McKinsey’s State of AI 2026 puts organisational AI adoption at 91%, up from 78% in 2024 and 55% in 2023. Pew Research, surveying actual US workers in October 2025, found just 21% use AI at work and only 10% use it daily. That gap is the most useful frame for understanding trends in workplace automation this year.
The conversation has moved past “should we” into “how, at what cost, and with what guardrails.” The hype reel is loud. The audited deployments are quieter and more interesting, and they are where the next 12 months of decisions get made.
Most 2026 trend pieces name no companies and cite no numbers. The nine shifts below are different. Each is anchored in named deployments and real figures from McKinsey, Deloitte, the World Economic Forum, the EU AI Act, Goldman Sachs, Klarna, Kärcher, Air Canada, and Fireflies. Where there is a useful playbook, you will find one. Where the trend has a public failure case (Klarna in customer service, Air Canada on chatbot liability), it is named.
Each section closes with a “what to do this quarter” beat for your team, whether you sit in IT, ops, HR, or a small business with no automation budget. The list is ordered roughly by how quickly each shift will show up in your week.
1. Agentic AI Moves from Pilot to Production
McKinsey’s 2026 State of AI puts the headline number plainly. 23% of organisations are scaling an agentic AI system somewhere in the enterprise, and another 39% are experimenting. Nearly two thirds of large companies now have agents in production or in evaluation.
Salesforce’s framing has become the working definition. Enterprise AI has moved “beyond simple prompts and reactive text generation into a new reality where digital agents don’t just talk, they act.” Microsoft’s 2026 posture is sharper: “Companies do not want or need more AI experimentation. They need AI that delivers real business outcomes and growth.”
What changed in 2026 is that copilots evolved into agents that update CRM records, file tickets, draft and send emails, and orchestrate multi-step workflows across systems. The mainstream stack now includes Microsoft 365 Copilot agents, Salesforce Agentforce, ServiceNow AI Agents, and UiPath agentic automation. Most enterprises already pay for at least one.
Marco Argenti, Goldman Sachs CIO, gives the sober counterweight that ties the next eight trends together. “AI tools may offer superhuman productivity, but I don’t know if they can be superhumanly smart.” Productivity is the easy half. Judgment is the hard half, and judgment is where the failures happen. The 23% scaling agentic systems are quietly investing as much in escalation paths and decision logging as they are in the agents themselves, because every agent that acts also creates an audit trail somebody will eventually be asked about.
For most teams in 2026, the right move is not to deploy your first agent. It is to define the autonomous-action boundary for the agents already inside the products you pay for: which actions can the agent take without human sign-off, which require review, and which are off-limits. Until that list exists, the agent that ships in your CRM upgrade next quarter operates in a vacuum that nobody owns.
2. AI Customer Service Settles into a Hybrid Pattern
In 2025, Klarna replaced roughly 700 customer service agents with AI and the CEO publicly claimed AI matched human quality. By mid-2025, customer satisfaction had dropped 22% and the company began rehiring humans under an “Uber-style” flexible model. The new posture: AI handles routine inquiries, humans take the complex cases.
The failure mode was specific. Klarna’s AI gave “confident-but-wrong answers about policy, fees, or payment terms,” especially on disputes, fraud, and hardship. Volume scaled. Judgment did not.
That arc now anchors almost every 2026 customer-service AI decision, and the lesson has hardened into a routing pattern. Map ticket types by complexity and emotional load. Route routine repetitive types (refund status, password reset, order tracking) to AI. Route complex, regulated, or emotional cases (fraud, hardship, disputes) to humans. Measure CSAT by ticket type, not aggregate, because aggregate hides the rollback risk.
The vendor landscape has split usefully along this line. Voice AI specialists like Vapi, Deepgram, and Ringly.io handle inbound and outbound call automation, often multilingual, with sub-second latency. Mainstream platforms layer AI on top of existing service estates: Zendesk AI, Intercom Fin, Salesforce Agentforce. None of the 2026 leaders pitch full replacement anymore. Even Klarna’s revised messaging now emphasises “AI plus human” rather than the original “AI instead of human” framing from 2024.
Keep a human escalation route inside every AI conversation flow. The Klarna numbers prove what AI does well at the edges of customer service. They also prove what one wrong confident answer can do to a CSAT chart.
Best for: high-volume, low-judgment ticket types where the cost of being wrong is low and resolution is verifiable.
Skip if: your tickets cluster around disputes, hardship, fraud, or regulated outcomes, where one wrong confident answer can cost more than a year of agent salary.
