Ten years of AI at work reveal just how much the ordinary working day has changed

In March 2016, Google’s AlphaGo defeated professional Go player Lee Sedol by four games to one. The achievement made international headlines because a computer had mastered a task many researchers believed would remain beyond machines for another decade. Yet most office workers could not ask that same artificial intelligence to prepare a client proposal, examine a spreadsheet or rewrite a difficult email.

Ten years later, those activities are routine applications of generative AI. A marketing professional can provide campaign results and request an analysis. A developer can assign a software issue to a coding agent. A manager can ask an assistant to extract decisions and responsibilities from a recorded meeting. Some AI systems can even navigate websites and operate software interfaces.

The fundamental change is that AI has moved from performing isolated, predefined tasks toward participating in connected work processes. In 2016, people usually adapted their activities to the limitations of AI software. In 2026, they increasingly describe an intended outcome and ask a system to determine how to produce it.

That does not mean AI can reliably replace the person responsible for the result. Its remarkable technical progress has introduced new problems involving accuracy, accountability, information security and the verification of machine-generated work.

The artificial intelligence of 2016 was already useful, but narrowly specialised

Artificial intelligence did not begin with ChatGPT. By 2016, machine learning had become an important component of search engines, recommendation systems, fraud detection, digital advertising, image recognition and speech processing. Businesses could buy access to specialised AI capabilities through cloud APIs or build custom models with substantial technical expertise.

Google’s Cloud Vision API, released in public beta in February 2016, could identify objects in photographs, recognise logos and extract printed text. Google’s Smart Reply feature suggested up to three short responses to incoming messages. Neural machine translation was beginning to improve the quality of automated language translation.

These systems addressed genuine business problems. Their limitations were principally about scope and flexibility. A model trained to recognise images could not automatically become a competent research assistant, financial analyst or website developer. Combining separate AI functions into an operational process generally required programming, integration work and conventional business software.

A small company might have used machine learning without even recognising it as AI. An online retailer could benefit from better product recommendations, while its employees still wrote descriptions, manually checked orders and prepared reports. Much of the intelligence was embedded inside the application rather than exposed as a conversational assistant.

That distinction explains why public experiences of AI changed so dramatically after the introduction of ChatGPT on 30 November 2022. Natural language became a practical interface for instructing general-purpose models, removing some of the technical barriers that had previously limited access.

Email changed from suggested sentences to understanding an entire conversation

A revealing example comes from Google’s Smart Reply. In March 2016, Google reported that the feature was already involved in 10% of mobile replies in Inbox by Gmail. Its purpose was straightforward: suggest a short response that someone could select, edit and send.

Imagine receiving an email from a customer asking whether a proposed meeting time is acceptable.

In 2016, the intelligent feature might offer three responses resembling “Sounds good”, “Thanks” or “I’ll get back to you”. That was useful because it eliminated a few keystrokes.

In 2026, a suitably connected AI assistant may be able to examine a longer exchange, identify unresolved requirements, summarise the customer’s concerns, draft a substantive answer and suggest suitable follow-up actions.

Consider an agency dealing with a client who has exchanged 18 emails about a delayed website launch. Instead of reading every message sequentially, the account manager can ask the assistant to identify the latest approved deadline, explain the disagreement over additional development work and draft a reply that separates contractual commitments from new requests.

The difference is substantial. The old system suggested language; the newer one can organise the information needed to decide what to communicate. However, the employee must still verify contractual facts, the status of work and any commitments made in the drafted response. A convincing summary is not evidence that every detail is correct.

Translation now involves context, terminology and entire documents

Machine translation was one of AI’s most visible commercial successes before the generative AI boom. In September 2016, Google announced its neural machine translation system, designed to improve translation quality compared with its previous phrase-based technology.

For someone translating a product description, the process was largely transactional. Enter source text, choose the languages, obtain a translated version and correct mistakes manually.

