AI Next WaveAugust 26, 2026
Meta Planned to Cut Some Teams by 60%. Then It Canceled the Second Wave.
Executives designed an “AI-native” workforce around autonomous agents. Internal data showed more code, more incidents and uncertain productivity. Google’s legal AI preview, Australia’s music rules and a medtech cyberattack offer a steadier way to judge technology at work.
The Short Version
Meta tried to design the organization it expected AI to make possible before the agents could reliably do the work.
Reuters reviewed internal documents showing that Meta’s Project OT, short for Organization Transformation, envisioned AI handling much of the daily work performed by thousands of employees. Executives explored shrinking many teams by as much as 60% through a combination of layoffs and reassignment. Meta says the most aggressive scenarios applied to some teams, excluded several major units and never contemplated cutting 60% of the entire company.
The first wave went ahead. Meta laid off about 10% of its workforce in May and reassigned roughly 7,000 employees to AI-focused initiatives. Hours before those cuts, Mark Zuckerberg called off planning for a second restructuring expected in November.
Internal measurements help explain the retreat. AI-assisted code changes on Meta’s internal platforms and infrastructure rose 220% from a year earlier, while changes that delivered new or upgraded features to users rose 36%. Major technical and security incidents increased 40%, and staff time spent firefighting them rose 70%, according to internal posts reviewed by Reuters.
Those figures are company data, not an independent audit. Meta declined to comment on the disruption metrics and says its redeployments have begun producing useful training data. Zuckerberg separately acknowledged in July that agent development had progressed more slowly than expected and that the organizational bets had yet to produce the anticipated results.
The lesson travels well. More output can coexist with weak productivity, heavier review and lower reliability. A staffing decision needs evidence from accepted work, customer results and operating cost.
Three other developments put that test into practice. Google released a preview of legal agents built to work inside firm permissions and matter systems. Australia’s recorded-music industry drew a public boundary between AI-assisted and AI-generated songs. Boston Scientific disclosed a cyberattack that disrupted systems used to process and ship medical-device orders.
Technology changes the work only when it performs inside the surrounding system. That includes the people, controls, records, customers and recovery paths that keep the work usable.
Meta Built the Org Chart Ahead of the Capability
Project OT treated expected productivity as available capacity
Meta executives began planning Project OT in January. The concept paired smaller groups of employees with virtual workers that would take over a substantial share of routine activity. Scenario exercises explored cuts and redeployments that would reduce many teams by up to 60%, according to two people familiar with the work and documents reviewed by Reuters.
This was internal scenario planning, not a public workforce commitment. Meta acknowledged that it considered two waves and that some aggressive scenarios reached 60% for particular teams. The company disputed any suggestion that it intended to remove 60% of its full workforce.
The operating sequence still created consequences. Meta cut about 10% of its global workforce in May and shifted roughly 7,000 employees into AI-related work. Headcount in some engineering units fell as much as 30% after the layoffs and transfers, Reuters reported.
Employees were also asked to generate training data for the agents. Meta installed software on US employee devices to capture mouse movements, clicks and keystrokes. The program was later paused amid staff objections. Some reassigned engineers created software-development puzzles used to train Meta’s coding models.
The productivity numbers expose the planning flaw. Code changes increased sharply. Customer-facing delivery grew far more slowly. Internal teams also warned that unchecked agents were causing large-scale disruptions that people were unlikely to create manually.
More code is a production count. It says little about whether a release solved a customer problem, reduced cycle time or lowered cost after review and repair. The reported increase in incidents and firefighting suggests that some apparent capacity became correction work elsewhere.
Meta’s internal employee-sentiment score also fell from 74% favorable to 55% favorable. The survey does not isolate a single cause. Layoffs, transfers, surveillance concerns, unclear communication and fear of replacement all occurred during the same period.
Meta says training data from the reassigned engineering group helped produce a model released in July. That is a concrete output, although its business return remains unknown. The company continues investing heavily in AI and has moderated parts of the organizational plan without abandoning the strategy.
Leaders can avoid the same mistake by reversing the sequence. Test an agent on one production workflow. Measure accepted results, defects, review time, incidents and customer impact. Confirm that saved time becomes lower cost, added capacity or better service. Then decide whether roles, staffing or responsibilities should change.
A headcount target built into the business case can pressure a team to declare success before the evidence exists.
Google Brings Agents Into Legal Matters
The product has controls, connectors and no published ROI yet
Google launched Gemini Enterprise for Legal in preview Tuesday. It combines legal skills, agents and connections to document, research and matter-management systems.
Google says the system can support contract review and redlining, legal research, citation verification, regulatory monitoring, privacy requests and drafting. Partners include Thomson Reuters, Harvey, Legora, Everlaw, Relativity, iManage, NetDocuments and DocuSign. Cleary Gottlieb, Freshfields, Weil and Williams & Connolly are among the early law-firm collaborators.
The connections are consequential. A useful legal agent needs access to client files, prior work, firm playbooks and current legal authority. That access also creates confidentiality, privilege, permission and conflict risks.
Google says existing firm permissions and ethical walls carry into the platform. It also says client data, custom agents, playbooks and model outputs remain private to the organization and are never used to train or fine-tune Google’s foundation models.
Those are product commitments from Google. The platform is in preview, and the launch materials provide no independently measured accuracy, cost savings, adoption rate or client outcome. Pricing and the cost of integrating each firm’s systems are also absent from the public announcement.
Legal teams should begin with a narrow matter type and a reference set already reviewed by experienced lawyers. A contract-review pilot could measure material clauses found, false alarms, missed issues, citation accuracy, attorney review time and cost per accepted document. The comparison should use the same contracts and acceptance standard for the current process and the AI-assisted process.
