A program of EffectiveSC
How We Use AISources checked October 24 to 29, 2026
Where things stand / Products built as AI
What the companies that build products as AI publish about how much they're used, how they're priced and who checks their work.
Robb Wilson, who leads this research, works at OneReach.ai, which sells an agent platform, one of the kinds of product this page covers. EffectiveSC, which publishes How We Use AI, plans to offer AI services of its own. This page doesn't rank vendors or recommend a product. Its cards report what each company publishes about its own product, in that company's terms. Its podcast accounts are sellers describing their own products, checked against transcripts, not the audio, and the judgment near the end is the project's own reading, not a measurement.
The clearest view of heavy use comes from a maker's own logs, and from one company whose bills for coding agents ran past its budget. Each source draws the line in its own place, so the gaps below don't compare across sources.
6x
Workers at the 95th percentile of ChatGPT Enterprise use send 6 times the median worker's messages, and 17 times as many coding messages.
Both ends of the comparison are ChatGPT Enterprise users at OpenAI's customers; people who don't use it aren't counted.
Disclosure: OpenAI's own usage data about its own product.
7x
Frontier firms send about 2x the median firm's messages per seat but 7x as many to GPTs; weekly users of Custom GPTs and Projects rose about 19x in 2025, and about 20% of Enterprise messages run through a Custom GPT or Project.
Frontier firms are OpenAI's top firms by use, set against its median firm; the report doesn't say the Custom GPTs and Projects are shared with anyone.
Disclosure: OpenAI's own usage data about its own product.
$1,500
Uber used up its full-year 2026 AI budget within about four months, driven by AI coding tools such as Claude Code, and now caps each employee at $1,500 a month per AI coding tool.
Reported by The Information and Bloomberg, from Uber's chief technology officer and a spokesperson; both reports are behind paywalls.
Disclosure: Claude Code is Anthropic's product, and Claude helped compile this research.
Some agents are sold to finish a job rather than help a person do it, and some are priced per resolution or per outcome. Each seller publishes its own rate under its own test for resolved, so the rates below aren't one series.
76%
Fin (formerly Intercom) states an average resolution rate of 76% across 12,000+ customers, improving approximately 1% every month; its benchmark page defines resolution rate as the share of AI-involved conversations resolved without human help, and its top-10 teams average 78% automation, 85% resolution and 91% involvement.
Fin counts a conversation as resolved when the customer confirms it or doesn't follow up, and only among conversations Fin takes part in. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
Disclosure: Fin's own figure about its own product.
$0.99
Fin's pricing page lists its AI agent at $0.99 per outcome, counted when a customer confirms the issue is resolved, doesn't ask for more help after Fin responds, or Fin completes a workflow (a Procedure), including one that ends in a handoff; nothing is charged if Fin can't answer or the customer asks for a person, and a reopened conversation is refunded.
An outcome is more than a resolution: a workflow that ends in a handoff is billed too, and nothing is billed when the customer asks for a person. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
Disclosure: Fin's own price page.
Salesforce agreed on June 15, 2026 to acquire Fin, formerly Intercom, for approximately $3.6 billion, and completed the acquisition on September 10, 2026.
Salesforce's own release describes Fin's rate on a different base, all support volume, from the one Fin's pages use. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
70%
HubSpot's product page states a 70% average resolution rate for its customer agent (top teams hit 90%), and says customers using it close 74% more tickets and have an 84% higher ticket resolution rate, on average; it charges $0.50 per resolution.
HubSpot counts a resolution when the agent gave support and no person took the conversation over for a set period afterward, or when the agent qualified a lead. The page gives no base, period or comparison group for its averages. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
Disclosure: HubSpot's own figures about its own product.
Salesforce said Agentforce on help.salesforce.com had handled 4.3 million inquiries and resolved 70 percent of them (June 25, 2026); its customer-zero page says Agentforce on Help has surpassed five million customer conversations, resolving more than 63% of them instantly.
Salesforce counts a conversation as resolved when Agentforce answered the customer's inquiry; the customer doesn't have to confirm it, and there's no waiting period. The two totals differ, and Salesforce doesn't say why. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
Disclosure: Salesforce's own count of its own agent on its own help site.