Modern AI systems can work with significantly more context. A user can supply a product catalogue, a terminology list, the intended audience and instructions concerning tone. The model can then translate text while attempting to preserve the distinctions between technical language, advertising copy and contractual terminology.

Consider a Slovak manufacturer preparing a catalogue for customers in Germany and the Czech Republic. Instead of translating isolated sentences, a team can provide the entire catalogue together with approved product names, measurement conventions and examples of existing corporate language.

The assistant can identify inconsistent terminology across descriptions and produce versions intended for different channels. One version might suit a technical specification sheet; another might be structured for an online product listing.

The improvement is not simply better translation. It is the ability to process language as part of a larger publishing assignment. Specialist terminology, legal meaning, numerical values and product claims still require verification. A model that writes fluent German can nevertheless mistranslate a technical requirement or inadvertently strengthen a marketing claim.

Marketing has moved from automated targeting to campaign production

In 2016, advertising platforms already used machine learning extensively. Algorithms helped determine which advertisements to display, how to rank them and which audiences were likely to respond. However, the marketer remained responsible for writing copy, preparing visuals, developing landing pages and interpreting most campaign reports.

A typical campaign required separate specialists or separate software tools. Copywriters created headlines, designers prepared banners, analysts examined results and campaign managers adjusted budgets.

Today, generative AI can contribute to nearly every one of those activities. A marketer can provide a campaign brief, customer research, product information and existing brand guidelines, then request alternative messages, creative concepts, landing-page structures and analytical interpretations.

Consider a company selling insulation materials. A campaign might need to explain thermal performance to homeowners, practical installation requirements to contractors and purchasing considerations to distributors.

In 2016, producing those three sets of materials meant manually creating content for separate customer groups. In 2026, an AI-assisted workflow can create initial versions of all three, propose variations and identify information that requires technical verification.

AI has dramatically reduced the effort involved in producing a first draft. It has not eliminated the need to know whether that draft is accurate, commercially appropriate or persuasive. A model may write convincing claims about energy savings without having evidence for a specific product or building.

For an agency such as Webiano, the practical opportunity lies in connecting content creation with research, search visibility, web development, distribution and performance measurement. Producing more text is relatively easy. Building a reliable process that delivers useful information to the right audience remains considerably harder.

Search engine optimisation has become a research and editorial challenge

Ten years ago, much SEO work concentrated on keyword research, technical optimisation, metadata, content planning, backlinks and interpreting search engine results. Those activities remain relevant, but generative AI has changed both production and discovery.

A content specialist can now examine hundreds of exported search queries, classify them according to user intent, identify overlapping article topics and suggest a more coherent editorial structure. AI may also assist with schema markup, internal-link planning and the analysis of technical reports.

A practical example is a publisher maintaining 500 articles about construction materials. Instead of manually reviewing every headline and spreadsheet row, the editor can give an assistant a content inventory and ask it to flag duplicates, outdated subjects and missing explanations.

The assistant might recommend consolidating several near-identical articles or creating an explanatory guide where multiple pages answer only fragments of the same question.

Yet publishing hundreds of automatically generated articles does not establish editorial authority. AI-generated content can reproduce inaccuracies, obscure original sources and create pages that appear comprehensive without offering independent value.

The central SEO responsibility has therefore become more demanding: decide which information deserves publication, verify it against authoritative material and make it useful enough that readers have a reason to trust it.

Image editing changed from recognising objects to creating them

Image recognition was an established AI application in 2016. Google’s Cloud Vision API could detect objects, read text and classify photographs. That enabled functions such as automatic tagging and photo-library organisation.

Generative image editing added a different capability. In May 2023, Adobe introduced Generative Fill in Photoshop, allowing users to add, remove and extend image content with text instructions.

A graphic designer preparing an advertisement can now work with an existing photograph, remove distracting objects, extend the background to a different aspect ratio or create visual alternatives. What previously required considerable manual editing can sometimes be attempted within seconds.

Imagine a furniture retailer with one photograph of a chair on a white background. The creative team can use generative tools to explore different settings for the product, such as a Scandinavian interior or a modern office.