The review burden deserves its own measure. An agent can produce a draft quickly while shifting difficult checking to a lawyer who is already carrying a full caseload. Time saved becomes useful capacity only when the firm can show where that time went and whether quality held.
Smaller firms may gain access to capabilities that once required large knowledge-management and technology teams. They also have fewer people available to configure permissions, evaluate vendors and investigate a bad output. A managed legal-AI service could help, but professional responsibility stays with the lawyer handling the matter.
Australia Draws a Line Around Human Music
AI-assisted tracks remain eligible for charts and awards
Australia’s recorded-music industry has adopted a public test for how AI-created music will be treated.
Starting with the ARIA chart dated August 31, a recording must be substantially human-made, free of chart-manipulation concerns and compliant with copyright and related laws. A lead vocal or primary instrumental performance generated by AI makes a track ineligible. Human-written and performed music can still use AI for minor backing elements, mastering, stem separation, effects and similar production work.
The same eligibility rule applies to the ARIA Awards.
The policy follows a 16-week chart run for an AI-assisted variation of Madonna’s “Like a Prayer,” which peaked at number two on ARIA’s Australian singles chart, according to the Associated Press. Its success brought an abstract industry debate into a ranking that can influence attention, bookings and careers.
ARIA’s rule avoids a blanket exclusion of AI tools. It focuses on who created and performed the central elements. It also creates an evidence problem. Artists and rights holders must accurately declare how a recording was made, and ARIA may ask for supporting material before excluding a track.
The organization provides an appeal route through its Chart and Marketing Committee and then its board. ARIA says it will contact a rights holder before making a decision, but it will not publish a list of assessed or excluded recordings.
Independent artists may carry a disproportionate documentation burden. A practical record could include session files, stems, performer credits, licenses, AI-service terms and a short description of where generative tools entered the process. That record can support chart eligibility, distribution, licensing and later copyright questions.
The rule also gives AI-music businesses a clearer product constraint. Tools that preserve provenance, permissions and human authorship may fit professional workflows better than services that deliver a finished recording with an uncertain training or rights history.
A Cyberattack Reaches Medical-Device Orders
Boston Scientific’s products and its order systems are different risks
Boston Scientific said Wednesday that a cybersecurity incident had disrupted global operations, including information systems used to process and ship customer orders.
The medical-device manufacturer detected the incident on August 25 and brought in outside cybersecurity specialists. It expects parts of the business to remain affected during recovery. The company has yet to determine the full scope, the nature of the attack or whether the incident is likely to have a material financial effect.
The disclosure concerns corporate information systems and business applications. Boston Scientific has not reported that implanted or hospital-operated devices were compromised. Treating the incident as a device-safety breach would go beyond the available evidence.
Order disruption can still reach patient care if a hospital, distributor or clinical team cannot obtain equipment on schedule. The risk depends on the specific product, available inventory, substitutes and duration of the outage. None of those effects has been quantified publicly.
Healthcare buyers should identify open orders, time-sensitive procedures, available stock and approved substitutes while keeping a manual contact route with the supplier. Manufacturers need an offline method to prioritize orders without losing product identifiers, destination records or traceability during recovery.
The useful measure is continuity, not the speed of a press statement. Track orders delayed, clinical schedules affected, manual errors, recovery time and the backlog cleared after systems return. A restored login can arrive well before normal operations.
Opportunity Radar
Independent AI productivity audits before workforce redesign
Companies are being asked to approve AI-driven staffing and operating changes while many internal measurements count prompts, licenses, generated code or hours estimated rather than completed value.
A consultant, internal audit group or workforce-analytics team could test one live workflow before a reorganization. The engagement would establish the existing cost, quality and cycle time, run the AI-assisted process against the same workload, and include review, correction, failures and downstream support.
Operations, finance, human resources and technology leaders could pay for a decision they can defend. The service must show whether capacity was genuinely released and where it went. It also needs employee participation and clear limits on monitoring. A productivity study built from surveillance data alone will create its own adoption problem.
What You Can Do With This
If you lead an AI transformation
Remove assumed headcount savings from the pilot result. Measure accepted work, customer outcomes, defects, review, incidents and capacity actually reused. Change roles after the evidence holds under normal workload.
If AI is entering your profession
Choose one recurring task with a clear standard and keep final accountability with a qualified person. Record what the system used, what it produced, what changed during review and why the final decision was accepted.
If you create with AI tools
Preserve session files, source material, performer credits, licenses and the terms of any AI service. A short provenance record can protect access to charts, clients, distributors and future licensing opportunities.
If a supplier’s systems go down
Identify time-sensitive orders, current inventory, approved substitutes and a manual contact path. Keep traceability intact while people work around the outage.
The Bigger Picture
Meta’s experience puts a number on a common failure. Activity expanded faster than useful delivery, while incidents and correction work also grew.
The response requires better operating evidence. Google’s legal platform enters preview with permissions, specialized connections and privacy commitments, but its return remains unproven. ARIA has defined which AI-assisted music can participate in its charts and what creators may need to demonstrate. Boston Scientific’s cyberattack shows how technology risk can interrupt a physical supply chain even when the affected system sits behind the product.
Each case depends on work surrounding the tool. Employees catch disruptions. Lawyers verify authority. Artists preserve provenance. Operations teams keep orders moving.
AI adoption becomes durable when an organization can show the accepted result, the human responsibility, the full cost and the recovery path. An org chart should follow that evidence.
References
Reuters: Google expands Gemini Enterprise for lawyers and law firms, August 25, 2026
ARIA: Chart eligibility rules for recordings made with generative AI, August 24, 2026
ARIA: Detailed guidance, evidence process and appeals for AI-assisted music