Avalerion Intelligence · Edition 019
Tuesday 20 October 2026 · AI, Data & Enterprise Transformation for the Nordics, UK & Ireland
Your next coworker has an ID and a spend cap
| Move | Who | Why it matters to you |
|---|---|---|
| Gemini agent: own identity, audit trail attributed to the agent, hard spend caps per project | Google (8 October) | Agent governance is now a product feature you can demand from every vendor |
| Haiku 5.5 costs about 75 percent less to run than Haiku 4.5 | Anthropic (7 October) | Routine agent steps get cheap, so volume, not price, becomes the budget risk |
| Ten model developers change data practices; call for evidence on agents closes 20 November | UK ICO (8 October) | Deployers, not only developers, are being asked how their agents handle personal data |
| Joule Work begins customer rollout after 110,000 internal users and 20 percent productivity gains | SAP (6 October) | The ERP vendor's AI layer arrives this quarter, grounded in its own data model |
| Trusted Agent Platform goes headless: any agent works under Workday permissions and audit trails | Workday (13 October) | HR and finance agents can run without anyone logging in, so the rules must travel with them |
| EU inference routes AI requests to European endpoints, and fails rather than leave the boundary | Lovable (12 October) | Where your AI runs is now a separate question from where your data is stored |
1. Top AI News
Google gave its agent an identity, a log and a budget.
What happened. On 8 October Google Cloud introduced the Gemini agent, one agent for knowledge work and code across Workspace and business systems. Each agent has its own identity, role-based permissions approved by security administrators, and an audit trail that attributes every action to the agent, not a person. A hard spend cap per project pauses the agent when hit. It routes jobs across Gemini and Anthropic's Claude models.
Why it matters. Whoever controls identity, logs and spend controls the agent. Expect every vendor to be asked for all three.
What to do. Add three questions to every agent procurement: whose identity does it act under, where is its log, and what stops the spend.
The routine agent step just got much cheaper.
What happened. On 7 October Anthropic released Claude Haiku 5.5, which it says costs on average about 75 percent less to run than Haiku 4.5. Input tokens fall from $1.00 to $0.10 per million for prompts up to 100,000 tokens, and Anthropic is halving the price of Sonnet 5.5.
Why it matters. Cheap steps invite volume, and the bill moves to the calls nobody counts.
What to do. Route simple, high-volume tasks to small models, and put a cap on each project before usage grows.
The UK privacy regulator has opened its file on agents.
On 8 October the ICO said ten foundation model developers, including OpenAI, Google, Microsoft and Anthropic, have changed or committed to change how they handle personal data. It also opened a six-week call for evidence on agentic AI, closing 20 November, aimed at deployers as well as developers. Its words: autonomy "is not an excuse for poor compliance".
2. Enterprise AI Trend: the ERP layer goes conversational
SAP rolls Joule Work out to customers.
What happened. At SAP Connect on 6 October, SAP said Joule Work, its AI interface across SAP and third-party data, is beginning its customer rollout. Inside SAP it is live with 110,000 employees, with reported productivity gains of 20 percent across finance, HR and procurement. SAP LeanIX now offers a central inventory of enterprise agents.
Why it matters. The vendor that holds your ledger wants to be its front door. The 20 percent is SAP measuring itself, on a clean SAP estate.
What to do. Before rollout, test Joule on your own messiest process, not the demo one, and add every agent to an inventory from the first day.
3. Data & AI Readiness: where the AI runs is a data question
A Stockholm company separates residency from inference.
What happened. On 12 October Lovable introduced EU inference for enterprise teams. Supported AI model requests are routed to designated European endpoints, and if no permitted endpoint can serve a request, it fails instead of leaving the boundary. Lovable is explicit about the limits: it covers processing, not storage, and web search still uses US services.
Why it matters. Most security reviews ask where data is stored. Few ask where prompts and context are processed. They are different answers.
What to do. Add "where is inference processed, and what happens when that region is unavailable" to your AI vendor questionnaire.
4. Business Process Spotlight: headless HR and finance
Workday: headless can't mean lawless.
What happened. At Workday Rising on 13 October, Workday made its Trusted Agent Platform headless: agents built by Workday, customers or partners can work with HR and finance data without anyone logging in, and each request carries Workday's permissions, approvals and audit trail. New autopilot agents will spot missing pay data before payroll runs.
