AI Enablement Radar week 24: from chat to governed workflows

The clearest signal this week is not that another model became stronger. It is that AI is getting workspaces, permissions, sandboxes, and routines around it. For a small Nordic team, that points to a simple next step: stop asking "which chat answers best?" and start asking "which recurring job can we connect to the right sources, the right limits, and a human who approves the final step?"

Top signals this week

• BBVA shows what broad adoption looks like in practice. The bank says more than 100,000 employees use ChatGPT Enterprise, with more than 70 percent monthly and weekly active usage, roughly three hours saved per employee per week, and more than 20,000 custom GPTs created by employees. The numbers matter, but the setup matters more: security, legal, and compliance were involved from the start.

  Source: BBVA puts AI at the core of banking with OpenAI

• LSEG is moving AI from internal experiment to customer-facing product work. London Stock Exchange Group describes product cycles that used to take three to six months moving to roughly two weeks for some AI-adapted products, with customer requests reaching production in about four weeks. Enablement is not measured in prompt volume. It is measured in a shorter path from question to reviewed outcome.

  Source: From data to decisions: how LSEG is scaling trusted AI

• Meta is making customer conversations more agentic for small businesses. According to Meta, more than one million businesses already use a Meta Business Agent on WhatsApp and Messenger, with expansion to Instagram. The agent can answer questions, recommend products, book appointments, qualify leads, and let a human step in at the right point.

  Source: Be There for Every Customer With Meta Business Agent

• Coding agents are getting real work environments. The GitHub Copilot app gathers agent work into a desktop experience with sessions, plans, sandboxes, and review. Local and cloud Copilot sandboxes are also in public preview, so agent commands can run with restricted access to the filesystem, network, and system resources.

  Source: GitHub Copilot app: The agent-native desktop experience and Cloud and local sandboxes for GitHub Copilot now in public preview

• Long-running agents need more than a giant prompt. OpenAI highlights Skills, shell, and compaction as building blocks for agent workflows that read data, run commands, write files, and leave artifacts. A Skill is essentially a versioned work instruction the model can load when needed. Compaction means the system preserves and summarizes context during long runs.

  Source: Shell + Skills + Compaction: Tips for long-running agents that do real work

• Rules and risk frameworks are becoming more operational. The EU AI Act describes risk levels and obligations for AI systems, while NIST says the AI Risk Management Framework is being revised and has a concept note for critical infrastructure. For everyday organizations, this does not mean turning every AI test into legal paperwork. It means knowing which system is used, for what purpose, with which data, and who reviews the output.

  Source: AI Act, European Commission and AI Risk Management Framework, NIST

What organizations are actually doing with AI

BBVA is this week's clearest example of training before magic. The bank started with a limited ChatGPT Enterprise rollout to 3,000 employees in 2024 and has since expanded to more than 100,000 users. OpenAI also says 250 senior leaders, including the CEO and chair, have been trained. Smaller teams often miss that part: if leadership does not learn how the work should be reviewed, AI becomes a side project for the people who were already convinced.

Source: BBVA puts AI at the core of banking with OpenAI

LSEG shows another pattern. It connects generative AI to trusted data, product teams, customer needs, and internal workflows. That sounds big, but the same principle works in a smaller organization: start with a recurring question where the sources already exist, such as proposals, course material, support tickets, or weekly reports. AI should not guess. It should work from a source pack and leave something a human can check.

Source: From data to decisions: how LSEG is scaling trusted AI

Meta Business Agent matters for smaller businesses because it starts in the channels where customer questions already happen. WhatsApp, Messenger, and Instagram are not abstract enterprise platforms. They are where bookings, simple questions, product choices, and service issues often start. The practical work is to write the answer book, escalation rules, and product boundaries before the agent gets more autonomy.

Source: Be There for Every Customer With Meta Business Agent

The tooling layer: platforms, agents, and workflows

An agentic workflow is an AI workflow where the system does more than write text: it plans steps, uses tools, and leaves an output that can be reviewed. This week's tooling signal is that vendors are building around that: workspaces, runtimes, memory, policy, and follow-up.

GitHub makes this visible for development teams. The Copilot app treats agent work as something you should be able to follow, pause, review, and return to. The sandboxes may matter even more for many organizations: the agent can run commands in an isolated environment instead of getting loose access to an entire machine or repository.

