OpenAI’s move to launch persistent “Dots” agents pushes artificial intelligence beyond the familiar chat window and deeper into everyday business workflows, raising a more practical question than whether agents can impress on stage: can companies trust software that keeps working after the user has moved on?
At its annual Developer Day in San Francisco on September 29, OpenAI introduced always-on agents called Dots, which the company said can pursue user goals across applications rather than waiting for a new prompt at every step. Reuters reported that the agents are powered by GPT-6 Astra and can work with services including Slack and Microsoft Teams as well as OpenAI tools such as Codex and ChatGPT Work. The launch puts persistence, rather than another incremental chatbot feature, at the centre of OpenAI’s enterprise pitch.
The shift is important because an agent that remains active changes the economics and governance of automation. A chatbot usually waits for a person to ask a question, reviews the immediate context and returns an answer. A persistent agent may instead monitor a project, revisit an objective, collect information from several applications and continue work over time. That can reduce repetitive coordination, but it also makes reliability, permissions and auditability part of the product rather than optional administrative details.
OpenAI’s own DevDay announcement established September 29 as the date of its 2026 developer gathering. The event is significant less because one company has declared the next software interface than because enterprise AI is moving toward a model in which the assistant is expected to act between conversations. That transition will be measured by completed work, not by the fluency of a demonstration.
The agent leaves the chat window
The defining idea behind Dots is continuity. In a conventional interaction, the user opens a tool, describes a task and waits for a result. Persistent agents are designed to remain associated with a goal even when the user is no longer actively supervising each individual step. That can make them useful for projects whose inputs arrive gradually, such as collecting sales updates, tracking a software release, preparing recurring reports or coordinating information across departments.
The attraction for employers is straightforward. Knowledge work contains a large amount of small administrative activity that sits between larger decisions: requesting a missing number, checking whether a file changed, reconciling comments, updating a status board or notifying a colleague that an input has arrived. Automating those transitions can save time even when the agent never makes an important business judgment on its own.
But persistence creates a new standard for usefulness. A system that is excellent for ten minutes and unreliable over ten hours may be less valuable than a simpler tool that performs a narrow function consistently. Enterprise buyers therefore need to evaluate agents on long-running task completion, recovery from missing data and the quality of handoffs to humans, rather than on single-turn benchmark performance alone.
Cross-application work is where the value is supposed to appear
Reuters reported that Dots can interact through workplace applications such as Slack and Teams while drawing on OpenAI products including Codex and ChatGPT Work. The commercial logic is clear: companies do not perform their work inside one AI interface. Information is divided among messaging systems, documents, source-code repositories, ticketing tools, calendars and databases. An agent becomes more useful when it can move across those boundaries without forcing the employee to copy every piece of context manually.
That integration is also where complexity rises. Each application has its own permissions, data structure and history. A message in a team channel may be informal discussion, while a record in a customer database may be authoritative. An effective agent has to understand not only what information says, but what role that information plays inside the organization. Treating every source as equally trustworthy can produce polished but operationally wrong results.
Enterprises will therefore need to decide which systems an agent can read, which it can modify and which require explicit approval. A persistent assistant that can summarize a project is fundamentally different from one that can change the project’s official status. The first is an information tool; the second participates in business process. Procurement and governance should reflect that distinction.
The product is competing with workflow software, not only other chatbots
OpenAI’s move places it in a broader contest over who becomes the coordination layer for business software. Reuters described Dots as part of an autonomous-AI push that increases competition with Meta and other technology companies building persistent agents. The relevant rivals also include established enterprise-software vendors, which already control the systems where employees approve expenses, manage customers, assign tasks and measure performance.
This competition matters because the agent that understands a workflow can become a new interface to many underlying applications. Instead of opening five systems, a user might state an objective once and receive a consolidated result. If that pattern becomes normal, the economic value may shift from the application screen toward the orchestration layer that decides which system to use and when.
Incumbent software companies are unlikely to surrender that role quietly. They have advantages in permissions, structured data and customer relationships. AI developers have advantages in general reasoning and natural-language interaction. The likely enterprise market is therefore not a simple winner-takes-all replacement of existing software, but a negotiation over where intelligence lives and which vendor controls the user’s most important workflow.