3. AI Meeting Notes and Knowledge Capture Become Their Own Category
The AI note-taking market was valued at $623.5M in 2025 and is projected to reach $740.4M in 2026, on track for $3.48B by 2035 at an 18.75% CAGR. The narrower AI meeting assistants segment hit $1.6B in 2024 and is projected at $6.2B by 2033. This is no longer a feature inside a copilot. It is one of the fastest-growing automation categories in the workplace.
Fireflies anchors the high end of the market with a $1B valuation, more than 20 million users, and 75% Fortune 500 adoption. Otter, Granola, Read AI, and NoteGPT.com round out the specialist landscape, and every major suite now ships a native option (Zoom AI Companion, Microsoft 365 Copilot Recap, Google Gemini meeting summaries).
The category is growing this fast for structural reasons. McKinsey has long estimated that knowledge workers spend roughly 19% of the workweek searching for and synthesising information. Meeting AI compresses the synthesis half by turning hours of recordings into structured summaries, decisions, and action items in minutes.
The choice between a suite-native tool and a specialist usually comes down to two factors. Suite-native wins on integration and incremental cost (you already pay for Microsoft 365 or Google Workspace). Specialists tend to win on transcription accuracy, multilingual coverage, summary quality, and depth of search across past meetings. Privacy and data-residency posture varies sharply by vendor and is the most overlooked variable in procurement.
We built NoteGPT.com for the version of this problem that crosses languages and source types (audio, PDF, YouTube, web), so it tends to fit teams with multilingual meetings or heavy reading loads.
A short rollout checklist that holds up across vendors:
- Pick one tool per team for a month, not five competing bots in the same call.
- Define what “a good summary” looks like before you start (action items, decisions, full transcript, or all three).
- Audit what gets shared, where transcripts are stored, and which models the vendor trains on, before scaling beyond a pilot.
4. Shadow AI Is Now the Dominant Adoption Pattern
Microsoft’s Work Trend Index puts the BYOAI numbers high enough to reframe every IT discussion. 78% of AI users at work bring their own AI tools. At small and medium businesses, the figure is 80%. The “why” is in the same dataset: 53% of employees say they lack the time or energy for all their tasks. Most IT teams discover this only after a leak.
Workers use AI under the radar because their company has no policy, and the policy that does exist tends to lag the tools by 18 months. The risk side is concrete. Data entered into consumer LLMs may be retained, surfaced in other accounts, or used to train future models, depending on the tool and the tier. PII, customer records, source code, and confidential documents are the most common leaks, and the highest-risk category in regulated industries is also the most commonly pasted: customer data inside support tickets and sales notes.
The 2026 shift is from banning to channelling. Five steps now show up in most credible playbooks:
- Survey actual usage. Most teams underestimate by two to three times.
- Publish a short approved-tools list with one paid enterprise option per category. Chat (ChatGPT Team, Claude for Work, Microsoft 365 Copilot Chat). Notes (NoteGPT.com, Otter, Fireflies). Code (GitHub Copilot, Cursor).
- Block specific data classes (PII, customer records, legal) from consumer LLMs. Let everything else through.
- Train every team monthly for 30 minutes. Measure adoption.
- Revisit quarterly. Shadow AI shrinks when sanctioned options are equally good.
Banning shadow AI does not work. Replacing it with sanctioned options that are equally good does, and the licensing math usually favours the channelled approach once you count the breach risk. Run the survey this quarter. The gap between what you think is happening and what is actually happening will tell you the budget conversation to have next.
5. Governance Shifts from Policy Deck to Operating Model
On 14 February 2024, the British Columbia Civil Resolution Tribunal ordered Air Canada to pay Jake Moffatt $812.02 in damages after the airline’s chatbot gave incorrect information about bereavement fares. The Tribunal explicitly rejected the airline’s argument that the chatbot was “a separate legal entity that is responsible for its own actions.” Cheap damages. Expensive precedent.
The regulatory layer landed alongside it. The EU AI Act was adopted on 1 August 2024 and becomes fully applicable on 2 August 2026. Under Annex III, workplace AI used for hiring, performance review, and worker monitoring is classified as “high-risk.” Schools and employers using these tools become “deployers” under Article 29, with documentation, transparency, and human-oversight obligations. US state laws now reinforce the same direction (Colorado AI Act, NYC Local Law 144 on automated employment decisions).
That combination has pushed governance out of the policy deck and into the operating model. Five elements show up in the playbooks that survive an audit:
- Define autonomous-action boundaries for every AI tool: what can it do without human sign-off?
- Set explicit escalation paths to a named human reviewer.
- Enable model and decision logging so you can reconstruct what happened.
- Track cross-system handoffs where one AI hands off to another.
- Review every 90 days, and shorten the loop after every incident.
