agreement to acquire Windsurf's product, intellectual property, brand, business, and team in July 2025.^[56] In June 2026, Cognition introduced Devin Desktop as the next generation of Windsurf, keeping IDE compatibility while adding a surface for managing local and cloud agents.^[57]

Coding agents vary in where code runs, which repositories they can access, whether commands require approval, how secrets are supplied, and whether network access is enabled. A remote agent with repository write permission and unrestricted command execution has a wider attack surface than an inline completion tool. Tests, static analysis, code review, and deployment controls remain necessary even when an agent says it has verified its work.

Generated code also creates maintenance questions. A patch may pass existing tests while violating an unstated architectural requirement. A model may introduce an insecure dependency, reproduce an obsolete API, or make a broad change to satisfy a narrow prompt. Repository instructions and examples can improve alignment, but the maintainer remains responsible for understanding the change that is merged.

Evidence on productivity and quality

Research does not support a single percentage for "AI productivity." Effects vary by task, worker experience, model, interface, and measurement method.

A controlled experiment asked developers to implement a JavaScript HTTP server. Participants with GitHub Copilot completed that bounded task 55.8 percent faster than the control group.^[58] This is evidence for one programming task and participant sample, not for all professional software development.

A 2025 randomized study by METR examined 16 experienced open-source developers completing 246 tasks in repositories they knew. In that setting, access to early-2025 AI tools increased completion time by 19 percent, even though participants believed the tools had made them faster.^[59] METR explicitly cautioned against treating the result as evidence about most developers or all tasks. The result also represents tools available during a specific period.

Outside programming, a preregistered experiment with 453 college-educated professionals found that ChatGPT reduced time on selected writing tasks by 40 percent and increased evaluated output quality by 18 percent.^[60] The customer-support field study found a 15 percent average productivity increase and substantial differences by worker experience.^[47] Together, these studies show that assistance can improve some bounded workflows while slowing others.

Quality measurement must accompany speed. For writing, that can mean factual review and task-specific scoring. For code, it can include tests, security checks, maintainability, and review time. For analytics, it includes query correctness and metric interpretation. Time saved generating a first draft may be offset by time spent detecting a subtle error.

Reliability, security, and governance

Incorrect or unsupported output

Large language models can produce fluent statements that are not supported by their input or by fact. A major research survey defines hallucination in natural-language generation as output that is nonsensical or unfaithful to the provided source and reviews causes, measurement, and mitigation across tasks.^[61] Product interfaces may call the same behavior an inaccuracy or confabulation.

Retrieval, citations, constrained output formats, and tool-based calculation can reduce some errors. None is a complete solution. A citation can point to a source that does not support the sentence. A generated query can be syntactically valid but use the wrong field. A meeting summary can omit a disagreement. High-impact output needs an appropriate reviewer or deterministic check.

Prompt injection and excessive agency

When software retrieves emails, documents, web pages, or tickets, those data can contain text that the model interprets as instructions. Researchers demonstrated this indirect prompt-injection problem in 2023 against model-integrated applications and showed possible effects including data theft and unintended tool use.^[62] The problem arises because natural-language systems can blur the boundary between data and commands.

Least privilege limits damage. An assistant that only needs to summarize email should not receive a tool that can send messages. An analytics agent should not obtain write access to production tables merely because its query tool supports it. High-risk actions should use explicit approval, and tool arguments should be validated by ordinary software rather than trusted solely because a model generated them. OWASP makes similar recommendations for excessive agency.^[50]

Privacy and data governance

AI features can expose previously fragmented data through one conversational interface. Even when the system enforces permissions correctly, a concise answer can make sensitive information easier to discover. Organizations should audit connectors, group membership, public links, old files, retention, and the scope of enterprise search before broad deployment.

Training policy is a separate question from inference privacy. A vendor may promise not to train a generative model on customer content while still processing that content to provide a requested feature. Retention, abuse monitoring, support access, and product-improvement programs need their own review.

Adobe's June 2024 terms update illustrates the importance of precise language. After customers objected to broad wording about content access, Adobe said it did not train generative AI on customer content and would clarify ownership and license terms.^[63] Its current general terms say customers retain ownership and that the license to operate the service does not permit Adobe to train generative AI on customer content unless it was submitted to Adobe Stock.^[64] This policy should be described from the legal and help-center text, not from an unsupported social-media impression.

