Intelligence Augmentation Weekly Review 2026-05-18

Week In Review

This week’s intelligence-augmentation news clustered around a single argument: the platforms that mediate everyday cognitive work are being rebuilt around AI agents, and the companies setting the terms are racing to lock in distribution before the next product cycle. On the consumer side, Google races to put Gemini at the center of Android ahead of next year’s Apple AI reboot, and the upcoming Google I/O 2026 keynote is expected to position a new Gemini agent and a refreshed Project Astra against OpenAI’s GPT-5.5 and Anthropic’s Mythos. On the enterprise side, Microsoft’s two weeks of Frontier-suite announcements continued to ripple — Microsoft Agent 365 reached general availability on May 1, Copilot Cowork went mobile with reusable skills, and Microsoft published a new “human agency” framing for how Copilot is supposed to interact with knowledge workers as agents proliferate.

A second theme was that the agent layer is starting to grow its own ecosystem. Norm Ai’s compliance agent for Microsoft 365 Copilot — an early third-party plug-in for regulated industries — is the kind of release that only makes sense once a control plane exists for it to sit on. OpenAI granting EU vetted teams access to GPT-5.5-Cyber suggests that even sensitive frontier deployments are now being routed through region-specific partner channels rather than the global API. Read together, these stories show the agent stack hardening into a multi-vendor market with platform controllers, app vendors, and managed deployments — much earlier than most observers expected.

On the neurotechnology side, Naveen Rao’s Neurotech Notables #54 used the May 1–15 window to underscore how lopsided the field has become: brain-computer interface companies have captured roughly 76 percent of the $2.27 billion raised across all neurotechnology over the past twelve months. Two specific items help illustrate where that money is going. SonoMind closed a €20M Series A to take focused ultrasound deep into the brain without surgery for treatment-resistant depression, and Prophetic publicly priced its Dual and Phase ultrasonic dream-shaping headbands for late-2026 and 2027 shipping. Together they map out a spectrum that runs from clinical neuromodulation at one end to consumer cognitive wearables at the other, with focused ultrasound emerging as the connective tissue.

The combined picture is that intelligence augmentation is becoming less of a single technology trend and more of a layered stack: agent-enabled productivity at the application layer, frontier model deployment governed regionally underneath, and a maturing neurotechnology supply chain pushing ultrasound, EEG, and implantable BCIs deeper into the workflow on the other end. The week’s news points to a more pluralistic, more regulated, and more clinically credible 2026 than the field looked like even six months ago.

Items

Google races to put Gemini at the center of Android

Google is restructuring Android around its Gemini model in advance of Apple’s expected 2026 AI overhaul, according to CNBC reporting on May 12. The push touches the Android launcher, the assistant layer, on-device inference, and the way third-party apps surface AI features, and is being driven from inside Alphabet by the same teams responsible for Gemini and the Pixel platform.

The strategic logic is that Apple’s installed base is a once-a-decade opening: if Google can convince users and OEMs that Gemini is the default cognitive layer of Android before Apple ships its own redesigned intelligence stack, it locks in the assistant layer for the next platform cycle. Distribution, not raw model quality, is the prize.

For intelligence augmentation specifically, the article matters because it signals that the assistant becomes a system service rather than an app — accessible from the launcher, the keyboard, and any running application without explicit invocation. That changes the unit of analysis from “are users using ChatGPT” to “are users always partially using an assistant.”

Source: CNBC

OpenAI grants EU vetted teams access to GPT-5.5-Cyber

CNBC reported on May 11 that OpenAI is rolling out GPT-5.5-Cyber — a variant of its latest frontier model tuned for cybersecurity workflows — to vetted security teams across the European Union in a limited preview, with Anthropic’s Mythos notably absent from a similar arrangement so far.

The cyber-tuned variant is interesting independent of the regional politics. By specializing GPT-5.5 for tasks like vulnerability triage, threat-hunt augmentation, and defensive code review, OpenAI is signaling that the way to make frontier models useful in enterprises is not just bigger context windows but role-specific behavior conditioning, distributed through audited channels.

The EU-only framing is also a milestone. It is the first public confirmation that the major frontier labs are willing to ship region-segmented versions of their flagship models to satisfy local regulation and procurement requirements. For intelligence augmentation in regulated industries, that segmentation is what makes deployment feasible at all.