Source: GitHub Copilot app: The agent-native desktop experience and Cloud and local sandboxes for GitHub Copilot now in public preview

OpenAI's Skills pattern points in the same direction, but more generally. Put stable routines, templates, examples, and stop rules into a reusable skill instead of pasting everything into every prompt. Then an agent can work more consistently, and the team can change the routine in one place.

Source: Shell + Skills + Compaction: Tips for long-running agents that do real work

Google Workspace shows how this can land in office work. Workspace Studio is getting reusable AI automation flows called skills, and Google also highlights a Workspace MCP Server. MCP, Model Context Protocol, is a way for AI tools to connect to external systems and data sources in a more standardized way. For small teams, the question is plain: which sources may AI read, which systems may it write to, and when must a human press approve?

Source: 10 more announcements for Workspace at Google Cloud Next 2026

Microsoft and Google both describe the larger platform direction: AI needs to be built, run, governed, measured, and improved inside a coherent system. That can sound like enterprise language. At Hammer scale it means this: have one place for sources, one place for instructions, one place for permissions, and one place where someone can see what the AI did.

Source: AI alone won't change your business. The system running it will. and Google Cloud Next 2026 Wrap Up

Governance and risk: what needs to be in place before scaling

AI governance does not mean slowing everything down. It means deciding what AI may do, with which data, under which permissions, and how the result gets checked. That matters more when AI starts using tools, reading internal documents, or writing back to business systems.

The EU AI Act uses a risk-based model with prohibited, high-risk, transparency, and low/minimal-risk systems. Many everyday workflows in small organizations will not fall into the heaviest categories, but the logic still helps: classify the use before connecting it to real customers, students, employees, or money.

Source: AI Act, European Commission

The EDPB and EDPS have also pointed to the need for stronger safeguards for fundamental rights when AI rules are simplified. That is a useful warning light for practical adoption: simplification is good, but it cannot make personal data, decision influence, or automated profiling fuzzy.

Source: Artificial intelligence, European Data Protection Board

For a smaller team, a clear starting setup is often enough: use environment variables or a secret manager for keys, give AI scoped API keys with least-privilege permissions, redact sensitive information when it is not needed, run risky steps in a sandbox, require human approval before external sends or payments, and keep a simple audit log of what the agent did.

Source: AI Risk Management Framework, NIST

This week's practical Hammer test

Try a "source-backed work card" in 30 to 45 minutes. Pick one job that already repeats every week: support handover, proposal draft, lesson planning, customer follow-up, or an internal weekly report.

Do this:

1. Put five to ten sources in one folder: previous examples, price list, policy, customer questions, notes, or course material.
2. Write what AI may do: read, summarize, suggest, compare, or draft.
3. Write what AI may not do without human approval: send, publish, change prices, book appointments, contact a customer, or write to a business system.
4. Run the task in your AI environment and ask it to leave a work card: sources used, assumptions, uncertainty, recommendation, and the next human decision.
5. Save the result. Next week, compare it: did the work get faster, better, or only more polished?

Copy this prompt:

You are helping us with a recurring workflow. Use only the sources in this folder and clearly mark anything that is missing.

Task: [describe the work]

Create a work card with:
1. Which sources you used
2. Which assumptions you made
3. Suggested draft or action
4. Risks, open questions, and what a human must check
5. The next decision point

You may not send, publish, book, change prices, or write to external systems. Only suggest the next step for human review.

This is a good first step for Tool Forge: not a huge AI project, just a small workflow with sources, permissions, approvals, and a measurable improvement.

Companies and tools to watch

• BBVA: shows how internal AI training, leadership support, and secure environments can scale usage in a regulated industry.
• LSEG: shows how generative AI can shorten the path from data and customer need to product change.
• Meta Business Agent: makes customer-service and sales agents more accessible in everyday channels for small businesses.
• GitHub Copilot sandboxes: make agent execution more concrete by limiting what the agent can access.
• Google Workspace Studio and Workspace MCP Server: point toward AI workflows built directly on top of office data and approved integrations.

If you want to build this kind of workflow without making it bigger than it needs to be, Hammer Automations Tool Forge helps you choose one real workflow, connect the right sources, and set permissions, approvals, and logging before you scale.