The History of AI Productivity Tools

By Victor Hoang · Updated Oct 24, 2026

In 1960 a researcher tracked his own working day. J. C. R. Licklider found it was mostly chores. He spent it finding information, plotting graphs and instructing an assistant how to plot them.

He wrote that about 85 per cent of his "thinking" time went to getting into a position to think. AI productivity tools are software that uses machine learning to take over such small, repeated chores. They check spelling, answer mail, draft text, summarize meetings and find things. The goal is older than the technology. Vannevar Bush described a desk-sized personal memory machine in July 1945. Ralph Gorin wrote the first interactive spelling corrector at Stanford in 1971. The tools most people use now arrived between November 2022 and 2023. Each generation started with the clerical work around thinking, then pushed that line closer to the thinking itself.

Two neighbors cover adjacent ground, so this article stays on the work itself. The general field is in The History of AI, and the assistant that sits inside your documents has its own account in The History of AI Copilots. Here the question is narrower: how did software come to help with everyday office work, and what did each step change for the people doing it?

One distinction runs through the article. Bush's memex, Licklider's partnership and the early spell checkers are precursors: they pursued the same goal without machine learning or without a language model. The concept as people mean it today (AI that drafts, answers and summarizes inside daily tools) took shape between 2015 and 2023.

Timeline at a Glance

The Memex Dream (1945-1969)

Three researchers imagine a machine that takes the chores

The story starts in a country coming out of a war. In July 1945, in the Atlantic Monthly, Vannevar Bush published "As We May Think". He proposed a future device he called a memex, defined in his own words as a device in which an individual stores all his books, records and communications, mechanized so it can be consulted with speed and flexibility. He called it "an enlarged intimate supplement to his memory." His target was the growing pile of records that no one could search by hand, and his remedy was selection by association instead of by index.

Fifteen years later, Licklider turned the idea from storage to partnership. His March 1960 paper "Man-Computer Symbiosis" is available in full, and it rests on that self-audit: most of his working time went to preparation, not judgment. The machine's job, he argued, was to take the preparation.

In October 1962, Douglas Engelbart's SRI report for the Air Force Office of Scientific Research set out a framework for what he called augmenting human intellect, which he defined as increasing the capability of a man to approach a complex problem situation. Notice what all three share. None of them describes a program that does your job. Each describes a machine that clears the path for the person who does.

Nothing in the 1940s or 1960s could carry that out in an ordinary office, so these are precursors. What they solved was the aim. What they left open was a machine that could handle even one small chore reliably.

What it teaches teams today: Licklider's method still works. Before you pick a tool, log a week of your own work and count how much of it is preparation, searching and reformatting.

The Rules Years (1971-2002)

Spell checkers and paper clips run on rules and odds

The first office chore to be automated well was the one every writer dreads: catching typos. A Stanford recording of Ralph Gorin's talk gives the lineage. Les Earnest built a first spelling check in 1961 as part of a handwriting recognizer, using a list of the 10,000 most common English words. In 1967 a student at the Stanford AI Laboratory wrote one for text files in LISP, slow and listing unknown words with their locations. In 1971 Earnest recruited Gorin to build an interactive version. Gorin wrote SPELL in machine language for speed and created what the record calls the first spelling corrector, suggesting words one letter off or with adjacent letters swapped. He made it publicly available, and it spread over the ARPAnet about ten years before personal computers became common.

SPELL did not learn anything. It matched words against a list. But it set the template for the next thirty years of office help: a small tool that notices an error and offers a fix, so the writer stays in charge. Our history of expert systems covers the rule-based AI of the same period, and the history of office software shows where such features lived once word processors arrived.

The boldest rules-era attempt reached the mass market in 1997. Eric Horvitz and colleagues at Microsoft Research presented "The Lumiere Project" at the 1998 Conference on Uncertainty in Artificial Intelligence. It described Bayesian models that infer a software user's goals and needs from their actions, and its abstract says the Lumiere prototypes served as the basis for the Office Assistant in Microsoft Office '97. That animated helper, widely remembered as Clippy, was the first AI-style assistant shown to a huge audience of office workers. It guessed, from probabilities and rules, what you were trying to do.

The lesson of the era was less about the math than about manners. A guess shown at the wrong moment feels like an interruption, and the Office Assistant became shorthand for help nobody asked for. A suggestion has to be right and well timed to be welcome.