Source: CNBC

Norm Ai launches compliance agent for Microsoft 365 Copilot

Norm Ai announced on May 12 the general availability of a compliance agent that runs inside Microsoft 365 Copilot, offering policy intelligence, real-time compliance review, and auditability features aimed at financial-services, healthcare, and other heavily regulated enterprises. According to the company’s release, the agent integrates with Copilot’s chat and document workflows so that policy checks happen during composition rather than at end-of-process review.

The product itself is incremental — there are many compliance-as-a-service vendors — but the distribution model is novel. By shipping as a first-class agent inside Microsoft’s just-launched Agent 365 control plane, Norm Ai gets access to enterprise tenants without each customer needing a separate procurement cycle. That mirrors the early SaaS marketplace pattern that Salesforce’s AppExchange pioneered two decades ago.

For intelligence augmentation, the lesson is that agents will not stay monolithic. The same workflow surface that handles drafting will route specialized agents in and out — a compliance reviewer here, a finance approver there — and end users will rarely notice the handoffs.

Source: PR Newswire

Google I/O 2026 keynote opens with new Gemini, Project Astra refresh

Tech Times reported on May 17 that Google’s I/O 2026 keynote, set for May 19, would unveil a new Gemini model that the company positions as roughly comparable to OpenAI’s GPT-5.5 but meaningfully short of Anthropic’s Mythos, alongside a substantial refresh of Project Astra — Google DeepMind’s universal-agent prototype that aims to act across applications and devices on the user’s behalf.

If the early reporting holds, the Astra demonstration will be the more consequential announcement. A 24/7 personal agent that can read screens, reason across applications, and take actions on the user’s behalf is the long-anticipated successor to the chat interface that has defined the first generation of consumer AI. For intelligence augmentation, that shift moves the locus of value from “I asked the model a question” to “the agent already did the thing.”

The competitive subtext is also worth noting. Google’s positioning of the new Gemini below Mythos and roughly even with GPT-5.5 is unusual public candor for a flagship release. It implies that Google is willing to compete on integration depth and platform reach rather than raw capability benchmarks — a bet that fits Android’s distribution advantage.

Source: Tech Times

Microsoft Copilot Cowork goes mobile with reusable skills

Microsoft announced on May 5 that Copilot Cowork, its agentic multi-step orchestration layer, is now available on iOS and Android, with a new “Skills” system that lets organizations capture reusable instruction sets for recurring tasks. The release also expands native integrations to include Power BI’s Fabric IQ, multiple Dynamics 365 applications, and third-party tools including LSEG, Miro, and monday.com.

The shift to mobile is more substantive than it sounds. Cowork is designed to run long workflows asynchronously — drafting a report, gathering data across systems, preparing a deck — and mobile delegation lets users hand work off during dead time (commutes, meetings, evening hours) and pick up results later. That decouples cognitive supervision from desk time in a way that earlier copilots could not.

The Skills system is the more important architectural change. By giving organizations a way to encode “how we do this kind of work here,” Microsoft is creating a corpus of company-specific procedural knowledge that the agent layer can consult — closer to apprenticeship than autocomplete.

Source: Microsoft 365 Blog

Microsoft Agent 365 enters general availability

Microsoft Agent 365 — the management and governance plane for AI agents inside Microsoft 365 — entered general availability on May 1 at $15 per user per month, according to the Microsoft Security Blog. Agent 365 gives IT a single place to observe, govern, and secure every agent that runs in an organization, regardless of which Microsoft or third-party product the agent originates from.

The product addresses what may be the practical bottleneck on enterprise AI adoption: not whether agents work, but whether they can be permissioned, audited, and shut down. By offering agent-level identity, role-based access, lifecycle controls, and observability, Microsoft is treating agents as a new class of corporate user that has to live inside the same governance framework as employees and service accounts.

For intelligence augmentation, the launch matters because it marks the point at which “deploying an agent” inside a regulated enterprise stops looking like a research project and starts looking like provisioning. That is the threshold at which deployment volumes typically take off.

Source: Microsoft Security Blog

Prophetic prices Dual and Phase ultrasonic lucid-dreaming headbands

Prophetic, a neurotech startup focused on sleep and dreaming, published pricing and ship dates for two products: Dual at $449 (first units shipping end of 2026) and Phase at $1,299 (shipping mid-to-late 2027). Both headbands deliver focused ultrasonic energy to the prefrontal cortex during REM sleep, with the goal of activating the frontoparietal network — a brain system associated with self-awareness and cognitive control that is typically quiet during dreaming — to induce lucid dreams.

