ChatGPT Work, Claude Cowork, and the Rest: Why an AI App Is the Wrong Question for Your Business

Within 48 hours, two of the world's most important AI providers made nearly identical announcements. On July 7, 2026, Anthropic expanded "Claude Cowork" from a desktop app to mobile and web. On July 9, 2026, OpenAI followed suit with "ChatGPT Work": an agent inside ChatGPT that pulls context from connected apps, breaks a goal down into sub-steps, and independently delivers finished documents, spreadsheets, presentations, or even web apps – powered by GPT-5.6 in its Sol, Luna, and Terra variants. Both companies are visibly pursuing the same strategy: chat, coding, and autonomous agents converging under one subscription, in one interface, for one business. The office super-app has arrived.

For many owners and IT leads at German SMBs, that sounds tempting. One solution, one provider, one price – finally an end to the patchwork of Copilot licenses, ChatGPT accounts in the marketing department, and a specialized platform on the shop floor. But that temptation is exactly where we should pause. Because the question "Which AI app should our business adopt?" is the wrong question. The right question is: what problem do we actually want to solve – and with which tool can we solve it most affordably and reliably? This article breaks down what ChatGPT Work and Claude Cowork can actually do, what the current research says about the benefits of such suites, and why the super-app bet is often the more expensive and riskier option for a 30-to-80-person business.

Two Days, Two Super-Apps: What Happened on July 7 and 9, 2026

Claude Cowork started out as a desktop-only feature for Anthropic's most expensive plan, Claude Max ($100/month). With the rollout to mobile and web, Anthropic removed an important technical hurdle: Cowork now keeps working in the background even when the device is offline, pinging the user via mobile notification whenever a decision is needed. Chat and Cowork now share a common interface – one tab, one sidebar, one shared projects and artifacts area. What was once a coding tool has now definitively become a general-purpose office agent.

OpenAI followed two days later with the launch of ChatGPT Work: an agent that works autonomously on a task for hours, pulls context from connected apps, and delivers a finished result at the end – a document, a spreadsheet, a presentation, or a small web app. The rollout initially targets Pro, Enterprise, and Edu users on web, mobile, and desktop; Plus and Business customers are to follow later. Notably, OpenAI had already merged Codex and ChatGPT into one app before this. ChatGPT Work is therefore not a one-off move, but the next step in a longer bundling strategy.

Both moves follow the same logic: whoever bundles chat, a coding agent, and an autonomous office agent into one subscription binds customers more tightly to their own platform and makes switching to a competitor harder. For the providers, that makes economic sense. For an SMB footing the bill at the end, though, a second look is worthwhile.

The Bundling Logic: Attractive for the Provider, Not Automatically for You

Selling a super-app is strategically sound for OpenAI and Anthropic: a single subscription per user raises average revenue per customer and makes a later switch more expensive, because workflows, prompts, projects, and file integrations are deeply embedded in the respective ecosystem. This lock-in risk has already become visible – and not at some small SMB, but at one of the largest software companies in the world. In May 2026, Microsoft cut Claude Code licenses for its Experiences and Devices division, effective June 30, 2026. Officially, this was justified as "toolchain unification," but according to reports, the real reason lay in sharply rising token costs from agentic workflows. If even Microsoft has to reverse a tool decision it already made because ongoing costs spiraled out of control, that's a clear warning sign for any business now making a similar bundling decision.

On top of that: both providers' pricing structures are built for heavy usage, not for the average needs of a mid-sized team. ChatGPT Business costs $20 per user per month on the annual plan ($25/month, minimum two seats) – that sounds manageable. But according to the rollout plan, the full feature set of ChatGPT Work is initially available only to Pro, Enterprise, and Edu users. And ChatGPT Enterprise isn't publicly priced but negotiated through sales; market observers put 2026 pricing at roughly $45 to $75 per user per month, often around $60, with a minimum of about 150 seats and annual prepayment. That works out to a cost floor of roughly $108,000 per year – for a plan that's often booked only because the desired feature (the autonomous agent) is tied to the most expensive tier.