That does not establish that the resulting setting, material or product details are authentic. A generated image can accidentally change the product being advertised, including its shape, proportions or components.

Professional review becomes particularly important for catalogues, real estate, engineering, healthcare and other uses where visual accuracy matters.

The distinction between representing an existing reality and generating a plausible alternative is now a basic part of responsible commercial design.

Video production is no longer limited to editing captured footage

AI-assisted video processing existed long before 2026, but its applications were generally narrower than today’s generative systems. Teams still depended heavily on recorded footage, conventional editing timelines, manual graphics and professionally produced narration.

Modern AI software can assist with transcription, subtitle preparation, speech processing, translation, storyboard development and the generation of visual sequences.

In May 2025, Google introduced Veo 3, a video-generation model capable of producing video with audio. Such tools can generate material that does not originate from a traditional camera recording.

For a small business, the practical difference is significant. A promotional campaign that requires several visual concepts may begin with generated storyboards or short video experiments before the company commissions a full shoot.

A training department can also use AI to organise a recorded tutorial, extract important sections, prepare subtitles and generate accompanying written instructions.

However, high production speed creates new responsibilities. Synthetic footage should not be presented as documentary evidence of an event that never occurred. Generated product demonstrations must not misrepresent performance, appearance or safety characteristics.

AI changes the economics of video production, but the distinction between authentic footage and synthetic media remains commercially and ethically important.

The practical comparison between 2016 and 2026

Table 1. Everyday work then and now

These are representative capabilities, not guarantees that every product or subscription provides them. Actual results depend on software integration, permissions, available information and human supervision.

Programming moved from code suggestions toward delegated development

Software development provides one of the clearest examples of AI’s evolution.

In 2016, developers already used sophisticated code completion, static analysis and automated testing. Machine-learning research was exploring software-related applications, but conversational general-purpose coding assistants were not yet standard development tools.

GitHub Copilot’s technical preview in June 2021 marked a visible transition. It introduced AI assistance capable of suggesting complete lines and functions while developers worked.

By May 2025, GitHub had introduced a coding agent capable of receiving an assigned issue, working in an environment powered by GitHub Actions and submitting changes as a pull request.

Consider a developer maintaining a large WordPress website. A customer reports that a page has stopped displaying after a plugin update.

An AI-assisted development process may involve examining relevant source files, reviewing error messages, proposing a code change, producing a test and preparing the result for review.

That is a different kind of interaction from asking an autocomplete system to finish a line of PHP.

The developer increasingly supervises the solution rather than manually producing every part of it. But supervision is not optional. Generated changes can introduce security vulnerabilities, break compatibility or misunderstand the architecture of an existing project.

A 2025 study by METR found that experienced open-source developers working in familiar repositories took 19% longer to complete assigned tasks when using the AI tools tested. Participants nevertheless believed the tools had made them faster.

The study does not establish that AI always slows programming. It demonstrates why measured productivity is more reliable than subjective impressions of speed.

Server administration illustrates the difference between advice and execution

System administration offers another practical comparison.

A Linux administrator investigating a failed service in 2016 might search documentation, inspect logs, examine configuration files and execute commands manually. Automated monitoring already existed, but troubleshooting frequently depended on the administrator connecting information from several sources.

In 2026, a capable AI assistant can explain errors, compare configuration files, interpret log entries and help prepare diagnostic commands. Where suitable tools and permissions are available, an agent may execute parts of that investigation.

Suppose a Debian server running Nginx and PHP-FPM begins returning HTTP 504 errors.

An administrator can provide the assistant with selected service logs, resource measurements, process information and relevant configuration. The assistant can investigate whether available evidence points toward exhausted PHP-FPM workers, long-running requests, database contention or another cause.

It can then suggest which measurement would distinguish those explanations.

The important improvement is the ability to reason across several technical observations. Nevertheless, an apparently confident diagnosis is not proof that the suggested configuration change is safe.