The technical premise is that targeted ultrasonic neuromodulation, until recently a laboratory-only technique, can now be miniaturized into a wearable form factor. The Phase model improves on Dual primarily through spatial steering of the ultrasound, which Prophetic claims gives finer control over which prefrontal regions are activated and therefore over dream content.

Independent scientific validation of consistent lucid-dream induction in consumer settings is still limited, and Prophetic itself frames the products as part of an ongoing research program rather than a clinically validated therapy. Even so, the launch is notable as one of the first consumer-priced products built around focused ultrasound for cognitive purposes — a delivery mechanism that, until 2024, lived almost exclusively in research labs.

Source: Prophetic

Microsoft reframes Copilot around “human agency” in the age of agents

In a companion post on May 5, Microsoft’s Microsoft 365 blog argued that as agents proliferate, the central design question shifts from “how do we get the AI to do more” to “how do we preserve human agency in the work.” The post outlines a set of principles — visible delegation, reversible actions, explicit confirmation for consequential decisions, and human-set goals — that Microsoft says will govern Copilot and Agent 365 going forward.

The framing is partly defensive against accusations that AI in the workplace deskills workers, but it also reflects a substantive design choice. By treating Copilot as a delegate the user supervises rather than a system that acts on the user’s behalf by default, Microsoft is taking a different stance than vendors pushing “fully autonomous” agents — and creating a posture that maps better to regulated industries’ supervisory requirements.

For intelligence augmentation, the post is worth reading as the first major vendor articulation of where the human stays in the loop once agents are competent enough to act unsupervised. The question is no longer whether they can, but whether and when they should.

Source: Microsoft 365 Blog

Neurotech Notables #54: brain-computer interfaces capture 76% of recent neurotech funding

Naveen Rao’s Neurotech Notables newsletter, in its May 1–15 roundup published May 18, reported that brain-computer interface companies have captured roughly $1.73 billion of the $2.275 billion raised across all neurotechnology over the past twelve months — about 76 percent of disclosed capital. North America accounted for $1.899 billion of that total across 16 deals.

The lopsided distribution reflects a few large concentration points: Neuralink’s $650 million Series E in April, Synchron’s $200 million Series D, and recent rounds for Merge Labs and Science Corporation. The roundup also flagged smaller but strategically interesting events from the same window: BlueWind Medical’s $47.8 million financing, Galvani Bioelectronics’ shutdown, and the EU MDR approval of PlatoCare’s wireless tDCS device for depression.

What stands out is the maturity gap. The BCI category is now investible at venture-capital scale despite a regulatory pathway that most participants concede stretches to 2028–2030 for commercial approval. That is unusual: capital is flowing well ahead of revenue, in part because the strategic value of an early BCI position is being priced by analogy to early platform plays in mobile and the cloud.

Source: Neurotech Notables (Substack)

SonoMind raises €20M to deliver focused ultrasound for resistant depression

SonoMind, a Paris-based neurotech spinout from Inserm and ESPCI Paris-PSL, closed a €20 million ($23.5 million) Series A led by Critical Path Ventures and Bpifrance to bring its focused-ultrasound neuromodulation platform into pivotal trials for treatment-resistant depression. The technology uses a custom acoustic lens to direct low-intensity ultrasound into deep brain structures without surgery or implants, delivered in outpatient sessions.

The clinical foundation matters. Preliminary results from Sainte-Anne Hospital in Paris, published in 2025 in the journal Brain Stimulation, reported that the approach produced more than a 60 percent reduction in symptom severity after five consecutive days of treatment, with no significant adverse events. If those results hold in a randomized pivotal study, focused ultrasound becomes a competitive option in a treatment space currently anchored by ketamine, TMS, and electroconvulsive therapy.

The company’s stated market authorization target is 2029, with anxiety, addiction, Parkinson’s disease, essential tremor, post-stroke recovery, and chronic pain as eventual indications. For intelligence augmentation specifically, SonoMind is part of a broader bet that the next decade of cognitive intervention will increasingly come from non-invasive, precisely targeted neuromodulation rather than pharmacology or implants.

Source: Mass Device