88,371
In the week of October 18, 2026, Agentforce on Salesforce's help site handled 88,371 conversations: 67,002 resolved, 6,302 abandoned, 11,353 handed to a human agent and 3,714 handed over at once at the customer's request.
Counts for one week on one help site, not a rate. Salesforce defines resolved as Agentforce answering the inquiry, and a handoff as a transfer to a human agent. Fin, HubSpot and Salesforce each define resolved differently, so their rates aren't one series.
Disclosure: Salesforce's own log of its own agent on its own help site.
The companies that make coding agents publish shares of their own code, and each counts a different event: code generated and approved, code merged, or code merged after an agent's review. The shares aren't one series, and none says how much work went out with no person checking it.
80%
As of May 2026, more than 80% of the code merged into Anthropic's codebase was authored by Claude, with engineers directing and reviewing; an automated Claude reviewer reads proposed changes before they can merge.
Anthropic counts lines merged to production that it can attribute to Claude, with an engineer directing and reviewing; it says lines of code overstate the productivity gain.
Disclosure: Anthropic makes Claude, which helped compile this research.
75%
Google said 75% of all new code at Google is now AI-generated and approved by engineers, up from 50% the previous fall.
Google counts new code generated by AI and approved by engineers, and publishes no method.
Disclosure: Google sells the models and coding tools it describes.
OpenAI's harness-engineering team said that over five months three engineers, later seven, directing Codex opened and merged roughly 1,500 pull requests in a repository where Codex wrote every line; humans may review pull requests but aren't required to, and agents often merge their own.
One team's repository, not OpenAI as a whole; the team chose to write no code by hand.
Disclosure: OpenAI sells Codex, the coding agent the team used.
70%
Qoder's lead at Alibaba said 70% of the team's code now merges directly after review by an agent, and that agents finish about 60% of the work in the loop in many areas, up from about 30%.
One team's account in a media interview, not an Alibaba page; an agent, not a person, reviews the code that merges.
Disclosure: Qoder is Alibaba's own coding product.
0% to 20%
More than half of 132 Anthropic engineers and researchers said they could fully delegate only 0% to 20% of their work.
A survey of Anthropic's own engineers and researchers, whose answers weren't anonymous; they read fully delegate as anything from no checking at all to light oversight.
Disclosure: Anthropic makes Claude, which helped compile this research.
62.1%
62.1% of engaged Copilot Business and Enterprise users are still primarily engaging with code completions; across thousands of enterprises, average merged pull requests were 78%, 41% and 151% higher from code first to agent first, agent first to multi-agent, and code first to multi-agent.
GitHub's code-first group includes agent mode inside the editor, and the gains in merged pull requests compare users already in each group; nothing assigned the tool.
Disclosure: GitHub's own telemetry about its own product.
Shoppers now reach stores from AI assistants, and retailers and search engines build assistants of their own. The figures come from the companies that run them, and none says how many users come back.
nearly 50%
In Shopify's Q1 2026 data, AI-referred sessions that begin on a product page converted at nearly 50% higher rates than organic search, AI-referred orders carried 14% higher average order values, referral sessions from AI chatbots grew more than 8x and AI-referred orders nearly 13x year over year; organic search still referred more sessions.
The gap is for sessions that begin on a product page. Shoppers who arrive from an assistant may differ from those who search, and Shopify counts referrals from Google's AI Overviews as organic search.
Disclosure: Shopify's own data about sales on its platform.
250 million
Amazon said 250 million customers used Rufus in 2025 and shoppers using it were 60% more likely to complete a purchase; in Q2 2026 Alexa for Shopping's active users close to doubled and interactions rose over 5x year over year, and US customers who use it spend over 40% more per order than those who don't.
Shoppers who use the assistant differ from those who don't before they open it, so the purchase and spending gaps are comparisons, not effects. Amazon defines no active user and gives no base.
Disclosure: Amazon's own figures about its own assistant.
Google said queries in AI Mode are on average three times longer than traditional searches, and one in five AI Mode searches globally use voice, images or documents.
