Will Claude Cowork still be my company's work tool in 2027 or 2028?

Nothing guarantees it, and this is not a question of product quality. Three layers Cowork rests on keep moving: the software harness (daily builds, no version-pinning mechanism), the underlying model (barely more than four months separate Sonnet 4.6 becoming the default model on 17 February 2026 from Sonnet 5 arriving on 30 June), and the MCP connectors, managed by third parties and able to change the tools they expose mid-session. A dedicated article on this site details this instability layer by layer. Cowork still remains a legitimate transitional tool: it lets a company discover what an agent can do, not carry a critical process in production over time.

What is generative UI, and how does it replace chat?

An interface generated on the fly for a specific intent, rather than picked from a fixed catalogue of applications or typed into a dialogue box. Google laid out this vision as early as November 2025 on its research blog, with an example that became well known internally: explaining the microbiome to a five-year-old requires different content and a different interface than explaining it to an adult. In February 2026, Google formalised the principle under the name Natively Adaptive Interfaces, an architecture built on multimodal models able to process voice, text and image in a single context, replacing a static navigation tree with agent-driven modules; at Google I/O in May 2026, this same generative-UI logic was built natively into Search's AI mode.

Do open protocols for building these generative interfaces already exist?

Yes, and adoption goes beyond a single vendor. AG-UI (Agent-User Interaction) standardises the real-time channel between an agent and a client-side application (messages, tool calls, state updates); it is adopted by Google, LangChain, AWS, Microsoft, Mastra and PydanticAI. The associated framework, the Generative UI Spectrum published by CopilotKit, distinguishes four levels of control: a fully pre-built interface where the agent only fills in already-validated components (Controlled), a vocabulary of composable blocks the agent assembles (Declarative), third-party surfaces embedded via MCP Apps, and an interface the agent generates entirely on its own (Open-Ended). The more freedom left to the agent, the more the risk of visual drift and security flaws grows with it.

Is the token cost of MCP connectors really an unresolved problem?

Yes, and Anthropic documents it itself. In a November 2025 engineering post, the company acknowledges that an agent connected to several MCP servers can burn through up to 150,000 tokens of tool definitions before even processing a request, a cost its own workaround (code execution instead of direct tool calls) brings down to 2,000 tokens in its example, a 98.7% saving. An independent analysis also counts more than a threefold increase in the number of MCP servers in six months, between August 2025 and February 2026, which multiplies the problem rather than solving it. The most recent academic literature still describes token allocation in agentic systems as an open problem. Giving an agent accurate, reliable, token-cheap data aggregates remains a research question, not a settled implementation detail.

So what is a tool like Cowork actually for today?

Buying a company a few months while its information system modernises. It is a useful buffer, not an end state: it enables exploration, framing and upskilling teams, while the business software itself gradually absorbs AI into where the work actually happens. Programming, in the sense of capturing a process, preparing it and tooling it, keeps its full value inside this window: it builds the durable architecture that will take over. A Deloitte survey of 501 executives published in August 2026 gives a sense of how long this window is likely to last: only 31% expect most of their business processes to be redesigned around agentic AI within two years, and only 5% describe themselves as highly prepared today.

AI Agents

Conversational AI cannot become a work interface

27 August 2026 · 15 min read · Paul-Antoine Tual

AI agents user interfaces Claude Cowork generative UI AG-UI Model Context Protocol human in the loop Service-as-Software Junyr Method

A question asked in training, never really disproved

In mid-2026, during an agentic AI training session, Paul-Antoine Tual asks an executive committee that has just rolled out Claude Cowork to its managers a question: “Do you really think this tool will still be yours in 2027? In 2028?” No one answers yes. Yet this is not a question about product quality. Cowork works, and its users say so. The problem lies elsewhere: the ground a company builds its automations on is still moving, and it is moving on three layers at once.

An article published on this site the week before details this instability layer by layer: a software harness that changes through daily builds, an underlying model subject to a retirement policy with a minimum 60-day notice, and MCP connectors managed by third parties, able to change the tools they expose mid-session [1]. Two facts, published since, add substance to this instability rather than repeating it. The first concerns the model: barely more than four months separate Sonnet 4.6 becoming the default model on claude.ai and Cowork, on 17 February 2026, from Sonnet 5 arriving on 30 June [2][3]. Anthropic itself describes Sonnet 5 as a model that “finishes complex tasks where previous Sonnet models would stop short” [3], and Opus 5 follows three weeks later, on 24 July [4]. A prompt calibrated for one is no longer really calibrated for the next: you are not changing a version, you are changing a colleague.

