AI news and DevDay: GPT-6.1 Sol, Sonnet 5.5 and new ChatGPT plans

What has changed since 23 September?

We reviewed official model announcements, product notes, help articles and API changelogs from OpenAI, Anthropic, Google and Microsoft. The focus is changes between 23 and 29 September 2026 that matter to businesses with roughly 5–250 employees. Previously covered products such as GPT-6 Sol and Luna, Claude Docs and Gemini 3.8 appear where there is a new development or a correction to make. The related guides provide the longer setup instructions.

Editorial cutoff: 29 September 2026, 21:18 CEST. The DevDay section covers all 25 product cards in the official recap and additional product and developer documentation. An event presentation does not automatically mean a capability is new or generally available. Prices are vendors’ US list prices, not lol.marketing offers; taxes and local billing may differ.

ChatGPT Pro 200 returns: have the limits been halved?

The current Pro help page confirms new $200 monthly subscriptions with less included usage. Eligible existing subscribers can retain their previous allowance while Pro 200 remains active through 29 October 2026, after which it decreases. The page gives neither a reliable percentage reduction nor a specific eligibility cutoff date. It therefore does not substantiate “almost halved”.

The Codex pricing documentation checked before DevDay showed Pro 5x and Pro 20x alongside estimated messages per five hours. These earlier estimates do not establish the new allowances. They are not fixed weekly quotas and cannot be equated with a percentage meter for different models or longer tasks. They do not justify calculating a new subscription multiplier.

For a purchasing decision, check your account information and any change notification. Record the plan, model, task type and reset window before comparing usage. Dollar prices here are vendor figures; check the actual German billing amount at checkout. A personal Pro subscription and a Business workspace are separate purchasing decisions.

Work, Codex and Office: plan usage together

OpenAI’s credits documentation describes a shared agentic usage pool for supported features on personal plans, including Codex, Work and Office extensions. Additional credits are separate from API credits. When an allowance runs short, investigate document work as well as coding tasks. Available purchase options depend on the account.

Our recommendation is to track the cost and handling time of recurring tasks for a week, including human review. Keep the model and reasoning setting constant. A more expensive subscription may reduce interruptions, but a larger allowance cannot replace clear task instructions. Measure the effort per accepted result.

DevDay: GPT-6.1 Sol, Pro 500 and Astra Ultrafast

On 29 September, OpenAI announced Pro 500 at US$500 per month. Initially available to subscribe to on the web, it provides 25 times the included usage allowance of Plus according to the DevDay recap, alongside Astra Ultrafast in Work and Codex. This multiplier is not a fixed token or message count. Among Pro subscriptions, Ultrafast is exclusive to Pro 500; buying credits does not unlock it on Pro 100 or Pro 200. Pro 200 returning with a smaller allowance is a separate change. Public sources still do not reliably quantify that reduction.

Ultrafast runs GPT-6 Astra at a faster service tier. Speed claims concern token generation and do not guarantee an equivalent reduction in every complete task. The Work/Codex documentation specifies that included usage is consumed at eight times the Standard rate; purchased credits and Enterprise pay-as-you-go usage use a sixfold rate. Eligible Enterprise and Edu plans can also qualify. Ultrafast starts disabled in Enterprise. Required inference residency outside the United States rules it out; the company’s location alone does not determine eligibility.

GPT-6.1 Sol is a new model for coding, computer use and multi-step professional work. While the launch post announces availability, the detailed help describes a gradual rollout: Pro first, followed by Plus, Business, Enterprise and Edu; Free and Go are excluded at launch. Enterprise and Edu require an administrator to enable the model. In ChatGPT it belongs to Work and Codex and is not yet available in ordinary Chat. Standard and Fast are supported; Ultrafast for GPT-6.1 Sol follows in the coming days. An SME pilot should reuse previously assessed tasks so quality, time and consumption remain comparable.

Dots: persistent agents with different regional rules

A dot uses GPT-6 Astra on its own cloud computer and can continue work between conversations. The user defines its goal, connected applications and permitted actions. A business might pilot a regularly updated project overview whose results are reviewed by an accountable person. This is a possible workflow, not a promise of an error-free autonomous employee.