The principle is similar with Claude: Cowork is included in every paid plan at no extra charge, but the difference lies in usage volume. Anyone wanting to genuinely use Cowork intensively as a background agent quickly ends up on Claude Max at $100 or even $200 per user per month – not on the entry-level Pro plan at $20. Claude Team Premium costs $125 per seat per month. Book the full suite for a ten-person team, and you're quickly paying a five-figure amount per year – for features that, according to Anthropic's own data, most users only tap into a fraction of.

What the Usage Data Actually Shows: It's Barely About Coding

One detail from Anthropic's own analysis is particularly telling. An analysis of 1.2 million Cowork sessions shows: over 90 percent of usage isn't coding at all, but classic business operations and content creation. The largest category, at 33.4 percent, is pure process work – pulling together scattered updates into a report, reconciling spreadsheets, compiling status updates. In second place, at 16.4 percent, comes content creation and copywriting.

That's remarkable, because Claude Cowork was originally positioned as an offshoot of coding-agent technology. In practice, most businesses use the super-app for generic office tasks – not for the highly specialized autonomous workflows that both providers' marketing materials tend to foreground. This lines up with an observation that, at its core, isn't surprising: most businesses book the full suite but only use a small slice of its feature range – say, only reporting or only content, never the full breadth of autonomous-agent capabilities the expensive tier is actually paying for.

This observation isn't unique to the AI industry. It follows a pattern that the Nexthink "Soft-WASTE" study described for enterprise software in general: roughly 50 percent of all software applications installed in businesses go entirely unused by employees, and 95 percent of IT leaders have no visibility into whether the tools they've rolled out are actually being used. That figure refers to software in general, not specifically to AI suites – but as an analogy for the risk of blindly booked licenses, it's highly relevant.

Only 21 percent of businesses currently have a mature governance model for autonomous agents – even though 74 percent plan to deploy agentic AI to at least a moderate degree within the next two years. This gap between adoption speed and control is the real story behind the super-app hype.

These figures come from Deloitte's "State of AI in the Enterprise 2026" report, a survey of 3,235 executives across 24 countries (fieldwork August/September 2025, published in 2026). They point to a structural problem: the speed at which businesses want to adopt autonomous AI agents outpaces the speed at which they learn to control them. For an SMB with a typically thin IT department, this gap weighs especially heavily – precisely because autonomous agents like ChatGPT Work or Claude Cowork independently access files, email, and calendars. Without clear internal rules on who is allowed to approve which decisions and how autonomous actions are documented, this isn't a theoretical risk but a very concrete matter of data protection and compliance – one that's becoming increasingly important given the regulatory requirements around AI in business.

The ROI Question: Why the Tool Is Rarely the Real Problem

The imbalance becomes even clearer when you look at return on investment. The MIT-NANDA study "The GenAI Divide: State of AI in Business 2025" – based on 150 interviews, 350 surveyed employees, and 300 analyzed deployments, published in July 2025 – reaches a stark conclusion: 95 percent of generative-AI pilot projects in businesses deliver no measurable ROI. According to MIT, the main cause is explicitly not the quality of the underlying model, but a lack of workflow integration, immature data, and unclear goal-setting before the project even starts. Notably, more than half of budgets flow into sales and marketing tools, even though the greatest ROI demonstrably lies in the back office.

Applied to the current super-app debate, that means: whether a business adopts ChatGPT Work or Claude Cowork is secondary as long as the real homework is missing – a clearly defined process with clean data and a measurable goal. A generic suite with a broad feature set doesn't automatically solve that underlying problem. On the contrary: the more features a tool offers, the easier it becomes to keep postponing the actual goal-setting, because "something useful" can almost always be found in a super-app – just rarely the thing that would move the needle most.

The German numbers confirm this picture. According to the Bitkom AI Study 2026 (as of March 2026), 41 percent of German companies already use AI productively; among SMBs, the figure is around 38 percent. At the same time, 33 percent of companies report higher AI costs than originally expected. Adoption and cost control are drifting apart here too.