The second fact concerns connectors. Anthropic itself documents, in a November 2025 engineering post, that an agent connected to several MCP servers can burn through up to 150,000 tokens of tool definitions before even reading the first request, a cost its own workaround (code execution instead of direct tool calls) brings down to 2,000 tokens in its own example, a 98.7% saving [5]. That the protocol’s own vendor documents this problem, and that an independent industry analysis counts more than a threefold increase in the number of MCP servers in six months between August 2025 and February 2026 [6], confirms a simple point: how to give an agent accurate, reliable, token-cheap data aggregates is not a settled question. It is an open construction site, one the industry itself describes as a growing pain still to be solved [7], documented as far as the most recent academic literature as an unresolved problem of token allocation in agentic systems [8]. This is not an implementation detail.

Three unstable layers stacked on top of each other do not make an automation foundation. They make a permanent prototype.

A transitional tool, not an automation tool

These tools are excellent for transformation, and Paul-Antoine Tual uses them heavily in training: synthesis, analysis, exploration and framing, upskilling teams. They are not, today, automation tools. You do not hand a stable process, let alone an industrialisation budget, to a stack that rewrites itself every six months.

The business model wrapping these tools is moving too, which confirms this is still a transition phase rather than a settled regime. Classic SaaS charges for access to a fixed interface, per user seat, regardless of the work actually done. Two concrete examples show the shift toward paying for the outcome, what trade press calls Service-as-Software. Salesforce launched an Agentforce offering as early as October 2024 charging 2 dollars per conversation an agent resolves, a model since supplemented by other formulas but never withdrawn [9]. Zendesk charges its automated resolutions between 1.50 dollars on committed volume and 2 dollars pay-as-you-go, a rate the vendor does not publish on its official price list but confirms contract by contract [10]. Paying per seat assumes a stable product whose usage is known in advance; paying for the outcome assumes, on the contrary, that you do not yet know how much work the agent will actually get through. It is another symptom of the same diagnosis: neither the tool, nor its interface, nor how it is billed, is pinned down yet.

The question a company should be asking is therefore not “which tool do I buy”, but “how will I automate, and what will I build on once it is stable”. The Junyr Method answers the first half of this question by distinguishing an organisation’s levels of AI maturity [11]. It does not, by itself, answer the second half, which touches a layer even further upstream than the harness, the model or MCP: the interface itself.

Chat is not a work interface

A second question, more uncomfortable than the first: nothing proves that a company’s employees will still be working conversationally in one year, two years, three years.

Work does not happen by chatting. Work does not happen over coffee. Chat is the lowest common denominator among interfaces: it speaks to many jobs at once, and so never really speaks to any of them. Arnab Bose, Asana’s chief product officer, puts it bluntly in an op-ed published in late July 2026: “Chat is a useful interface, but a poor operating model.” He backs it with a figure: 73% of surveyed IT decision-makers attribute failed or stalled AI initiatives to missing organisational context, exactly what a generic dialogue box fails to capture [12].

Today Sales Production Finance Human resources Support

Chat one, generic interface

Tomorrow Sales Voice, on the road

Production Image, quality control Finance Diagram, modelling Human resources Dashboard, steering Support Direct manipulation, technical role A chat has to speak to five jobs at once, so it never really speaks to any of them. An interface fitted to the job's own gesture replaces free text with each role's natural mode of expression.

Figure 1. Chat as the lowest common denominator, against generative interfaces fitted to the job. Source: notes [12], [13], [14], [16].

What work actually requires are more intuitive, faster interfaces, fitted to the job’s own gesture: voice for a salesperson on the road, an image for quality control on a production line, a diagram to model a financial scenario, direct manipulation of an object for a technical role. One interface per job, not one dialogue box for everyone.

The biggest vendors document this themselves, without waiting for academic consensus. Google published its vision of generative UI as early as November 2025, an interface generated on the fly for each intent rather than picked from a fixed catalogue of applications, with an example that became well known internally: explaining the microbiome to a five-year-old requires different content and a different interface than explaining it to an adult [13]. Three months later, in February 2026, Google formalised this principle under the name Natively Adaptive Interfaces, an architecture built on multimodal models able to process voice, text and image in a single context, replacing a static navigation tree with agent-driven modules [14]. At Google I/O in May 2026, this same generative-UI logic was built natively into Search’s AI mode [15]. Anthropic follows a different but converging path: since 7 May 2026, Claude for Excel, Word and PowerPoint has been generally available, and Claude for Outlook in public beta, not as a chat pasted into a spreadsheet cell, but as an embedded module carrying its context from one application to the next [16]. In both cases, the answer to the chat problem is not a better chat. It is chat’s disappearance as the default work surface.