Increasing pm.max_children, for example, may worsen memory pressure if the server lacks sufficient RAM. An assistant that modifies a production system without adequate checks can turn a temporary performance problem into an extended outage.

The safest arrangement separates observation, diagnosis, proposed changes, testing and authorised execution. Backups, rollback procedures and change records remain essential.

Meetings can now produce structured work instead of just recordings

In 2016, organisations could record conferences, store documents and use transcription services. Turning a lengthy meeting into a reliable list of decisions still commonly required someone to take notes or review the recording.

Microsoft’s documentation for Copilot in Teams now describes the ability to summarise meeting discussions, identify speakers, suggest action items and answer questions about the conversation, subject to the relevant settings and access.

Consider a 70-minute meeting covering a website redesign, advertising expenditure and the release date of a product catalogue.

A manager could ask an AI assistant to organise the discussion into decisions, unresolved questions and proposed responsibilities. The resulting record might identify that the designer needs revised images, the developer needs approved specifications and the client still has to confirm a campaign budget.

In this case, the assistant’s value is not its ability to transcribe words. It is its ability to turn conversational information into an initial operational record.

There is a critical limitation. If meeting participants never agreed on a deadline, the AI should not manufacture one simply to make the resulting action list look complete. The person responsible for the meeting must distinguish confirmed decisions from suggestions and assumptions.

Research has shifted from finding documents to examining collections of evidence

Research workflows have changed considerably, particularly when people work with large collections of material.

In 2016, a researcher might use search engines, digital libraries, reference managers and specialist analytical software. These were powerful resources, but synthesising evidence across documents remained a labour-intensive activity.

Modern AI tools can examine uploaded reports, compare sections, identify recurring themes and draft structured explanations. Google’s NotebookLM, for example, introduced Audio Overviews in September 2024 and later expanded supported source types to include audio and public YouTube material.

Imagine a business analyst reviewing 25 market reports, several interview transcripts and internal documents. Rather than beginning with an empty document, the analyst can ask an AI system to identify agreement and disagreement among those sources.

It may find that several reports forecast industry growth while others warn about declining profitability. That discrepancy is more useful than a polished summary pretending that all sources support the same conclusion.

The essential skill is now verifying the relationship between a claim and its underlying evidence. A system that generates citations or source references must still be checked. Incorrect attribution can be more dangerous than an obvious factual mistake because it gives unsupported claims the appearance of authority.

Customer service provides measurable evidence of productivity improvements

Customer support demonstrates that the practical benefits of AI are not limited to spectacular demonstrations.

Researchers Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined the introduction of a generative AI assistant among 5,179 customer-support agents. Their research found an average productivity increase of 14%, measured by issues resolved per hour. Less experienced and lower-skilled workers experienced particularly large improvements.

This matters because customer service depends heavily on retrieving information, interpreting customer requests and communicating clearly.

In 2016, a support system could classify requests, direct customers through scripted menus or retrieve predefined answers. Complex inquiries frequently required an employee to search documentation and reconstruct the situation manually.

A generative assistant can now suggest a response based on the customer’s question and available knowledge material. An employee can review the recommendation, adapt its wording and decide whether to send it.

For example, a hosting provider dealing with a customer’s email-delivery problem may use AI to compare the reported symptoms with documented troubleshooting procedures and prepare diagnostic questions.

The assistant can reduce repetitive work without taking responsibility for the final technical conclusion.

The research supports productivity gains in the setting studied, not a universal percentage improvement for every company. Performance depends on employee experience, the quality of the knowledge base, system integration and the nature of customer requests.

Autonomous agents represent a more consequential transition

The emergence of AI agents is among the most important developments separating 2016 from 2026.

A conventional chatbot produces an answer. An agent may be equipped with tools that allow it to inspect information, choose a sequence of actions and work toward a defined outcome.

In July 2025, OpenAI introduced ChatGPT agent with capabilities combining research, a visual browser and other tools. Google introduced a Gemini Computer Use model in October 2025 to support agents interacting with software interfaces.