A live demonstration is not an operating record
Reuters noted that some live demonstrations at DevDay encountered technical glitches. That is not unusual at a technology event, but it is relevant to the product category. The promise of an always-on agent is based on continuity and reliability. A system that stops unexpectedly does more than interrupt a conversation; it can leave a multi-step process in an uncertain state.
Businesses will want to know whether an interrupted agent can resume without repeating actions, whether it can identify which steps were completed and whether a human can reconstruct the sequence afterward. These are ordinary engineering questions, but persistent AI makes them visible to a new group of buyers. Reliability is no longer only a cloud-infrastructure metric; it becomes part of the agent’s apparent competence.
The strongest evidence for adoption will therefore come from production use rather than stage performance. Organizations should look at completion rates, correction rates, human intervention, latency and the cost of failed tasks. A system that saves fifteen minutes on nine tasks but creates an hour of cleanup on the tenth may be less productive than a more limited assistant whose boundaries are clearer.
Permissions become a product feature
Reuters reported that OpenAI is giving users controls intended to regulate what Dots can do and to prevent unauthorized changes to data. This is central to the enterprise case. Once an agent can act across applications, permission design becomes part of the user experience rather than something hidden in an administrator’s console.
The most practical model is likely to be graduated authority. An agent may be allowed to read project messages automatically, draft an update without approval, request confirmation before sending it to clients and be completely blocked from changing financial data. Such distinctions preserve much of the convenience of automation without giving every agent a universal ability to alter the systems it can see.
Companies also need temporary permissions. A project agent may legitimately need access to a data room for two weeks and no reason to keep that access afterward. If agent credentials accumulate indefinitely, an organization can recreate the same access-control problems that already exist with forgotten user accounts and service credentials, only at a larger scale.
Business data rules will influence adoption
According to Reuters, OpenAI said business data is not used for training without consent. That assurance addresses one of the most common barriers to enterprise AI: companies may be willing to use a model on internal information, but unwilling to have that information become part of a future general-purpose training process. Contractual terms and technical controls will still matter because different products and deployment arrangements can have different data flows.
Persistent agents add another layer to the question. A system that works over time may retain task history, cached context or summaries that allow it to continue a project. The organization needs to understand where that information is stored, how long it remains available and whether deleting an underlying document also removes the agent’s derived memory of it.
Data residency can matter as well. Multinational businesses operate under different legal and regulatory requirements, particularly for employee, customer and financial information. An agent platform that spans applications may touch several categories of data during one task. The enterprise product has to make those movements visible enough that customers can map them to their own compliance obligations.
The enterprise market is already large enough to matter
Reuters reported that OpenAI cited more than 35 million weekly users across Codex and ChatGPT Work and about 1.2 billion consumer ChatGPT users. Those figures are company-provided and describe different products and usage definitions, so they should not be treated as directly comparable. They nevertheless show why OpenAI is trying to turn widespread experimentation into a more durable workplace relationship.
A separate Reuters report on September 29 said OpenAI’s annualized recurring revenue was nearing $70 billion, according to a source familiar with the matter, with enterprise sales having doubled since July. Annualized revenue is a run-rate measure rather than audited full-year revenue, but it reinforces the scale of the commercial stakes behind the product announcements.
The enterprise opportunity is valuable because workplace software can produce recurring revenue and long customer relationships. It is also expensive to win. Large customers expect security reviews, integration support, service guarantees, administrative controls and predictable pricing. The difference between a popular consumer assistant and a trusted enterprise platform is therefore organizational as much as technical.
The cost equation changes when the agent keeps working
A chatbot interaction has a relatively visible unit of consumption: a user submits a request and receives a response. An always-on agent can consume computing resources while performing background work, checking new information or revisiting a plan. That makes pricing and budget controls more important, particularly when thousands of employees can launch agents at the same time.
Enterprises will need limits that correspond to business value. A sales-research agent may justify more computation when it supports a large contract than when it is preparing a low-priority internal note. A software agent may need substantial resources during a release and very little afterward. Flat user subscriptions, usage charges and task-based pricing each create different incentives for how freely employees delegate work.