The gradual rollout is for eligible adults. Pro initially excludes the EEA, Switzerland and the UK, including Germany. The dot help page provides Business Premium access across supported ChatGPT regions. Enterprise receives an administrator-enabled beta that starts disabled. For the first month, eligible users' dot usage will not count toward plan allowances; subsequent terms have not yet been published. This is not a permanent free-usage promise.

Dots can proactively read connected-app information and form memories from it. Disconnecting an app does not automatically erase information already obtained. Local computer access starts disabled and requires separate authorization. Setup takes place in the desktop app or a desktop browser; the texting beta is limited to eligible Pro users in the United States. A first business pilot should use a limited dataset, clear approval rules and an outcome that can be checked.

OpenAI also announced Specialist Dots for organizations: specialized agents with their own enterprise identity and deliberately assigned permissions. These begin with selected Enterprise pilots. The announced collaboration with Microsoft aims to bring these dots into Microsoft Agent 365 governance and security controls. This is neither general availability of these organizational agents nor an already fully delivered Agent 365 integration. The useful principle for SMEs is a specific task, an accountable person and only the permissions required for that work.

Space, Pages, slides and Meetings connect documents and teamwork

ChatGPT Space brings together Pages, files and shared work for eligible Pro, Business and Enterprise accounts on web and desktop. It replaces Library for accounts with access; existing Projects remain separate. Enterprise sharing is an opt-in preview. Pages support collaborative writing, comments, tables and interactive content. A team could use one to document a project decision alongside its evidence, open questions and responsibilities.

Sharing a Page does not expose private chats or memories. Material copied into the Page is visible to its readers, however. Connected source files retain their own permissions, but a copied summary becomes newly shared content. For client projects, teams should therefore review excerpts as well as source-file access before sharing.

The new Meetings plugin turns conversations into personal summaries and action items in Space. Its detailed help describes a gradual beta rollout for Pro and Business in the macOS desktop app. Enterprise initially receives a limited alpha through account managers. Other platforms are planned for later. Audio is deleted after processing and cannot subsequently be replayed. Sharing notes does not automatically share the transcript. Before a pilot, establish consent, recording authorization and recipients; the announcement does not explicitly guarantee availability in every region.

Collaborative presentation creation and editing is another DevDay announcement for Space. These collaborative slides are planned for Pro, Business and Enterprise in the coming weeks, so they are not a generally available feature on 29 September. A possible future workflow would turn an agreed Page into a client presentation that the team can refine together. Until access appears in their accounts, businesses should continue to plan around their existing presentation process.

Edit Sites, schedule updates and share profiles

Plus or Pro owners and editors can reopen a published ChatGPT Site from a desktop browser, review changes and publish an update. Owners can also create recurring cloud schedules from Automations. These additions apply where Sites is supported. They do not establish a new German rollout: the reviewed Codex help page still excludes the EEA, Switzerland and the UK, while the general Sites help page has no equivalent country list.

A separate announcement enables embedding supported ChatGPT plugins in custom-built Sites. The DevDay recap lists Business, Enterprise, Healthcare and Edu for this capability. Colleagues can use the same Site with their own connected data and permissions. This organizational announcement does not establish plugin integration for personal Plus or Pro accounts. Existing workspace and data permissions still apply. The Sites help page also states that Sites does not support data or inference residency. A hosted internal workflow therefore needs a review of its content, access and operating requirements.

Shareable ChatGPT profiles can showcase selected Sites. Personal profiles start private and visitors must sign in to ChatGPT. Business profiles follow workspace settings. The reviewed sources disagree on Enterprise availability: the DevDay recap lists Business and Enterprise as available, while the release notes or detailed help still describe Enterprise as a later rollout. Organizations should therefore check actual access in their workspace. Showcasing a Site does not expand its access permissions.

Sign in to third-party apps with ChatGPT and authorize usage

Sign in with ChatGPT initially provides identity sign-in to participating services and is globally available according to the help page, subject to administrator policies. The partner receives a name, email address and profile picture if present. This does not disclose private conversations or memories. Separately, eligible Plus and Pro users in a limited preview can permit individual partner applications to consume their existing ChatGPT model allowance and set an application-specific limit.

That permission creates no extra allowance and does not require a personal API key. Spending purchased credits requires a separate opt-in, and the partner can still charge for its own services. Budget planning must distinguish identity sign-in, data permissions and authorization to consume usage. A practical starting point is one approved application with a low limit and a documented review of its actual consumption.