Dan Pitman of Redwood frames the underlying shift well. “Efficiency gains from isolated automation fade quickly. Resilience comes from orchestrating processes across domains, not optimizing tasks in isolation.”
A short pre-August 2026 checklist:
- Inventory every AI tool that touches customer-facing or employment-decision data this month.
- Confirm which fall under EU AI Act high-risk classification or NYC AEDT requirements.
- Add logging and a human review step where either is missing, before the August 2026 deadline.
6. Developer Productivity Through AI Coding Assistants
Goldman Sachs has the cleanest audited numbers in the space. In January 2025, the firm rolled out its GS AI Assistant to more than 10,000 employees, on top of GitHub Copilot already deployed to roughly 12,000 developers. Internal measurement: 20% productivity gain on routine coding, up to 55% on specific tasks. The next phase, now in motion, is “thousands of autonomous AI software engineers” projected at three to four times further gains.
The 2026 coding stack is broader than Copilot. GitHub Copilot, Cursor, Claude Code, Google Gemini for coding, and Amazon Q Developer all ship agentic features now: multi-file edits, test generation, and autonomous pull requests. The shift from “autocomplete on steroids” to “agent that opens a PR” became table stakes this year.
Where the gains actually land is more uneven than the headline suggests. AI coding assistants accelerate routine work (boilerplate, refactors, test scaffolding), help substantially on greenfield and prototyping, and offer less leverage on legacy debugging and architectural decisions. Argenti’s framing applies directly: the point is to make experienced humans faster, not to replace judgment.
Most teams overstate gains because they measure the wrong thing. Lines of code goes up. Time-to-merge, defect rate, and customer-visible incident frequency are the numbers that matter, and only teams that built measurement in from day one (Goldman among them) can speak to those credibly. Without that baseline, the productivity story becomes a survey of how good developers feel about their tools rather than a measurement of what shipped.
Best for: engineering teams with a measurable change in PR cycle time and test coverage to compare before-and-after.
Skip if: your headline metric is “Copilot adoption” rather than a business outcome. Junior engineers benefit most on routine tasks, while senior engineers benefit most on prototyping and tricky debugging.
7. Boards Demand ROI: The Investment-Impact Gap Widens
Deloitte’s 2025 enterprise AI work captures the central frustration of 2026 in three numbers. 85% of organisations increased AI investment in 2025. 91% planned further increases in 2026. Only about one in five qualify as “AI ROI leaders,” and one in four cite inadequate data infrastructure as the top barrier. Eight in ten firms report no bottom-line impact yet.
Pricing pressure compounds the problem. Forrester finds that more than four in five tech leaders expect AI features in SaaS products to push software costs up over the next year. AI is becoming a line item, not a free upgrade. Microsoft 365 Copilot at $30 per user per month, Salesforce Einstein add-ons, and Google Gemini for Workspace tiers all add to a budget that was already growing.
Boards have noticed. The new question is no longer “are we doing AI” but “did approval cycles shorten, did cost per workflow improve, or did we just buy more software.” That is harder to answer than most slide decks suggest.
ROI leaders behave differently in identifiable ways. They measure at process level, not task level. They tie each AI deployment to a specific business KPI before purchase. They treat governance and reskilling as part of the AI budget, not afterthoughts. They kill underperforming pilots within 90 days rather than letting them drift into the next budget cycle.
A quick comparison of the two camps:
| ROI leaders (~20%) | The 80% with no bottom-line impact |
|---|---|
| Define one business metric per deployment before purchase | Pick the tool first, look for metrics later |
| Measure cycle time, cost per ticket, sales reps activated | Measure tool adoption and licence count |
| Budget includes governance and reskilling from day one | Governance and training added after rollout (or never) |
| Kill pilots at 90 days if metric does not move | Pilots drift into next year’s budget |
The order of operations is the difference. Metric, then tool. Not the other way around.
8. No-Code and Low-Code Agentic Workflow Builders Go Mainstream
Kärcher, the German cleaning-equipment manufacturer, built virtual agent teams using Google Workspace Studio for feature assessment and user-story drafting. The reported result: a 90% reduction in drafting time. A real outcome from a real, named manufacturer, on a builder a non-engineer can configure.
The 2026 stack splits cleanly by company size. SMB and team-level: Zapier AI, Make, n8n, and Google Workspace Flows. Mid-market and enterprise: Microsoft Power Automate, UiPath, ServiceNow. Google’s Gemini-powered Workspace builder and Microsoft’s Copilot Studio both moved the no-code agent layer firmly into the mainstream this year.