The productivity calculation should also include supervision. If every autonomous task requires a lengthy review, the organization has moved work rather than removed it. The highest-value agents will probably be those operating in processes where results can be checked quickly and objectively, allowing humans to supervise outcomes instead of redoing the underlying task.
Persistent memory can be useful and dangerous in ordinary ways
Continuity often depends on memory. An agent following a project needs to know previous decisions, unresolved questions and the meaning of recurring names or files. Without that context, the user would have to rebuild the project state in every session, undermining much of the benefit of persistence.
Yet stored context can become stale. A project assumption that was correct last month may no longer be valid, while an old instruction may conflict with a new policy. Persistent agents therefore need ways to distinguish durable facts from temporary guidance and to show users which assumptions they are carrying forward. Otherwise, memory can make an error more consistent rather than making the system more intelligent.
Corporate records also have retention rules. Some information must be preserved; other information must be deleted after a defined period. Agent memory should fit inside those policies rather than becoming an informal archive outside them. That may require administrators to manage agent histories with the same seriousness as email, documents and database records.
The natural-language interface does not eliminate process design
One attraction of agents is that employees can describe goals in normal language instead of configuring complex automation. That lowers the barrier to creating workflows, but it does not remove ambiguity. “Keep the project on track” is not an operational definition. The agent still needs rules for deadlines, priorities, escalation and which source is authoritative when two systems disagree.
Companies that treat the agent as a substitute for defining a process may discover that it reproduces organizational confusion at machine speed. The strongest deployments will likely begin with workflows whose inputs, outputs and decision rights are already understood. Natural language can simplify interaction while the underlying governance remains explicit.
This is similar to earlier waves of business automation. Robotic process automation worked best where organizations first understood the process being automated. Agents can handle more variation, but they still benefit from clear definitions of success and failure. Intelligence reduces the need for rigid scripts; it does not eliminate the need for accountability.
Human intervention must be designed rather than assumed
Persistent agents are often described as systems that know when to ask for help. In practice, that threshold is a design decision. An agent that escalates every uncertainty becomes an expensive notification system. One that rarely escalates may continue too far after a key assumption has failed.
A useful escalation rule should consider consequence as well as confidence. An uncertain draft can be inexpensive to review later; an uncertain payment or external commitment should stop earlier. The user interface has to make the reason for escalation clear enough that a human can decide quickly, otherwise every handoff becomes a new investigation.
Organizations can also learn from intervention data. If humans repeatedly correct the same category of task, the workflow may not be ready for autonomous operation. If intervention falls as the agent gains better context and tooling, the organization can expand its authority gradually. Deployment becomes a controlled progression rather than a one-time decision to “turn on autonomy.”
Competition will move to measurement
As more companies release agents, product comparisons will become harder if every vendor uses its own demonstrations and benchmark definitions. Enterprise customers will need measures tied to their own work: how many support cases were resolved correctly, how many code changes passed review, how quickly research tasks were completed and how often employees had to reverse an agent’s action.
These operational metrics may matter more than general intelligence benchmarks. A model can score highly on reasoning tests and still be poorly suited to a specific corporate workflow because the integrations are weak or the cost is too high. Conversely, a less powerful model can deliver more value if it has access to the right structured information and operates inside well-defined boundaries.
Vendors will therefore compete on the surrounding system: connectors, permissions, memory, monitoring, administration and support. The language model remains the engine, but the enterprise product is the vehicle built around it. Dots is notable because OpenAI is explicitly moving further into that systems layer.
Software developers are the first proving ground
OpenAI’s inclusion of Codex in the Dots ecosystem reflects a wider industry pattern: software development is one of the earliest areas where agents can perform long sequences of semi-structured work. Code can be tested, compared and reviewed in ways that many business documents cannot. That makes the field a useful proving ground for persistent automation.
Even there, success depends on the environment. An agent that writes a feature has to understand the repository, testing requirements, deployment process and organization-specific conventions. A strong general coding model without those inputs can produce technically plausible work that does not fit the product. Context integration is therefore part of performance.