Codex Cloud, refreshed CLI, Code Review and Security Cloud

Codex Cloud uses prepared environments containing repositories, tools and dependencies. Each task receives its own workspace and can continue while a laptop is closed; results are accessible from web, mobile or desktop. The environment is configured, tested and published before use. Current documented limitations include no computer or browser use, GitLab or self-hosted GitHub Enterprise Server in these environments. A team could test its web application overnight while retaining a separate production approval step.

The Code Review plugin supports examining pull requests, summaries, diffs and findings in supported desktop and web experiences. Repository and workspace permissions still govern access. Codex Security Cloud additionally checks connected GitHub repositories for security issues on demand, on a schedule and in new commits. It investigates findings and prepares potential fixes for review; proposed patches are not automatically applied to the repository.

The DevDay recap lists Pro, Business, Enterprise and Edu for Codex Security Cloud. Appropriate permissions, repository setup and billing are also required; the plan list alone does not confirm access in every account or region. Existing Security customers must opt in to paid usage. The Daybreak Blue access provided through this feature applies inside Security Cloud and is not general model access. For an SME maintaining software, a limited scan of an approved repository followed by a review of substantiated findings is a practical first step.

The Codex CLI also receives a refreshed interface: voice input helps users describe work, and a new view through “/agents” makes multi-agent work accessible. This product announcement complements the individual maintenance releases of 25 and 28 September. It is particularly relevant to a development team working on several clearly bounded tasks in parallel. Voice control and an agent overview do not replace repository permissions or review before a merge or deployment.

Teams and Team Tasks: shared ownership of recurring work

Teams inside ChatGPT group colleagues in a Business or Enterprise workspace. Pages and Spaces can then be shared with that group. Shared Team Tasks run in the cloud on a schedule or supported app events. Authorized members can review results and refine instructions. OpenAI is rolling out these features gradually. A weekly project report based on explicitly selected documents would be a useful SME pilot.

A task uses configured team connections and their account permissions. It does not automatically inherit its creator’s personal chats, memories or app access. New team members can see earlier task runs, and membership does not automatically synchronize with Slack or Microsoft Teams. The admin documentation specifies workspace credits and separate team spending limits. Define sources, membership, a budget and an accountable owner before starting the pilot.

@ChatGPT joins work directly in Slack and Microsoft Teams

After admin setup, @ChatGPT can be mentioned in enabled channels, threads and direct messages. The product page lists Business, Enterprise and Edu. Colleagues can participate in enabled channels without an individual ChatGPT license. A project discussion could become a collaboratively refined brief. Access must be provided by the organization; an arbitrary Slack account is insufficient.

Connections and permissions determine data access. Replies are visible to conversation participants, while source files may have different sharing settings. Start with a limited project channel, a clearly defined task and a result that the team can verify.

OpenAI Marketplace connects enterprise commitments with partner software

OpenAI Marketplace serves eligible enterprise customers. Part of an existing contractual OpenAI spending commitment can count toward approved partner products. Eligibility and the amount depend on the agreement, product and program. Customers contract with and pay the software partner directly. For SMEs, this is primarily a procurement consideration when an enterprise agreement already exists; an ordinary ChatGPT subscription does not establish eligibility. Confirm supplier terms, licensing scope and actual costs before purchasing.

GPT-6.1 Sol: a new API model with cheaper cache reads

GPT-6.1 Sol has been available in the API as gpt-6.1-sol since 29 September. OpenAI positions it for demanding coding, computer use and professional work. For requests with up to 272,000 input tokens, Standard processing costs $2 for input, $0.10 for cache reads, $2.50 for cache writes and $10 for output per million tokens. Cache reads therefore cost half as much as GPT-6 Sol's; this does not halve the total cost of an application. The context window is 1.05 million tokens, with a maximum output of 128,000 tokens. Requests exceeding 272,000 input tokens incur twice the input and cache rates and 1.5 times the output rate across the entire request. API billing is separate from a ChatGPT subscription.

Migration may require more than changing the model name: tool calls require the Responses API, while Chat Completions supports this model only without tools. The none and minimal reasoning settings are unavailable. Supported efforts are low, medium, high, xhigh and max, with medium as the default. Existing sampling and logprob parameters must also match the new configuration. Germany is a supported API country. EU data residency is documented for Sol 6.1 with Standard, Flex and Batch processing, but not Fast. Before switching, compare ten real proposal reviews for accuracy, turnaround time and total cost.