What the new builders do, beyond classic Zapier-style triggers, is accept a multi-step workflow described in plain language, generate the steps, run them with conditional logic, and call an LLM at any point. Concrete examples now in production:
- A Workspace Flow that drafts a weekly customer email from CRM updates, then routes it to a human for sign-off.
- A Power Automate flow that triages and routes inbound support tickets by sentiment and topic.
- A Zapier AI agent that summarises new Slack messages by channel and posts a daily digest.
The boundary is real. Low-code is excellent for team-level automation. It breaks down at enterprise orchestration, where logging, identity, audit, and SLAs matter. That is where UiPath, ServiceNow, and the Power Platform earn their cost, and where a 30-day Zapier proof-of-concept will not pass an IT review.
Best for: an ops or team lead who can name the exact workflow that wastes four or more hours a week on the team.
Skip if: the workflow touches regulated data or systems of record, where the audit trail of a no-code tool will not hold up in an internal review or external audit.
9. The Reskilling Crunch: Net Positive Jobs, Distribution Problem
The honest workforce trend is neither “AI takes everything” nor “AI creates infinite jobs.” The World Economic Forum’s Future of Jobs 2025 puts the picture in numbers. By 2030: 92 million jobs displaced, 170 million created, net gain of 78 million. 22% of all jobs disrupted. 39% of workers’ core skills will change. 59% of workers (around 590 million globally) need reskilling. 11%, more than 120 million people, are unlikely to receive it.
The categories matter as much as the totals. WEF’s fastest-declining roles are cashiers, administrative assistants, and now graphic designers (the last is new in 2025, driven by generative AI). The fastest-growing roles are big data specialists, fintech engineers, and AI and machine learning specialists. AI and data processing alone create 11 million new roles and replace 9 million.
Worker sentiment lags the data. Recent surveys show 43% of workers fear automation may replace their job within two years, and worker confidence fell 18% across the same period. Both feelings are reasonable given the headline coverage, even if the net number is positive.
The WEF framing on the distribution problem is the line worth printing on the wall: net plus 78 million jobs by 2030, but only if reskilling actually happens. The 11% miss rate is the real story.
Companies handling this well share three habits. They name declining roles internally and offer paid reskilling pathways before the role is cut. They rotate high-volume routine workers into AI-supervision and AI-training roles. They treat AI literacy as both a hiring criterion and a recurring 30-minute monthly training, not a one-off course.
Two practical moves close the trend. As a worker, pick one of the WEF growth categories adjacent to your current skill (analytics, AI operations, change management, fintech) and learn it this quarter. As a leader, name the declining role on your team and put a paid reskilling pathway in front of the people in it before the headcount conversation, not after.
Frequently Asked Questions
What are the biggest workplace automation trends in 2026?
Agentic AI scaling from pilot to production, AI meeting and knowledge capture as its own category, governance moving from policy deck to operating model, BYOAI/shadow AI normalising as the default adoption pattern, AI customer service settling into a hybrid model after Klarna’s rollback, and the reskilling crunch behind WEF’s projected net 78 million new jobs by 2030.
What is agentic AI in the workplace?
Agentic AI refers to systems that take actions rather than just answer questions: updating records, creating tasks, sending messages, and running multi-step workflows. McKinsey reports 23% of organisations are scaling agentic systems and 39% are experimenting. Successful deployments pair the agent with explicit autonomy boundaries and a named human escalation path.
Will AI replace office workers by 2030?
Not on net. WEF’s Future of Jobs 2025 projects 92 million jobs displaced and 170 million created by 2030, a net gain of 78 million. Displacement concentrates in cashier, administrative-assistant, and (newly in 2025) graphic-designer roles. Growth concentrates in big data, AI/ML, and fintech. 11% of workers are unlikely to receive any reskilling.
Are companies liable for what their AI chatbot says?
Yes. The 2024 Moffatt v. Air Canada ruling at the BC Civil Resolution Tribunal confirmed companies remain liable for chatbot misinformation. Air Canada was ordered to pay $812.02 and the “separate entity” defence was rejected. EU AI Act deployer obligations from August 2026 reinforce the same principle.
Should we let employees use ChatGPT and other consumer AI at work?
They already do. Microsoft Work Trend Index shows 78% of AI users bring their own tools to work, 80% at SMBs. The fix is a short approved-tools list with one paid enterprise option per category, data-class restrictions on PII, and 30 minutes of monthly training. Banning does not work.
Why are so many workplace AI investments not delivering ROI?
Deloitte found 85% of organisations increased AI investment but only one in five qualify as ROI leaders, and 80% see no bottom-line impact. Common causes: AI deployed in isolation, governance in policy decks rather than workflows, and success measured at task level rather than process level.