Provenance and intellectual property

Generated media and code can raise questions about training data, similarity, attribution, and authorized use. Vendor indemnity or "commercially safe" marketing applies only under stated contract conditions and does not prove that every output is free of third-party rights.

Content Credentials can preserve signed provenance assertions through supported editing workflows.^[30] They can help a recipient inspect claimed origin or changes, but they are not a universal detector. Governance also includes keeping the original source, prompt or brief, approvals, and applicable license information when those records matter.

Evaluation and change management

Model-backed software can change without the application receiving a conventional major-version number. A vendor may replace a default model, alter a system prompt, add tools, or modify retrieval. Evaluation should therefore continue after launch.

A practical evaluation uses representative tasks and measures both success and failure. Useful records include task completion, factual or query accuracy, false-positive and false-negative rates, time including review, escalation rate, user overrides, harmful-action attempts, and subgroup performance where relevant. Red-team testing is valuable for foreseeable misuse, but production monitoring and an incident process are still needed.

Users also need a clear fallback. If an assistant is unavailable, inaccurate, or not authorized for sensitive data, the underlying application should remain usable. Human escalation is particularly important in customer support, security response, financial analysis, and publishing.

Packaging and product evaluation

AI software is sold through several mechanisms: inclusion in a base subscription, a per-user add-on, usage credits, capacity-based billing, or metered model consumption. Vendors change these mechanisms frequently. Google moved core Gemini features into commercial Workspace plans in 2025,^[23] while Power BI Copilot remains tied to supported organizational capacity.^[39] Historical prices are useful only when explicitly dated.

A procurement comparison should separate at least five dimensions:

1. the workflow and measurable task the feature is expected to improve;
2. the sources and permissions it can access;
3. the actions it can take and the approvals those actions require;
4. the evaluation evidence and known failure modes;
5. the full cost, including model usage, storage, connectors, review, security, and administration.

Feature checklists alone can be misleading because two products may both advertise "agents" while offering very different tools, execution environments, and control boundaries. A short pilot on representative work, with predefined quality and safety criteria, is more informative than a vendor demonstration.

See also

• Software ChatGPT Plugins
• Software Development
• App Development
• Web Development
• Productivity
• GitHub Copilot
• Microsoft 365 Copilot
• Adobe Firefly
• Notion AI
• Apple Intelligence
• Gemini
• Salesforce
• Salesforce AI
• Copy.ai
• Sierra AI