Astra Ultrafast: premium speed without EU processing

GPT-6 Astra gains the Ultrafast API tier on 29 September. Developers select it with service_tier=ultrafast in Responses requests. For up to 272,000 input tokens, the pricing table lists $60 for input, $6 for cache reads, $75 for cache writes and $300 for output per million tokens. API users have access within their rate limits. Global processing and US data residency are supported; EU and other regional endpoints are excluded. Comparing it with Standard may make sense for time-critical analysis, while the premium is harder to justify for overnight batch work. GPT-6.1 Sol Ultrafast is announced for the coming days.

A browser for the Agents API and Multi-agent in Responses

The 29 September addition is Computer Use in the Agents API: an OpenAI-hosted browser can operate web interfaces. The application must handle website-access approvals and sign-in. A small business could use it to test an internal ordering workflow in a test environment. The Agents API itself launched in public beta on 10 September. Billing covers models, tools and applicable container usage, without an additional Agents API fee. Its current data residency is limited to the US; Zero Data Retention is unsupported even when using a self-hosted sandbox.

Separately, the Responses API supports GPT-6.1 Sol through a Multi-agent beta; the GPT-5.6 models are also documented. A root agent can delegate independent tasks and combine the results. For a proposal comparison, three agents could separately inspect scope, deadlines and exclusions. The beta requires explicit activation and can consume more tokens. Shared writes and strictly sequential tasks are less suitable. Application developers still need to control permissions and spending.

Decisions API: a limited preview for bounded decisions

The Decisions API uses Luna to answer questions from a predefined, finite set of options, using text or images as context. Routing support requests to billing, technical support or sales would be a potential pilot. Access is limited to a preview on 29 September, with broader release planned in the coming days. The primary sources reviewed do not yet establish firm prices, quotas or specific German/EEA availability.

Private Intelligence: distinguish safety processing from private inference

Zero Data Retention with Private Safety Processing enables automated safety reviews in a protected, hardware-backed environment without human access to the protected content. Encrypted safety records are retained for at least 30 days in customer-controlled cloud storage. Organizations already approved for ZDR can configure it per project; storage region, permissions and any customer-managed keys must be correctly configured. Businesses should account for these operational responsibilities before enabling it.

Private Inference is a separate announced component: a preview using confidential computing is planned for the fall. This does not establish general availability, a particular European region or a price today. Assess each capability against its actual data flows and eligibility requirements.

Bedrock Managed Agents: OpenAI agents within AWS

The DevDay recap highlights Bedrock Managed Agents: orchestration and model inference run in Amazon Bedrock, while commands and tools use AgentCore Runtime or self-hosted compute. AWS still labels the service a limited preview; its original announcement dates back to April. Businesses already using AWS may want to assess it as an operating option. It is not a drop-in replacement for the OpenAI endpoint: authentication uses AWS IAM and SigV4, and interfaces and capabilities may differ. Specific regions, including Frankfurt, prices and account access still need confirmation for this preview.

Plugins as work surfaces and automation triggers

Plugin Extensions add sidebar apps, panels beside conversations and custom file viewers and editors to ChatGPT. A software provider could place a project dashboard next to a customer conversation. The detailed documentation specifies important limits: extensions on the web are coming later for Free and Go, while mentioning plugin content in the composer is desktop-only. OpenAI also announces Plugin Creator, revised submission with clearer feedback and improved discovery. Publication still requires review and appropriate access controls.

MCP Events connects plugins to event-driven automations, for example when an approved project board reports a new task. Users specify what to monitor and how to respond. The documented integration requires MCP 2.0 protocol version 2026-07-28, persistent subscriptions and signed HTTPS webhooks. It does not support every mode in the draft Events specification; polling and streaming are excluded here. Drafts and internal summaries make useful initial business pilots, with outputs reviewed before external actions.

Sonnet 5.5: what is new and what does it cost?

Anthropic released Sonnet 5.5 on 28 September. API list prices remain $2 per million input tokens and $10 per million output tokens. Claimed savings come from using fewer tokens, rather than lower rates. Anthropic reports output more than 30% faster and task costs up to 30% lower than Sonnet 5. These are vendor findings, not a guarantee for your workflow.