Consider a business owner preparing a competitor analysis.

A conversational assistant might produce a research plan and suggest what information to collect. An appropriately equipped agent may navigate supported websites, gather publicly available material, organise findings and prepare a spreadsheet or presentation.

In another example, a company might connect an agent to authorised business applications so that it can identify invoices requiring review, draft follow-up messages and prepare an accounting summary.

The defining difference is the progression from recommending actions to performing authorised actions. This introduces operational risks that were less prominent when AI primarily generated text.

An agent with access to company systems can affect records, communications and external services. Organisations therefore need explicit permissions, audit trails, checks before consequential actions and clear responsibility for mistakes.

An AI agent is not equivalent to an autonomous employee who understands every business obligation. It remains software operating within imperfect models, permissions and technical constraints.

Faster work does not always mean better work

Table 2. What research has actually measured

These findings should not be combined into a single productivity estimate. They describe different kinds of work, measurement methods and technological conditions.

Stanford’s 2026 AI Index also reported substantial progress in computer-use benchmarks, while noting that even strong agents continued to fail a meaningful share of structured tasks. Technical capability and operational reliability are not interchangeable.

The biggest new responsibility is checking work that looks finished

Perhaps the most misleading aspect of modern AI is the professional appearance of its output.

A generated report may contain attractive charts, well-structured paragraphs and confident recommendations. A program may compile successfully. An image may look indistinguishable from a photograph. None of those characteristics guarantees correctness.

The US National Institute of Standards and Technology published its Generative AI Profile in 2024 as a resource for organisations managing risks associated with generative systems.

For businesses, verification needs to reflect the consequences of error. A draft social media caption may require an editorial check. A financial calculation needs reconciliation against the underlying records. A legal document requires professional examination. A production-server change requires technical testing, access controls and a recovery procedure.

There is also a new information-security concern. When employees upload confidential material or grant agents access to external systems, they may expose information beyond its intended audience if services are configured incorrectly or permissions are too broad.

European regulation has developed alongside these capabilities. The EU AI Act applies progressively, with obligations concerning general-purpose AI, transparency and particular high-risk uses following different implementation dates.

AI literacy is consequently becoming a matter of understanding limitations, permissions and accountability, not merely knowing how to write a prompt.

The economic shift is from producing individual outputs to managing entire processes

The 2016 approach to AI was largely defined by specialised features. A company purchased an application to solve a recognisable problem: classify an image, predict customer behaviour, translate a sentence or improve search results.

In 2026, the commercial proposition is frequently broader. AI may participate in the flow of information between research, planning, production, analysis and execution.

Consider a small digital agency with responsibility for websites, editorial content, advertising and technical maintenance. Ten years ago, each activity might have required its own production process, specialist software and considerable manual coordination.

Today, AI can assist with preparing research, drafting articles, examining campaign exports, producing creative variants, analysing logs and documenting technical changes. The greatest business gains may come from reducing the delays and repetitive handovers between those activities.

This changes what employers and clients should measure. The number of generated articles, advertisements or lines of code is an incomplete measure of value. Better indicators include the accuracy of completed work, the time required for verification, the frequency of rework, customer outcomes and the reliability of the resulting processes.

For smaller organisations, the ability to handle more specialised activities without expanding every department is attractive. But a company that automates poor decisions or publishes unchecked information can also produce mistakes at a much greater rate.

The responsibility for deciding what deserves to be done has become more important precisely because producing an output has become easier.

Ten years have changed the meaning of working with artificial intelligence

A professional using AI in 2016 generally interacted with a feature designed to perform a particular operation. A professional using AI in 2026 may interact with a system capable of processing documents, interpreting instructions, generating several kinds of content and performing sequences of actions.

The difference is visible in ordinary work. Short email suggestions have developed into contextual correspondence assistance. Translation tools can participate in multilingual publishing. Image recognition has been joined by image generation. Code completion has developed into agents that propose software changes. Meeting recordings can become structured summaries and action lists.