This shift is not just product rhetoric. An open protocol dedicated to this kind of interface already exists and carries broad industry backing: AG-UI (Agent-User Interaction), which standardises the real-time channel between an agent and a client-side application (messages, tool calls, state updates), is adopted by Google, LangChain, AWS, Microsoft, Mastra and PydanticAI [17]. It carries an explicit hierarchy of control levels, from strictest to freest: a fully pre-built interface where the agent only fills in already-validated components (Controlled), a vocabulary of composable blocks the agent assembles (Declarative), third-party surfaces embedded via MCP Apps, up to an interface the agent generates entirely on its own (Open-Ended) [18]. The more freedom left to the agent, the more the risk of visual drift and security flaws grows with it: unlike chat, which offers total freedom with no guarantee at all, these protocols make that trade-off explicit rather than something you absorb after the fact.

Chat speaks to every job, and so to none.

What it is for, right now: a buffer, not an end state

Conversational AI remains a useful buffer, and that is not nothing. It buys a company a few months while its information system modernises, while its business software gradually absorbs AI into itself, where the work actually happens, rather than next to it, in a dialogue window.

Figure 2. The transition window separates discovery through chat from durable integration into business software. Source: note [19].

This is exactly why programming, in the sense of capturing a process, preparing it and tooling it, remains highly valuable within this window: it prepares what comes after, while chat serves as a testing ground. The Deloitte survey published on 12 August 2026, of 501 executives, gives a sense of how long this window is likely to last: only 31% of respondents expect most of their business processes to be redesigned around agentic AI within two years, and only 5% describe themselves as highly prepared today [19]. The transition is not a flip of a switch. It is a ramp.

And it must be said to the end: human in the loop remains extremely present

Believing that automation can happen without it within a few months is wishful thinking. An article on this site already sets out the principle that an agent’s guardrail lives in code, not in the prompt [20]. The generative-UI protocols themselves are starting to standardise this principle at the level of the visual component: a preview card before an irreversible action, an explicit accept-or-reject toggle, an audit trail to roll back to a prior state [18]. The same Deloitte survey confirms it from the executives’ own side: 61% expect most of their agents to remain generally autonomous but under human oversight, and 75% judge that collaboration between humans and agents creates more value than automation alone [19]. This is not a temporary resistance to change. It is a condition of the system working at all.

The ground will keep moving, on the three layers of harness, model and MCP, on the interface itself, and on how it gets billed. The stakes are not finding the tool that will stop moving. They are knowing what to build while it does.