What it teaches teams today: the cheapest AI feature is often the narrowest. A tool that fixes one known error reliably beats a tool that interrupts constantly with guesses.

The Assistant Project (2003-2014)

DARPA funds a learning aide, and a descendant lands on the iPhone

In 2003 the U.S. defense research agency DARPA funded a big attempt to build an assistant that could learn its owner's habits. SRI International led the project, called CALO, for Cognitive Assistant that Learns and Organizes. SRI describes it as a collaboration of more than 300 researchers from 22 research institutions over five years from 2003, part of DARPA's PAL program. Its offshoots included CALO-MA, a meeting assistant, an early sign that meetings were an obvious target for automation.

The best-known descendant came from SRI itself. SRI says it spun off Siri, Inc. in 2007, that Apple acquired Siri in April 2010, and that Siri was unveiled as part of the iPhone 4S in October 2011. Apple's own release at the time said the phone introduced Siri, an intelligent assistant that helps you get things done just by asking.

For the first time, millions of people could ask a machine to set a reminder, send a message or look something up in plain speech. But these assistants worked from fixed commands and fixed integrations. If your request fell outside what the designers had built, the assistant said so, or answered something else. Our history of early AI chatbots traces the conversational line of this idea back further.

The era proved demand and exposed the ceiling. The assistant could act on a few things well but could not write, and it could not cope with an open-ended request.

What it teaches teams today: a tool that handles twenty commands well still leaves the other thousand tasks to people. When you evaluate an assistant, list the tasks it will not cover.

The Suggestion Years (2015-2021)

Machines start writing the next sentence

The change from commands to writing began in a place every office worker visits all day: the inbox. On November 3, 2015, the Gmail team announced Smart Reply in Inbox. The post says it suggests up to three responses to an email, using machine learning to spot messages that need a reply and to generate the text. The stated motive was practical: replying on a phone is slow.

In June 2017 a team of eight authors posted "Attention Is All You Need", introducing the Transformer, an architecture that relies on attention alone and drops recurrence and convolution. It did not look like a productivity paper. But it became the base of the language models that later powered writing tools, and our history of large language models follows that thread.

On May 16, 2018, Yonghui Wu of the Google Brain team described Smart Compose, which offers sentence completions as you type, for new messages and replies alike. The post says it builds on technology developed for Smart Reply, and it names a latency target of about 100 milliseconds per keystroke, since a suggestion that arrives late is useless.

In ACE terms (Ingest, Analyze, Predict, Generate, Execute), these tools sat at Predict: they guessed the next words, and the person pressed a key to accept. They did not act for anyone. The new problems were sameness (everyone's email starting to sound alike) and trust in text no one had chosen word by word. Our overview of AI writing assistants maps today's version of this category.

What it teaches teams today: accepting a suggestion is a decision. If your team uses autocomplete for customer email, set a norm for what a person must read before sending.

The ChatGPT Moment (2022-2023)

A free chat window turns a research model into a daily tool

On November 30, 2022, OpenAI released ChatGPT, a conversational model open to the public. For office workers, the difference from Smart Compose was the open prompt. You could ask for a cover letter, an email about a reorganization or a plan for analyzing customer notifications, and get a draft.

Researchers moved quickly to measure it. In a study issued in April 2023 and revised that November, Erik Brynjolfsson, Danielle Li and Lindsey Raymond tracked 5,179 customer support agents given a generative AI assistant. Productivity, measured as issues resolved per hour, rose 14 percent on average and 34 percent for novice and low-skilled workers, with minimal impact on experienced and highly skilled ones. MIT's Shakked Noy and Whitney Zhang published a writing experiment in Science on July 14, 2023. According to MIT Economics' summary, 453 college-educated professionals who used ChatGPT cut their time on writing tasks by 40 percent while independent evaluators rated quality 18 percent higher. Both results come with limits: one is a single support operation, the other short, self-contained tasks.

The suite makers followed with the assistant in the document. On March 16, 2023, Jared Spataro announced Microsoft 365 Copilot, describing a first draft in Word, trends and charts in Excel, presentations from a prompt in PowerPoint, summaries of long threads in Outlook and key points during meetings in Teams. At that point Microsoft was testing with a small group of customers. The September 21, 2023 post said enterprise customers could buy it on November 1, after use by tens of thousands of enterprise users in an early access program, and Microsoft's November 15 post recorded that it had become generally available on November 1, 2023. So the announcement and the shipping date are two different events, nearly eight months apart. That same November post described meeting features still to come: a recap of key points, action items and decisions for missed meetings was slated to begin in December 2023. The meeting side of the story is in AI meeting notes and summaries.