Good first comparisons include a proposal outline from approved notes, a presentation revision or a small bug fix. Use identical inputs and assess completeness, factual errors, editing effort and total cost. Keep demanding or poorly specified work on a proven process initially. A faster first draft provides little benefit if someone then has to repair its figures. Haiku 5.5 is announced for the coming weeks, but is not a confirmed release at this cutoff.

Adopting Sonnet 5.5: distinguish selection from migration

Selecting a model in Claude differs from migrating an API application. The developer model ID is claude-sonnet-5-5. Applications that disabled thinking can use between_tools for no up-front thinking, while thinking blocks may still occur between tool calls. This option supports low, medium and high effort; xhigh and max require adaptive thinking. Follow the migration guide for the remaining restrictions. Test tool use, structured outputs and stored conversations before changing production. A valid model name alone does not establish integration compatibility.

For a German business, record the access path you actually use: the Claude app, a direct API or a cloud provider. Availability on several platforms does not establish a particular German deployment region, subscription entitlement or data-processing arrangement. Retain the previous model path as a fallback and assign responsibility for accepting or rejecting results.

OpenAI: account security, image understanding and Codex updates

ChatGPT added Security History for past account events on 25 September. It complements active-session checks; location and device information can be imprecise. On the same day, OpenAI reported a fix to image encoding in GPT-6 Sol and Luna. Repeat affected evaluations involving screenshots, charts or computer use. The fix is not a new model generation.

For installed macOS clients, the documented security update to version 26.924.20706 matters. Codex CLI releases on 25 and 28 September add model support, MCP authentication options and fixes to network and approval behaviour. Update through the official installation channel and check the connections you use. An integration capability does not grant permission for every dataset or action.

For completeness, the 28 September ChatGPT notes also describe new content-selection options in Health. We do not treat that as a workplace employee-health workflow. Smaller mobile interface changes are grouped here rather than turned into a separate guide for every maintenance release.

Claude Tag and plugins: what changes for teams?

Personal connectors for Claude Tag in Slack channels began rolling out to Team plans on 24 September, with Enterprise to follow. A response can draw on a personal connection and then become visible to the whole channel. Source access and permission to share the result are separate checks. Start with review before posting, for example when summarising sales information from your CRM. Personal connectors apply to your own requests; unattended routines continue to use channel connections.

On 25 September, Anthropic also opened a plugin submission portal to developers with paid Claude plans. An MCP server or bundle can be submitted for review and published after approval. This matters to providers of reusable customer solutions; subscribing does not guarantee acceptance. Internal teams should still assess available plugins for data access, maintenance and ownership.

Are Claude Docs and Slides now available for free?

We cannot confirm general free availability at this review date. Official app release notes include Free, while dedicated help pages and the pricing table still describe paid-plan access. We flag that conflict in our existing guides as well. An edited release-note entry establishes neither the change date nor access for every account. Enterprise owner activation remains relevant.

Before including Docs or Slides in a workshop, test the feature with the account participants will actually use. Check exports, collaboration and source permissions too. Successfully creating an ordinary artifact does not prove access to a particular document template.

Claude for developers: migration, caching and billing

Platform notes list Cache Diagnostics as generally available from 23 September, an addition to our previous cutoff. Developers can investigate cache misses on the Claude API; Bedrock and Google Cloud are not included. On 24 September, billing changed for some refusals before output, alongside Enterprise compliance interfaces. Do not assume every refused request is free, and check existing export parsers against current documentation.

Claude Code releases on 25 and 28 September add tighter model controls, project-instruction auditing and Sonnet 5.5 support. Include these updates in a managed maintenance window. Verify that existing tasks start, approvals remain effective and outputs retain the expected format.

The Compliance API now covers sessions from the Excel, PowerPoint, Word and Outlook add-ins where Enterprise prerequisites are met. Filenames and titles are absent from the Activity Feed from 24 September, so existing reports should not depend on those fields. Retrieving further information may require metadata permissions.

Google: Live Avatar extends Gemini 3.8 Live

Google introduced Live Avatar for Gemini Enterprise on 24 September, combining real-time conversation with a streamed avatar. Custom avatars require additional Enterprise approval. The model documentation lists an EU multi-region; that does not establish included Workspace licensing or access for every German account. This is a Live capability, not a confirmed Gemini 3.9 launch.