Sources

[1] Paul-Antoine Tual, “Agentic design: why Cowork does not build durable agents”, paulantoinetual.fr, 27 August 2026. /en/blog/cowork-agents-ia-durables [2] Anthropic, “Introducing Claude Sonnet 4.6”, 17 February 2026: default model for Free and Pro plans on claude.ai and Cowork. https://www.anthropic.com/news/claude-sonnet-4-6 [3] Anthropic, “Claude Sonnet 5”, 30 June 2026. https://www.anthropic.com/news/claude-sonnet-5 [4] Anthropic, “Claude Opus 5”, 24 July 2026. https://www.anthropic.com/news/claude-opus-5 [5] Anthropic Engineering, “Code execution with MCP: Building more efficient agents”, 4 November 2025: example reducing token usage from 150,000 to 2,000, a 98.7% saving. https://www.anthropic.com/engineering/code-execution-with-mcp [6] Bloomberry, “I analyzed 1,400 MCP servers, here’s what I learned”, data as of late February 2026: 1,412 servers counted versus 425 at the end of August 2025, a 232% increase in six months. https://bloomberry.com/blog/we-analyzed-1400-mcp-servers-heres-what-we-learned/ [7] The New Stack, “MCP’s biggest growing pains for production use will soon be solved”, 2026. https://thenewstack.io/model-context-protocol-roadmap-2026/ [8] Ou Wu (Hangzhou Institute for Advanced Study), Yingjun Deng (Hefei Institutes of Physical Science, CAS), “Computational Challenges in Token Economics: Bridging Economic Theory and AI System Design”, arXiv:2605.17410, 17 May 2026: token allocation in agentic systems presented as an open credit-assignment problem. https://arxiv.org/html/2605.17410v1 [9] SaaStr, “Salesforce Now Has 3 Pricing Models for Agentforce, and Maybe Right Now That’s the Way to Do It”, 2026: October 2024 launch of the 2-dollar-per-resolved-conversation charge, never withdrawn since despite additional pricing formulas being added. https://www.saastr.com/salesforce-now-has-3-pricing-models-for-agentforce-and-maybe-right-now-thats-the-way-to-do-it [10] eesel AI, “Understanding Zendesk AI pricing: a complete pay-per-resolution guide (2026)”: 1.50 dollars per automated resolution on committed volume, 2 dollars pay-as-you-go; rate confirmed contract by contract, absent from the official public price list. https://www.eesel.ai/blog/understanding-zendesk-ai-pricing-a-complete-pay-per-resolution-guide [11] Paul-Antoine Tual, “The Junyr Method ladder: five levels of AI maturity in SMEs”, paulantoinetual.fr. /en/blog/echelle-methode-junyr-cinq-niveaux-maturite-ia-pme [12] Arnab Bose (Asana), “Why Enterprise AI Needs More Than Chat: The New Business Model Of Agentic Work”, Forbes Technology Council, 30 July 2026. https://www.forbes.com/councils/forbestechcouncil/2026/07/30/why-enterprise-ai-needs-more-than-chat-the-new-business-model-of-agentic-work/ [13] Google Research, “Generative UI: A rich, custom, visual interactive user experience for any prompt”, 18 November 2025. https://research.google/blog/generative-ui-a-rich-custom-visual-interactive-user-experience-for-any-prompt/ [14] MarkTechPost, “Google AI Introduces Natively Adaptive Interfaces (NAI): An Agentic Multimodal Accessibility Framework Built on Gemini for Adaptive UI Design”, 10 February 2026. https://www.marktechpost.com/2026/02/10/google-ai-introduces-natively-adaptive-interfaces-nai-an-agentic-multimodal-accessibility-framework-built-on-gemini-for-adaptive-ui-design/ [15] Forbes, Janakiram MSV, “Google I/O 2026 Turned Gemini Into An Agent Platform”, 21 May 2026. https://www.forbes.com/sites/janakirammsv/2026/05/21/google-io-2026-turned-gemini-into-an-agent-platform/ [16] The New Stack, “Claude can now follow users across Outlook, Word, Excel, and PowerPoint”, May 2026: general availability for Excel, Word and PowerPoint, and public beta for Outlook, since 7 May 2026. https://thenewstack.io/claude-word-excel-powerpoint-outlook-microsoft-office/ [17] AG-UI Protocol, official documentation and ag-ui-protocol/ag-ui repository: open, event-based protocol standardising the real-time channel between an agent and a client-side application, adopted by Google, LangChain, AWS, Microsoft, Mastra and PydanticAI. https://docs.ag-ui.com/introduction [18] CopilotKit, “The Generative UI Spectrum”: four levels of control, Controlled Generative UI, Declarative Generative UI, MCP Apps, Open-Ended Generative UI, with AG-UI acting as the foundational protocol underneath these four patterns. https://www.copilotkit.ai/generative-ui-spectrum [19] Deloitte, “AI Agents are Only the Beginning: Deloitte Survey Examines the AI Readiness Gap…”, press release of 12 August 2026: 501 respondents from senior manager to executive committee level, fielded April-June 2026; 31% expect most of their processes redesigned within two years, 5% describe themselves as highly prepared, 61% expect generally autonomous agents under human oversight, 75% judge human-agent collaboration more value-creating than automation alone. https://www.prnewswire.com/news-releases/ai-agents-are-only-the-beginning-deloitte-survey-examines-the-ai-readiness-gap-and-reveals-how-enterprises-can-prepare-for-agentic-success-302848848.html [20] Paul-Antoine Tual, “Engineering agentic systems: golden rules, architecture and security of the Human-in-the-Loop”, paulantoinetual.fr, 4 July 2026. /en/blog/ingenierie-systemes-agentiques-human-in-the-loop