References

1. ^OpenAI. "Introducing ChatGPT." November 30, 2022. openai.com/...chatgpt
2. ^Chen, Mia Xu, et al. "Gmail Smart Compose: Real-Time Assisted Writing." KDD 2019. research.google/...pose-real-time-assisted-writing
3. ^Salesforce. "Salesforce Introduces Salesforce Einstein: Artificial Intelligence for Everyone." September 19, 2016. salesforce.com/...ficial-intelligence-for-everyone
4. ^Lewis, Patrick, et al. "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks." Advances in Neural Information Processing Systems 33, 2020. proceedings.neurips.cc/...bc26945df7481e5-Abstract
5. ^GitHub. "GitHub Copilot is generally available to all developers." June 21, 2022. github.blog/...nerally-available-to-all-developers
6. ^Notion. "Notion AI, for everyone." February 22, 2023. notion.com/...2023-02-22
7. ^Microsoft. "Introducing Microsoft 365 Copilot: your copilot for work." March 16, 2023. blogs.microsoft.com/...pilot-your-copilot-for-work
8. ^Microsoft. "Announcing Microsoft 365 Copilot general availability and Microsoft 365 Chat." September 21, 2023. microsoft.com/...ailability-and-microsoft-365-chat
9. ^Adobe. "Adobe Unveils Firefly, a Family of New Creative Generative AI Models." March 21, 2023. news.adobe.com/...ly-of-new-creative-generative-ai
10. ^Google Workspace. "Announcing new generative AI experiences in Google Workspace." March 14, 2023. workspace.google.com/...generative-ai
11. ^Zoom. "Zoom introduces Zoom AI Companion." September 5, 2023. news.zoom.com/zoom-ai-companion
12. ^Salesforce. "Salesforce Unveils Agentforce." September 12, 2024. salesforce.com/...agentforce-announcement
13. ^Salesforce. "Salesforce's Agentforce Is Here." October 29, 2024. salesforce.com/...eneral-availability-announcement
14. ^GitHub. "GitHub Copilot: The agent awakens." February 6, 2025. github.blog/...github-copilot-the-agent-awakens
15. ^GitHub. "GitHub Copilot coding agent in public preview." May 19, 2025. github.blog/...ilot-coding-agent-in-public-preview
16. ^Notion. "Notion 3.0: Agents." September 18, 2025. notion.com/...2025-09-18
17. ^Zoom. "Meet ZoomMate: Zoom's next chapter in AI." June 24, 2026. zoom.com/...zoom-ai-companion-zoommate
18. ^Apple. "Introducing Apple Intelligence for iPhone, iPad, and Mac." June 10, 2024. apple.com/...-intelligence-for-iphone-ipad-and-mac
19. ^Apple Security Research. "Private Cloud Compute: A new frontier for AI privacy in the cloud." June 10, 2024. security.apple.com/...private-cloud-compute
20. ^Slack. "Guide to AI features in Slack." Accessed July 28, 2026. slack.com/...2548883-Guide-to-AI-features-in-Slack
21. ^National Institute of Standards and Technology. "How AI Risks Differ from Traditional Software Risks." AI RMF 1.0 Appendix B. airc.nist.gov/...r-from-traditional-software-risks
22. ^National Institute of Standards and Technology. "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile." NIST AI 600-1, July 2024. nvlpubs.nist.gov/...NIST.AI.600-1.pdf
23. ^Google Workspace. "Empowering businesses with AI." January 15, 2025. workspace.google.com/...owering-businesses-with-ai
24. ^Apple. "Apple Intelligence is available today on iPhone, iPad, and Mac." October 28, 2024. apple.com/...vailable-today-on-iphone-ipad-and-mac
25. ^Adobe. "Adobe Unveils Future of Creative Cloud With Generative AI as a Creative Co-Pilot in Photoshop." May 23, 2023. news.adobe.com/...ve-Co-Pilot-in-Photoshop-default
26. ^Adobe. "New AI Innovation in Adobe Premiere Pro." April 2, 2025. news.adobe.com/...new-ai-innovation-in-industry
27. ^Adobe. "Adobe Firefly generative AI approach." Accessed July 28, 2026. adobe.com/...gen-ai-approach
28. ^Canva. "Introducing Magic Studio: the power of AI, all in one place." October 4, 2023. canva.com/...magic-studio
29. ^Blackmagic Design. "Blackmagic Design Announces DaVinci Resolve 20." April 4, 2025. blackmagicdesign.com/...20250404-02
30. ^Coalition for Content Provenance and Authenticity. "Content Credentials, C2PA Specifications 2.4." Accessed July 28, 2026. spec.c2pa.org/...ContentCredentials
31. ^Microsoft Support. "Recap in Microsoft Teams." Accessed July 28, 2026. support.microsoft.com/...recap-in-microsoft-teams
32. ^Google Workspace. "AI Note Taking and Take Notes for Me in Google Meet with Gemini." Accessed July 28, 2026. workspace.google.com/...ai-note-taking