A possible pilot is assisted explanation of a product or internal procedure. Assess clarity, response time, incorrect claims and handover to a person. A convincing avatar can make an incorrect answer seem more credible. Approved source material and a narrow task scope should therefore be acceptance criteria.

Workspace: notebook sources, Vids and the Meet rollout

Two additions to our previous cutoff were announced on 23 September: Google Docs can use existing Gemini Notebooks as sources, and Google Vids gains Omni 1.1 Flash with better scene extension, 1080p output and more precise clip duration. Docs attaches notebooks through @, subject to an eligible plan and enabled AI features, including certain personal Google AI plans. Omni was already in Vids; the version and capabilities are new. Usage limits depend on the account.

For proposals or training materials, a curated source collection may be more useful than additional disconnected chats. Sample the output to check that figures and claims really come from the selected documents. Generated video needs brand approval, rights to source materials and factual review before publication.

The previously announced rollout of new automatic-note settings in Meet also starts on 29 September and may take more than 15 days. Administrator defaults differ by edition. Starting today does not mean enabled everywhere or identical defaults for all businesses. Check the setting before recurring meetings are automatically documented.

Microsoft: Home, Code and Autopilot — what is available?

Microsoft announced a new Copilot with Home, Code and Autopilot on 25 September. Home brings together Chat, Cowork and editable Word, Excel and PowerPoint files, while Code targets small applications and workflows. Autopilot renames the personal cloud agent previously called Scout. Home and Code are planned for Frontier previews, with Autopilot initially in an expanded private preview. This does not establish access for every business or consumer account on 29 September. A shared plugin directory is also rolling out gradually.

One change explicitly described as generally available is Fabric IQ for Chat and Cowork, supplying curated business data and Power BI semantics as additional context. Its integration into Code is planned for a later Frontier preview. For data-driven reporting, start by testing your existing, maintained metric definitions.

The related Managed Runtime was announced in public preview to host and manage Copilot-generated code. A business still needs to decide who approves an application, what data it can access and how changes are rolled back. Start with a small internal application and an explicit acceptance test.

Microsoft also confirms Sonnet 5.5 in Foundry from 28 September. That is an access path for custom applications and agents, not proof of the same model in Microsoft 365 Copilot. Work IQ with additional business context remains an announcement at our cutoff: public preview is scheduled to begin on 30 September. We do not count that future date as a completed rollout.

Copilot costs: assess licensing and consumption together

Microsoft’s pricing explanation distinguishes everyday licensed use from more extensive consumption-based agent work. It means neither an extra charge for every prompt nor that every Cowork, Code or Autopilot task is included in a seat price. Related FinOps capabilities aim to improve spending visibility and controls. Check actual entitlements and euro pricing in your contract.

Set a pilot budget, an accountable owner and a stop condition for unexpected consumption. Track accepted results and rework alongside credits or spending. Purview and Entra controls discussed in September do not remove the need to correct inappropriate access. We do not present vendor comparison calculations as promised savings for your business.

Foundry: voice agents and better diagnostics

On 24 September, Microsoft also introduced Voice Agents and Insights in Foundry. Voice Agents brings together development, tools, deployment and evaluation for voice agents. Insights analyses agent traces and proposes evidence-based improvements; it is explicitly in public preview and does not deploy changes itself. Model and telemetry usage may incur additional costs.

For a business voice assistant, this is a reason to review an existing pilot for interrupted conversations, tool failures and human handover. The Voice Agents announcement does not establish a clear general-release status or complete regional list. We do not infer guaranteed German availability. Its planned VS Code toolkit should also be distinguished from portal functionality already described as usable.

What should a small business do now?

Start with changes that affect current work: review Pro 200 transition terms, update installed clients, repeat visual GPT-6 evaluations and test any Sonnet 5.5 migration. Optional pilots with Live Avatar, document sources or Copilot Code come next. Begin with existing licences and clearly defined datasets.

Should we switch immediately because there are new models? No: compare real tasks, including expert review. Are all these features available in Germany? Country support, an EU model region and workspace access are three separate claims. Does a Pro or Team subscription automatically cover API use? No; check that billing path separately.

If the choices become difficult to navigate, start with the workflow: which recurring task takes time, which sources are approved and how will you recognise a correct result? The AI System Check helps organise those prerequisites and narrow down a useful first application. Examples in this article are suggested tests, not claimed customer outcomes.

Decision matrix