AI Enablement Radar week 31: choose a model per step and measure the outcome

This week, AI work became easier to divide and measure. OpenAI is cutting prices for parts of the GPT-5.6 family, Google has made agent evaluation generally available, and Notion can make meeting notes trigger the next step. For a smaller organization, that means one workflow can use different models, measure each stage, and hand clear decisions to a person.

Top signals this week

• OpenAI cut GPT-5.6 Luna API pricing by 80% to $0.20 per million input tokens and $1.20 per million output tokens. Terra became 20% cheaper. Model choice at each workflow stage is now a budget decision teams can test, not a technical footnote. Source: Advancing the price-performance frontier with GPT-5.6.

• Google made Agent and Model Evaluations in Gemini Enterprise Agent Platform generally available. It includes more than 20 ready-made metrics for quality, grounding, tool use, and the path an agent takes through a task. Source: Agent and Model Evaluations in Gemini Enterprise Agent Platform are now GA.

• Singapore's National University Health System shows why bounded tools work. Four separate AI tools help pharmacists with medication reconciliation, triage, and verification. Initial validation ranged from above 88% to 100% accuracy, while pharmacists keep the clinical decision. Source: NUHS unveils novel AI platform to strengthen medication safety and streamline pharmacy workflows.

• Notion Custom Agents can now start when an AI Meeting Note has been summarized. A meeting can lead to an updated project status, a Slack recap, or new tickets without someone moving text between systems by hand. Source: AI Meeting Notes can now trigger Custom Agents.

• The EU AI Omnibus entered into force on July 27. Timelines for some high-risk systems moved, but the AI Act Article 50 transparency requirements apply from August 2, 2026. Organizations now need to explain when people meet AI and how covered AI-generated content is marked. Sources: AI Omnibus enters into force. Guidelines on transparency obligations for providers and deployers of certain AI systems.

What organizations are actually doing with AI

SaaStr has three full-time employees and more than 20 AI agents. Its sales agent Hexi went live in three weeks, starting with 1,250 old leads. Salesforce reports that the work created $3.5 million in pipeline and contributed to $2.7 million in closed revenue. Hexi cannot book meetings, go below a discount floor, or promise speaking slots. That detail matters: the agent handles volume, while people own negotiations and promises.

Source: Agentic outreach turns SaaStr's warm leads into $2.7 million in sales.

Cognizant has trained more than 30,000 employees on Claude and is embedding the model in its own platforms. In two client deployments, Anthropic reports that contract review became up to 40% faster and an underwriting tool saved roughly eight hours per user per week. The figures are provider-reported, but the operating model is useful: training, domain knowledge, and evaluation travel together.

Source: Cognizant and Anthropic expand their partnership to bring Claude to enterprise clients.

Oracle plans to make Gemini models available through Oracle AI Agent Studio for Fusion Applications, with planned uses across finance, HR, supply chains, and customer service. This is a roadmap announcement, not proof that every integration is live. The direction still matters for smaller teams: AI is moving into systems where approvals and transactions already happen.

Source: Oracle to make Gemini models available to thousands of enterprise applications customers.

The tooling layer: platforms, agents, and workflows

An agentic workflow is a process where AI can choose and complete several steps with tools instead of only writing an answer. MCP, the Model Context Protocol, is an open way to connect AI to tools and context through a shared interface.

GitHub Copilot code review can now use repository-specific SKILL.md instructions and read context from MCP-connected issue trackers, documentation systems, and service catalogs. MCP calls during review are read-only, and comments disclose when a skill or MCP source contributed. Other organizations can borrow the pattern: connect sources with read access first, measure the output, and open write actions after the review rule proves useful.

Source: Copilot code review: Agent skills and MCP now generally available.

Make published several connector changes that can break existing workflows when ignored. Stripe connections need a new permission, GoHighLevel is moving from API keys to OAuth 2.0, and Google Sheets changed input options in four modules. An AI workflow is never just a prompt. Connections, permissions, and test records are part of the work.

Source: Action required for Stripe users, new Gemini module, and more app updates.

Shopify now lets apps add event annotations to analytics charts, with the app identified as the source. An annotation records what happened around a change but does not prove that the change caused the result. It is still better than reconstructing campaigns and operational changes from memory.

Source: View app-added annotations on your analytics charts.

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

AI governance means the decisions, roles, and controls that determine how AI may be used. Two parts became unusually concrete this week.

First, EU transparency rules apply from August 2. Providers must clearly inform people when they interact directly with AI. Covered AI-generated or manipulated content also needs machine-readable marking. For a customer workflow, a useful starting point is to record where AI meets the user, who reviews public text, and how marking evidence is retained.

Second, evals are repeatable checks of AI results, and they need to follow the work from testing into production. Google's new service can run consistent metrics on local experiments and production traces. A smaller organization does not need twenty metrics on day one. Pick an outcome measure, a factuality measure, and a stop condition. Log the model version, sources, cost, and human decision for every run that can affect real work.

Integrate professionally: keep keys in environment variables or a secret manager, give each connection scoped permissions, redact sensitive fields before the model step, and require approval before external or hard-to-reverse actions. AI can then do useful work without placing passwords or accountability in a chat transcript.

This week's practical Hammer test

Choose a recurring workflow that ends in a human decision, such as a customer reply, meeting follow-up, or proposal review. Give the test 40 minutes.

1. Write down the result the recipient needs and what must be true before it is accepted.
2. Split the workflow into four stages: gather evidence, analyze, draft, and act or send.
3. Choose the model by stage. Use a fast, low-cost model for sorting or format checks. Reserve the stronger model for uncertain analysis where better reasoning changes the outcome.
4. Put one measure on each stage: missing sources, correct format, handling time, cost per case, or edits after human review.
5. Add an approval immediately before sending, booking, publishing, or changing a system. Run three real but reviewed examples.
6. Finish with a short run receipt: model by stage, sources, cost, errors, human decision, and the next change.

Copy this instruction into your AI assistant:

Help me map an AI workflow for [work task]. Split the work into gathering evidence, analysis, drafting, and action. Suggest the type of model each stage needs and explain the choice in terms of quality, speed, and cost. Add a measurable acceptance condition to every stage. Mark which connections should have read access, which actions require human approval, and what belongs in the run receipt. Ask questions when evidence or ownership is missing. Do not write to external systems.

Companies and tools to watch

• OpenAI GPT-5.6: Lower prices make model routing worth retesting, one workflow stage at a time.
• Google Gemini Enterprise Agent Platform: It puts local test cases and production monitoring in the same evaluation layer.
• Notion Custom Agents: Meetings can become the start of a reviewed operational workflow.
• National University Health System: Its platform shows how bounded AI tools can support a sensitive process while people keep final responsibility.
• EU AI Office: Transparency becomes a concrete product and communications requirement from August 2.

If you want to make this type of workflow reusable inside the business, Hammer's Tool Forge can help with connectors, model choice, measurement, and approval steps. Start with a workflow where the benefit is visible after three reviewed runs.