Frequently asked questions

Will Claude Cowork still be my company's work tool in 2027 or 2028?: Nothing guarantees it, and this is not a question of product quality. Three layers Cowork rests on keep moving: the software harness (daily builds, no version-pinning mechanism), the underlying model (barely more than four months separate Sonnet 4.6 becoming the default model on 17 February 2026 from Sonnet 5 arriving on 30 June), and the MCP connectors, managed by third parties and able to change the tools they expose mid-session. A dedicated article on this site details this instability layer by layer. Cowork still remains a legitimate transitional tool: it lets a company discover what an agent can do, not carry a critical process in production over time.
Is the software's business model changing too, or just the interface?: Both are moving together, which confirms this is a transition phase rather than a settled regime. Classic SaaS charges for access to a fixed interface, per user seat, regardless of the work actually done. Salesforce launched an Agentforce offering in October 2024 charging 2 dollars per conversation an agent resolves, never withdrawn since. Zendesk charges its automated resolutions between 1.50 dollars on committed volume and 2 dollars pay-as-you-go, a rate the vendor confirms contract by contract without publishing it on its official price list. Trade press calls this shift from paying per seat to paying for the outcome Service-as-Software: one more sign that neither the tool, nor its interface, nor how it is billed, is pinned down yet.
Why isn't chat a good everyday work interface?: Because a generic dialogue box has to speak to every job at once, and so never really speaks to any of them. Arnab Bose, Asana's chief product officer, puts it bluntly in a July 2026 op-ed: "Chat is a useful interface, but a poor operating model." His supporting data point: 73% of surveyed IT decision-makers attribute failed or stalled AI initiatives to missing organisational context, exactly what a free-text window fails to capture. Real work happens through different gestures depending on the job (speaking on a sales visit, looking at a part on a production line, sketching a diagram to model a financial scenario), and a single interface flattens all of these gestures to the lowest common denominator.
What is generative UI, and how does it replace chat?: An interface generated on the fly for a specific intent, rather than picked from a fixed catalogue of applications or typed into a dialogue box. Google laid out this vision as early as November 2025 on its research blog, with an example that became well known internally: explaining the microbiome to a five-year-old requires different content and a different interface than explaining it to an adult. In February 2026, Google formalised the principle under the name Natively Adaptive Interfaces, an architecture built on multimodal models able to process voice, text and image in a single context, replacing a static navigation tree with agent-driven modules; at Google I/O in May 2026, this same generative-UI logic was built natively into Search's AI mode.
Do open protocols for building these generative interfaces already exist?: Yes, and adoption goes beyond a single vendor. AG-UI (Agent-User Interaction) standardises the real-time channel between an agent and a client-side application (messages, tool calls, state updates); it is adopted by Google, LangChain, AWS, Microsoft, Mastra and PydanticAI. The associated framework, the Generative UI Spectrum published by CopilotKit, distinguishes four levels of control: a fully pre-built interface where the agent only fills in already-validated components (Controlled), a vocabulary of composable blocks the agent assembles (Declarative), third-party surfaces embedded via MCP Apps, and an interface the agent generates entirely on its own (Open-Ended). The more freedom left to the agent, the more the risk of visual drift and security flaws grows with it.
Is the token cost of MCP connectors really an unresolved problem?: Yes, and Anthropic documents it itself. In a November 2025 engineering post, the company acknowledges that an agent connected to several MCP servers can burn through up to 150,000 tokens of tool definitions before even processing a request, a cost its own workaround (code execution instead of direct tool calls) brings down to 2,000 tokens in its example, a 98.7% saving. An independent analysis also counts more than a threefold increase in the number of MCP servers in six months, between August 2025 and February 2026, which multiplies the problem rather than solving it. The most recent academic literature still describes token allocation in agentic systems as an open problem. Giving an agent accurate, reliable, token-cheap data aggregates remains a research question, not a settled implementation detail.
So what is a tool like Cowork actually for today?: Buying a company a few months while its information system modernises. It is a useful buffer, not an end state: it enables exploration, framing and upskilling teams, while the business software itself gradually absorbs AI into where the work actually happens. Programming, in the sense of capturing a process, preparing it and tooling it, keeps its full value inside this window: it builds the durable architecture that will take over. A Deloitte survey of 501 executives published in August 2026 gives a sense of how long this window is likely to last: only 31% expect most of their business processes to be redesigned around agentic AI within two years, and only 5% describe themselves as highly prepared today.
Will human in the loop disappear as agentic automation spreads?: No, and nothing in the available data points that way. The same Deloitte survey from August 2026 shows that 61% of respondents expect most of their agents to remain generally autonomous but under human oversight, and 75% believe human-agent collaboration creates more value than automation alone. The generative-UI protocols themselves are starting to standardise this principle at the level of the visual component: a preview card before an irreversible action, an explicit accept-or-reject toggle, an audit trail to roll back to a prior state. A dedicated article on this site already sets out the principle that an agent's guardrail lives in code, not in the prompt: the same logic applies here. Believing that automation without human supervision is a matter of months is wishful thinking, not a realistic roadmap.

Paul-Antoine Tual

AI Transformation Leader · Junyr Method™ · Transition manager specialising in AI for French SMEs and mid-caps. Engineer from the École des Mines de Nantes, lawyer, developer since 1993.

Free AI Maturity Audit →

← All articles