33. ^Otter.ai. "Otter.ai Takes Aim at Goliath with New Meeting GenAI Launch." February 13, 2024. otter.ai/...ot-to-redefine-ai-meeting-intelligence
34. ^Google. "Chrome: The browser you love, reimagined with AI." September 18, 2025. blog.google/...chrome-reimagined-with-ai
35. ^OpenAI. "Introducing ChatGPT Atlas." October 21, 2025. openai.com/...introducing-chatgpt-atlas
36. ^Microsoft. "Copilot+ PCs are the most performant Windows PCs ever built, now with more AI features." April 25, 2025. blogs.windows.com/...es-that-empower-you-every-day
37. ^Microsoft Support. "Privacy and control over your Recall experience." Accessed July 28, 2026. support.microsoft.com/...er-your-recall-experience
38. ^Microsoft Learn. "Manage Recall." Accessed July 28, 2026. learn.microsoft.com/...manage-recall
39. ^Microsoft Learn. "Copilot for Power BI overview." Accessed July 28, 2026. learn.microsoft.com/...copilot-introduction
40. ^Tableau. "Explore Your Data with Tableau Agent." Accessed July 28, 2026. help.tableau.com/...web_author_einstein
41. ^Google Cloud. "Conversational Analytics in Looker overview." Accessed July 28, 2026. docs.cloud.google.com/...tional-analytics-overview
42. ^Databricks. "Genie Agents." Updated July 21, 2026. docs.databricks.com/...genie-agents
43. ^Snowflake. "Cortex Analyst." Accessed July 28, 2026. docs.snowflake.com/...cortex-analyst
44. ^HubSpot. "HubSpot Launches New AI, Breeze." September 18, 2024. ir.hubspot.com/...ze-plus-hundreds-product-updates
45. ^Intercom. "Introducing Fin: Intercom's breakthrough AI chatbot, built on GPT-4." March 14, 2023. intercom.com/...announcing-intercoms-new-ai-chatbot
46. ^Zendesk. "Zendesk announces powerful AI designed exclusively for intelligent CX." May 10, 2023. zendesk.com/...zendesk-ai
47. ^Brynjolfsson, Erik, Danielle Li, and Lindsey Raymond. "Generative AI at Work." The Quarterly Journal of Economics 140, no. 2, 2025. doi.org/...qjae044
48. ^Microsoft. "Microsoft Copilot for Security is generally available on April 1, 2024." March 13, 2024. microsoft.com/...pril-1-2024-with-new-capabilities
49. ^CrowdStrike. "CrowdStrike Introduces Charlotte AI." May 30, 2023. ir.crowdstrike.com/...tte-ai-deliver-generative-ai
50. ^OWASP Foundation. "OWASP Top 10 for Large Language Model Applications 2025." 2025. owasp.org/...OWASP-Top-10-for-LLMs-v2025.pdf
51. ^Cursor. "Agent overview." Accessed July 28, 2026. docs.cursor.com/...overview
52. ^Cline. "Cline." GitHub repository and documentation. Accessed July 28, 2026. github.com/...cline
53. ^Aider. "Aider Documentation." Accessed July 28, 2026. aider.chat/docs
54. ^Cognition. "Introducing Devin, the first AI software engineer." March 12, 2024. cognition.com/...introducing-devin
55. ^Cursor. "Background Agents." Accessed July 28, 2026. docs.cursor.com/background-agent
56. ^Cognition. "Cognition's acquisition of Windsurf." July 14, 2025. cognition.com/...windsurf
57. ^Cognition. "Introducing Devin Desktop." June 2, 2026. cognition.com/...introducing-devin-desktop
58. ^Peng, Sida, et al. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." 2023. arxiv.org/...2302.06590
59. ^Becker, Joel, et al. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." METR, July 10, 2025. metr.org/...early-2025-ai-experienced-os-dev-study
60. ^Noy, Shakked, and Whitney Zhang. "Experimental evidence on the productivity effects of generative artificial intelligence." Science 381, no. 6654, 2023. doi.org/...science.adh2586
61. ^Ji, Ziwei, et al. "Survey of Hallucination in Natural Language Generation." ACM Computing Surveys 55, no. 12, 2023. doi.org/...3571730
62. ^Greshake, Kai, et al. "Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection." Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security, 2023. doi.org/...3605764.3623985
63. ^Adobe. "Here's what to know about Adobe's Terms of Use updates." June 10, 2024. blog.adobe.com/...updating-adobes-terms-of-use
64. ^Adobe. "General Terms of Use." Accessed July 28, 2026. adobe.com/...terms
65. ^CapCut. "CapCut AI Editing Features." Accessed July 28, 2026. capcut.com/features
66. ^Opera. "Opera unveils integrated browser AI: Aria." May 24, 2023. press.opera.com/...eils-integrated-browser-ai-aria
67. ^Brave. "Leo, Brave's AI browser assistant, is now available to desktop users." November 2, 2023. brave.com/...leo-launch
68. ^Microsoft Learn. "Welcome to Dynamics 365 Customer Service." Updated April 21, 2026. learn.microsoft.com/...overview