For a fuller look at the copilot label and rival suites, see The History of AI Copilots.

The era solved the blank page. It created a review burden: someone has to check the draft, the summary and the numbers, and that work does not appear in a "time to first draft" metric. The broader debate is in the AI productivity paradox, which echoes an older argument in our productivity paradox history.

What it teaches teams today: the gains in both studies were largest for the least experienced workers. Pair new hires with these tools, but have an experienced person review what goes out.

The Age of Agents (2024-present)

Assistants connect to your tools and take whole tasks

The newest change is that some assistants stopped waiting for the next keystroke. A prerequisite was plumbing. On November 25, 2024, Anthropic announced that it was open-sourcing the Model Context Protocol, an open standard for connecting AI assistants to the systems where data lives, with pre-built servers for tools such as Google Drive, Slack and GitHub. The idea was that an assistant should reach your files and messages through one shared connector, not a custom integration per app.

Then came tools that act. On May 19, 2025, Thomas Dohmke described a GitHub Copilot coding agent: you assign it an issue, it works in the background and pushes commits to a draft pull request, and it tags you for review. The post lists the controls around it. It can push only to branches it created, the developer who asked for the pull request cannot approve it, and workflows do not run until a human approves them. In ACE terms this moves from Generate to Execute, with the human placed at the approval step. The longer line is in The History of AI Agents.

What has the arrival of all this actually changed? One careful answer comes from Denmark. In a working paper dated May 9, 2025, Anders Humlum and Emilie Vestergaard linked adoption surveys covering 25,000 workers in 11 exposed occupations to employer-employee records. They report no significant impact on earnings or recorded hours in any occupation, and average time savings of 3 percent. Set that beside the 14 and 40 percent figures from task-level studies and the lesson is plain: gains on a measured task do not automatically become gains across a job. Our article on AI versus traditional productivity software compares the two kinds of tool.

The same sequence of questions has followed every earlier tool, as the pillar history of productivity shows: first what the tool can do, then how work has to change before it pays off. Two sibling histories pick up the surrounding context, the history of collaboration tools and the history of remote work.

What it teaches teams today: with agents, the design question is who approves what. Decide the review step before you grant a tool permission to act.

Key People

Key Facts

Key Facts: AI Productivity Tools History

• Bush's memex essay ran in The Atlantic Monthly in July 1945, and Licklider's March 1960 paper put 85 per cent of his thinking time into getting ready to think.
• Gorin's SPELL, written in 1971, spread over the ARPAnet about ten years before personal computers became common.
• SRI's CALO ran with more than 300 researchers from 22 institutions, and Siri is its most famous descendant.
• A study of 5,179 support agents found a 14 percent average rise in issues resolved per hour, and 34 percent for novices.
• In a Science experiment, ChatGPT users cut writing time by 40 percent with 18 percent higher rated quality across 453 professionals.
• A Danish study of 25,000 workers reports average time savings of 3 percent and no significant effect on earnings or hours.

The Pattern Behind the History

We call this the Chore-to-Colleague Pattern. It is our own synthesis of the timeline above, not a published model.

1. A chore is named first. Bush's records, Licklider's graphs and Gorin's typos were all clerical work around the real job.
2. A narrow tool wins trust. SPELL fixed one error type reliably before anything tried to guess intent.
3. Guessing arrives before understanding. The Office Assistant and the first assistants ran on rules and probabilities, so help was uneven.
4. A general model removes the ceiling. Smart Reply, the Transformer and ChatGPT moved tools from fixed commands to open requests.
5. Autonomy follows, and so does review. Agents take whole tasks, and the safeguards sit at the approval step.

What It Teaches Teams Today

• Start with the chore, not the label. Log where the week goes before choosing a tool.
• Match the claim to the sample. A 14 or 40 percent gain was measured on one kind of task, and a 3 percent average came from a whole workforce.
• Pair beginners with tools, and keep a reviewer. Both field studies found the biggest gains for the least experienced.
• Design the approval step early. The more a tool does alone, the more the human check carries.
• Fix the surrounding work. Earlier tools paid off only after work was reorganized around them, as the history of knowledge work and the history of the to-do list both show.

The tools changed, but the trade stayed the same: hand the chore to the machine, keep the judgment, and decide who checks the result.