Why it matters. The risk Workday names is real: an agent pointed at a copy of HR data in a data lake can read salaries it should never see.
What to do. Find every place an agent or integration reads HR or finance data through a system account or a copy, and move it behind the system's own permissions.
5. Nordic Technology: Swedish privacy tooling meets UK health data
AI Sweden takes LeakPro into Genomics England.
What happened. On 7 October AI Sweden signed a research partnership with Genomics England, owned by the UK Department of Health and Social Care. Its LeakPro tool will be tested to flag privacy risks in AI models trained on genomic data, inside the review that controls what leaves Genomics England's trusted research environment.
Why it matters. Sharing a model trained on sensitive data is a privacy decision in itself, and regulators will expect the risk to be measured.
What to do. If you train or fine-tune models on personal data, ask what a model could reveal before it leaves your environment, not only what data went in.
Join us in Stockholm
Thursday 5 November 2026, 12:00 to 13:30, central Stockholm. Free.
A working lunch with your laptop open: you build a real lifecycle flow in Customer.io, guided by our team and Customer.io's, and leave with it working in your own environment. Not a demo, not a panel. Seats are kept small; the venue goes to confirmed seats.
Do this next. Thirty minutes with us: which agents you are about to switch on, whose identity they act under, who reads their logs and who owns their budget. No pitch, no deck.
Recommended reading: What AI readiness actually means for Nordic companies and Why most enterprise AI projects still fail, at avalerions.com/insights
Frequently asked questions
What is Google's Gemini agent?
The Gemini agent, announced by Google Cloud at Gemini at Work on 8 October 2026, is a single agent for knowledge work and coding that works inside Workspace and connects to business systems. Each agent has its own identity and role-based permissions, every action is logged to the agent, and projects can set hard spend caps that pause the agent when reached.
What did the ICO announce about agentic AI?
On 8 October 2026 the UK Information Commissioner's Office said ten foundation model developers, including Amazon, Anthropic, Apple, Google, Meta, Microsoft and OpenAI, had made or committed to data protection changes. It also opened a call for evidence on agentic AI for developers and deployers, closing on 20 November 2026, to inform a statutory code on AI and automated decision-making.
What does Workday's headless Trusted Agent Platform do?
Announced on 13 October 2026, it lets agents built by Workday, customers or partners work with Workday HR and finance data without a user logging in. Every request carries Workday's permissions, approvals and audit trail, using Model Context Protocol tools any compatible agent can call. Workday Data Cloud and Agent Harness are generally available.
What is Lovable's EU inference?
EU inference, launched by Stockholm-based Lovable on 12 October 2026 for Enterprise teams, routes supported AI model requests to designated European endpoints while building and in published apps. If no permitted endpoint can serve a request, it fails rather than leave the boundary. It covers model processing, not data storage, and web search still runs on US services.
Sources
- Google Cloud, "Welcome to Gemini at Work 2026: Introducing the Gemini agent", Thomas Kurian, 8 October 2026: cloud.google.com
- Reworked, "Google debuts its all-purpose Gemini agent for work", 9 October 2026: reworked.co
- Anthropic, "Introducing Claude Haiku 5.5", 7 October 2026: anthropic.com
- ICO, "ICO secures changes from leading AI developers as scrutiny extends to AI agents", 8 October 2026: ico.org.uk
- ICO, Agentic AI call for evidence: ico.org.uk
- SAP News, "SAP Puts the Autonomous Enterprise to Work", 6 October 2026: news.sap.com
- SiliconANGLE, "SAP expands Joule into an agentic work layer as Autonomous Enterprise goes live", 6 October 2026: siliconangle.com
- Workday, "Workday's Trusted Agent Platform Now Delivers Headless Access to HR and Finance", 13 October 2026: newsroom.workday.com
- Workday, "Workday Launches Autopilot Agents That Get Real Work Done Across HR, Finance, and IT", 13 October 2026: newsroom.workday.com
- Lovable, "EU inference: control where AI processes your requests", 12 October 2026: lovable.dev
- AI Sweden, "New partnership with Genomics England on privacy-preserving AI", 7 October 2026: ai.se
- Event page (re-checked 13 October 2026): avalerions.com
Avalerion Intelligence is a weekly briefing on AI, Data & Enterprise Transformation for the Nordics, UK and Ireland, published by Avalerion Consulting AB. Subscribe by email