The lessons can transfer to other functions. Finance, legal, marketing and operations each have their own equivalent of a test suite: reconciliation checks, approval requirements, brand rules or service-level metrics. The more clearly those checks are expressed, the easier it becomes to delegate routine work without delegating final accountability.
The agent platform could reshape workplace interfaces
If persistent agents become reliable, employees may spend less time navigating menus and more time specifying outcomes. A project manager could ask for an updated risk summary rather than opening each underlying system. A salesperson could request a briefing assembled from approved sources. A developer could delegate a maintenance task and review the resulting change instead of performing each intermediate step manually.
That does not necessarily mean traditional applications disappear. They remain systems of record, sources of structured data and places where specialized work is performed. The change would be in how often users interact with them directly. The agent becomes a broker between human intent and software interfaces.
Such a shift would also change software distribution. If an agent chooses which application to call, vendors may compete to be the best tool for machine users as well as human users. Well-documented interfaces and clear permissions could become more important than visual design for some back-office functions.
Reliability will decide whether “always on” feels helpful or intrusive
Persistence can create value because the agent notices changes without being asked. The same characteristic can become intrusive if users receive constant updates, unnecessary questions or actions they did not expect. Product design therefore needs to control frequency as carefully as capability.
The best agent may be quiet most of the time. It should surface information when a decision is needed, when a deadline is at risk or when its own confidence falls below an agreed threshold. A constant stream of low-value activity undermines trust even if the underlying model is technically competent.
This makes notification design a business issue. Employees already manage email, chat and application alerts. Adding an autonomous assistant that generates its own stream of work can either reduce fragmentation or create another source of interruption. The distinction depends on whether the system consolidates decisions or simply produces more messages.
The economic test is whether the agent changes the unit of work
The long-term significance of Dots will depend on whether organizations begin buying outcomes rather than AI interactions. Today, many enterprise AI products are sold per user or per quantity of model usage. Persistent agents create the possibility of measuring completed tasks, maintained processes or resolved exceptions instead.
That could alter staffing and software budgets in ways that are difficult to infer from adoption statistics alone. A company may have more AI users while spending less human time on coordination, or it may add expensive agent subscriptions without changing the amount of work employees perform. Productivity has to be measured at the process level.
Managers should therefore resist interpreting the presence of agents as evidence of transformation. The useful question is what changed: cycle time, error rate, capacity, customer response or cost. If those measures do not improve, persistent automation has not yet justified its complexity.
DevDay marks a move from assistant to infrastructure
OpenAI also used DevDay to highlight a broader product portfolio, and Reuters reported new models and subscription offerings alongside Dots. The pattern suggests a company trying to serve consumers, developers and large organizations through a common AI layer while differentiating the level of capability and service around it.
For enterprise customers, the most important consequence is that AI is becoming infrastructure that must be administered. Persistent agents need identities, permissions, logs, budgets and lifecycle management. Those requirements may sound less dramatic than autonomous reasoning, but they determine whether the technology can survive contact with real corporate governance.
That is also why the next phase will not be decided by one keynote. Dots will have to prove that persistent agents can stay useful through changes in data, software and human priorities. Competitors will be judged by the same standard. The breakthrough is not an agent that can continue working after the user closes the window; it is an agent whose work remains predictable enough that the organization is comfortable letting it continue.
The next contest is trust at operating scale
The enterprise AI market has moved quickly from drafting emails to generating code and now toward continuous task execution. Each step removes a small amount of direct human control in exchange for greater potential productivity. The trade becomes acceptable only when organizations gain compensating visibility into what the system is doing.
OpenAI’s Dots announcement is therefore best understood as a test of operational trust. The product can be technically impressive and still fail commercially if companies cannot govern it, price it or measure its contribution. Conversely, an agent that appears less dramatic but reliably handles ordinary coordination could become deeply embedded in daily work.