Yet the evidence also reveals limits. Productivity gains differ across tasks and employees. Experienced developers can lose time reviewing AI-generated changes. Agents remain fallible. Generated information can appear authoritative even when its underlying claims are wrong.

The most successful organisations will not necessarily be those that purchase the largest number of AI subscriptions. They will be those that identify suitable tasks, provide reliable information, measure actual outcomes and retain appropriate human control.

In 2016, the central question was often whether a machine could perform a particular task. In 2026, the more demanding question is whether an organisation can trust the result, explain how it was produced and remain accountable for what happens next.

Author:
Jan Bielik
CEO & Founder of Webiano Digital & Marketing Agency

This article is an original analysis supported by the sources cited below

AlphaGo’s ultimate challenge: a five-game match against the legendary Lee Sedol
Google’s 2016 account documents the five-game competition and AlphaGo’s 4–1 victory over Lee Sedol.

Smart Reply comes to Inbox by Gmail on the web
Google’s March 2016 announcement explains short automated reply suggestions and reports their adoption in mobile email responses.

A Neural Network for Machine Translation, at Production Scale
Google Research describes its September 2016 transition toward neural machine translation.

Google Cloud Vision API enters Beta, open to all to try!
Google’s February 2016 announcement documents object recognition, optical character recognition and logo detection.

Introducing ChatGPT
OpenAI’s original November 2022 announcement establishes the launch and conversational capabilities of ChatGPT.

Introducing GitHub Copilot: your AI pair programmer
GitHub’s June 2021 announcement describes its initial coding assistant and code-completion capabilities.

Hello GPT-4o
OpenAI’s May 2024 announcement documents advances in multimodal interaction involving text, audio and visual information.

Adobe Unveils Future of Creative Cloud With Generative AI as a Creative Co-Pilot in Photoshop
Adobe’s May 2023 announcement establishes the introduction of Generative Fill and its image-editing functions.

NotebookLM now lets you listen to a conversation about your sources
Google’s September 2024 article explains source-based research assistance and Audio Overviews.

NotebookLM adds audio and YouTube support, plus easier sharing of Audio Overviews
Google’s September 2024 update documents additional source formats and research-related capabilities.

Catch up on meetings with Microsoft Copilot in Teams
Microsoft’s product documentation describes meeting summaries, discussion analysis, action items and relevant operational requirements.

GitHub Copilot: Meet the new coding agent
GitHub’s May 2025 announcement describes assigning development tasks to an agent and reviewing proposed code changes.

Introducing ChatGPT agent
OpenAI’s July 2025 announcement documents the introduction of agent capabilities combining research and actions performed through tools.

Introducing the Gemini 2.5 Computer Use model
Google’s October 2025 announcement describes AI agents interacting with software user interfaces.

Fuel your creativity with new generative media models and tools
Google’s May 2025 announcement introduces Veo 3 and other generative media capabilities.

Experimental evidence on the productivity effects of generative artificial intelligence
The 2023 research by Shakked Noy and Whitney Zhang reports measured effects on completion time and evaluated quality in professional writing tasks.

Generative AI at Work
Research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examines productivity outcomes among 5,179 customer-support agents.

Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity
The METR study documents a controlled experiment in which experienced developers took longer to finish assigned tasks when using the AI tools tested.

The 2026 AI Index Report
Stanford University’s 2026 report provides evidence concerning AI adoption, technical performance, economic effects and remaining limitations.

Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile
The US National Institute of Standards and Technology provides guidance for understanding and managing generative AI risks.

Timeline for the Implementation of the EU AI Act
The European Union’s official implementation timeline identifies the progressive application of AI Act requirements, including transparency and general-purpose AI provisions.

This article was prepared with the assistance of artificial intelligence tools. The content underwent expert human review, and Webiano Digital & Marketing Agency assumes editorial responsibility for its final version and publication.