The SMB Reality: Shadow AI, Duplicate Licenses, and No Cost Visibility

A typical scenario at German SMBs, according to industry analyses, looks like this: three to five parallel AI tools per business – say, GitHub Copilot in development, ChatGPT Enterprise in administration, plus one or more specialized platforms in individual departments. On top of that come shadow accounts that individual departments set up without IT approval, and the absence of any central cost visibility per task. Redundant, duplicate licenses often run unnoticed in the five-figure euro range per year.

It's exactly into this gap that shadow AI pushes. A Bitkom survey from October 2025 of 604 companies puts unapproved AI tool usage at 42 percent; a 2026 Mitel study puts it as high as around 50 percent of employees. Gartner forecasts that by 2030, more than 40 percent of businesses will experience a security or compliance incident due to unauthorized shadow AI. These figures aren't a fringe phenomenon – they're a strong indicator that a centrally mandated super-app doesn't solve the underlying demand problem. If employees are already seeking out their own point solutions today because the official corporate suite doesn't cover their specific workflow, then booking yet another blanket suite typically just adds extra cost alongside the shadow usage that keeps running anyway – not an end to it.

It's also telling that even large providers are now recognizing that SMBs don't need a monolithic suite, but tailored building blocks. Through its new "Accenture Edge" division, Accenture, together with Google Cloud, unveiled an agentic-AI suite on July 7, 2026, specifically for mid-market companies with $300 million to $3 billion in annual revenue – with six clearly delineated solution areas ranging from customer intelligence to cybersecurity to agentic business operations. For the typical German SMB, which is often well below this revenue threshold, it remains an open question whether this offering is even reachable or affordable. Still, the message behind it is instructive: even the largest providers are moving away from one-size-fits-all solutions toward modular building blocks tailored to the use case.

The Right Question: Not "Which App," but "Which Process"

Put these figures together, and a clear picture emerges: the decision between ChatGPT Work, Claude Cowork, or a comparable super-app is rarely the decision that determines whether your AI initiative succeeds or fails. What matters is whether you've done the real homework beforehand: which process in your business currently costs the most time or money? What data does an agent need to reliably take over that process? And what does a control structure look like that prevents an autonomous agent from unsupervised making decisions that a human should actually be accountable for?

A narrow agent tailored to your core process can, in practice, achieve more impact than a generic suite with many unused extra features – simply because it's trained and integrated precisely on your business's data, systems, and approval processes, rather than on the lowest common denominator of every possible use case. According to industry analyses, investment in AI agent projects at SMBs ranges from €25,000 to €80,000 initially for SaaS approaches, from €150,000 for enterprise custom implementations, and up to €500,000 for complex multi-agent systems; the break-even for SaaS solutions typically comes after three to six months. These figures come from consulting and industry sources, not an independent primary study – but they provide a realistic order of magnitude against which you can measure a planned super-app investment.

Concretely, for your decision, that means: before signing a company-wide license for ChatGPT Work, Claude Cowork, or a comparable bundle, an honest stocktaking is worthwhile. How many of your existing AI tools are actually being used? Which department genuinely needs an autonomous agent with access to email, calendar, and file system – and which would do just as well, or better, with a much leaner, purpose-built tool? And: who in your business is responsible for making sure an agent that independently sends documents or reschedules appointments does so within clearly defined limits?

Bottom line: ChatGPT Work and Claude Cowork are technically impressive products, and both providers have shown in a short time how quickly autonomous agent capabilities are evolving. But for an SMB, the real question isn't which of these two super-apps is better. The real question is whether a company-wide all-in-one suite is even the right tool for your specific problem – or whether a purpose-built agent, integrated precisely into your process, gets you there faster, cheaper, and with less governance risk. The research speaks clearly: 95 percent failed ROI pilots, only 21 percent mature governance models, roughly half of software licenses unused. These numbers aren't a reason to avoid AI agents. They're a reason to flip the order of the decision: process first, then tool – not the other way around.