The technology industry has spent several years asking whether generative AI can answer questions, write documents and produce software. Persistent agents move the question one step further: can AI hold responsibility for a continuing piece of work without losing the boundaries set by its users? The answer will emerge from deployment records, not slogans. If Dots and its rivals succeed, the defining workplace interface of the next cycle may be not a chatbox that responds, but a supervised digital worker that remains on the job.
Procurement will move from model choice to control architecture
The arrival of persistent agents also changes the questions asked during technology procurement. A buyer evaluating a conventional software product can focus on features, price, support and integration. An agent platform adds a second set of questions about authority: which actions can be delegated, how those rights are granted, what happens when a task crosses an organizational boundary and which evidence remains available after the agent finishes.
This favors vendors that can explain their control architecture in concrete terms. Customers will want administrative policies that apply across many agents, not a collection of individual settings configured by each employee. They will also want to distinguish read-only access from the ability to create, update or delete records. A broad promise that “the user stays in control” is less useful than a clear description of which actions require confirmation and which can occur automatically.
Procurement teams may also demand portability. If an organization builds important workflows around one agent platform, switching costs can rise quickly because permissions, memories and process logic become embedded in the system. Open standards for connectors and exportable audit records would reduce that dependence. The commercial battle will therefore involve not just model quality but how much of a company’s operating structure becomes tied to a vendor’s agent layer.
Agents will expose the quality of corporate data
Generative AI is often described as a way to make fragmented information easier to use, but an agent cannot fully solve contradictions inside the underlying records. If a customer has two different addresses in two systems, or a project has several competing deadline documents, the agent still needs a rule for choosing. Automation can make inconsistencies visible, but it cannot decide which source should be authoritative unless the organization defines that hierarchy.
That creates a less glamorous requirement for successful deployment: data governance. Companies that have maintained clear ownership, current records and consistent identifiers will find it easier to let agents operate across applications. Companies with duplicated files and ambiguous process ownership may discover that the AI spends much of its time asking questions that employees previously resolved informally.
The problem can become a benefit if deployments are staged carefully. A pilot may reveal that a workflow depends on information nobody officially owns, or that employees routinely bypass an outdated system. Those findings are valuable even if the first agent is not ready for full autonomy. In that sense, persistent AI can act as a diagnostic for organizational design as much as a tool for automating it.
The labor question is more granular than replacement
Always-on agents will inevitably revive debate about employment, but the immediate workplace effect is likely to be uneven. A job is usually a bundle of activities rather than one task. An agent may take over monitoring, drafting and coordination while leaving negotiation, judgment and accountability with the employee. The result can be a redesigned role rather than the disappearance of the role itself.
Whether that change improves work depends on how the saved time is used. Removing repetitive administration can give employees more capacity for customers, analysis or creative work. It can also lead employers to increase the number of cases each person is expected to supervise. Productivity gains and workload reductions are not the same outcome, and organizations should measure both when they assess the human effect of automation.
Training will matter because supervising an agent is a skill. Employees need to know how to formulate objectives, interpret uncertainty and recognize when a system is operating outside its useful range. That is different from simply learning a new interface. The more responsibility an agent receives, the more important it becomes that the human operator understands the process well enough to challenge the result rather than accepting it because it arrived quickly.
A mature market will distinguish autonomy levels
The term “agent” currently covers a wide range of products, from assistants that call one tool after explicit approval to systems that can pursue a broad objective for hours. That makes comparisons difficult and encourages marketing language to outrun operational reality. Enterprise buyers would benefit from describing autonomy in terms of observable permissions and time horizons rather than relying on the label alone.
A low-autonomy agent might draft actions but require confirmation for every change. A medium-autonomy system might perform routine work inside predefined limits and stop at exceptions. A higher-autonomy system could coordinate several applications over a longer period with only periodic review. None of these levels is inherently better. The appropriate choice depends on the consequence of error and the maturity of the surrounding process.
Clear autonomy levels would also make performance data easier to interpret. An agent that completes ninety percent of a task while asking for frequent approval is not directly comparable with one that completes eighty percent without supervision. Both may be useful in different environments. As the market develops, customers are likely to demand evidence matched to the level of authority they are actually buying, which would make the category more disciplined than today’s broad claims of digital coworkers and autonomous assistants.