These describe how people use AI Mode, not how many use it or whether they come back.
Disclosure: Google's own figures about its own product.
A public index grades the makers of AI agent products on what their own pages say, not on how well the products work.
About half of agent vendors fully document a control that lets a person approve an agent's actions: 554 of 1,082 (51.2%) in Agentic Index's data of October 29, 2026 (this project's count from a workbook export of the index), and 483 of 955 (50.6%) on the index's feature page, measured October 7, 2026.
Counts what vendors document on their own public pages, not whether the control works in practice.
Disclosure: The index doesn't list OneReach, the agent platform company where this research's lead works.
About one in five of Agentic Index's agent platforms fully document all four governance basics it sets for enterprise buyers: SOC 2 or ISO 27001 with single sign-on and role-based access control; self-hosted or private VPC deployment with control over where agent data is processed; tracing and audit logging; and human approval before an agent action executes. Its governance page gives 118 of 554 (21.3%), its home page 118 of 561 (21.0%), and the workbook's lanes 145 of 615 (23.6%).
Counts what vendors document on their own public pages, not whether they meet the basics in practice.
Disclosure: The index doesn't list OneReach, the agent platform company where this research's lead works.
Founders and executives describe their own products on podcasts: how they price them, how they check them and where they fall short. Each card paraphrases one person's account, with the episode and the time it was said. These are claims by people with a stake, not measurements.
Lassie co-founder Steijn Pelle said on The a16z Show that Lassie charges five figures for its first agent, which does about 30 of the roughly 200 hours of admin work a dental practice has each month, reconciling insurance payments and billing patients; he said it took years to build because an agent that takes over a job has to err toward correctness, and that the first agent does 98% of that work, leaving a sliver it can't yet do.
One person's account on a podcast, with no definition for the share of work; checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described, on a show run by Andreessen Horowitz, which says it may invest in the companies it discusses.
Bedrock Robotics' chief executive said on Bloomberg Tech that contractors pay for its autonomous excavators in different ways: some license the system like an operator added to their fleet, and some pay by productivity, in cubic yards of earth moved; he said this helps contractors tighten the uncertain length of their projects.
One person's account on a podcast, checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described.
Decagon co-founder and president Ashwin Sreenivas said on The a16z Show that about 90% of Decagon's agent workflow now runs on open-source models its research team tunes, chosen mainly to cut latency for its voice agents, with frontier models kept for new projects and products; co-founder Jesse Zhang said the tuned smaller models do better than large frontier models on their specific tasks and are cheaper and faster.
One person's account on a podcast, checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described, on a show run by Andreessen Horowitz, which says it may invest in the companies it discusses.
Decagon co-founder and president Ashwin Sreenivas said on The a16z Show that Duet, a second, bigger and slower agent, now does the setup work Decagon's staff used to do for its customer-service agents: from a customer's transcripts and documentation it drafts the agent's operating procedures, integrations and tests, and once the agent is live it reads the conversations, flags topics handled poorly and drafts improvements.
One person's account on a podcast, checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described, on a show run by Andreessen Horowitz, which says it may invest in the companies it discusses.
Warp chief executive Zach Lloyd showed on How I AI how Warp's software factory, which Warp sells and runs on its own engineering, takes a request tagged in a public Slack channel, opens an issue, implements the change, opens a pull request and checks the result with computer use; reading Warp's dashboard, the host said a change took 35 minutes from kickoff to pull request and three and a half hours from pull request to first human review, and Lloyd said every pull request still gets human code review.
One person's account on a podcast; the two times are the host's reading of Warp's own dashboard. Checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described.
Factory co-founder and chief technology officer Eno Reyes said on The Twenty Minute VC that Factory's engineers scope a project, share a bid of what it should cost and send it off, and that Factory sells a product, agent effectiveness, that shows how many credits each project used so a business can check whether it got the outcome; he said it makes more sense to allocate spending to projects run by agent systems than to map agents one to one to people.
One person's account on a podcast, checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described, and the host's fund invests in it.
A co-founder of Basis said on The Cognitive Revolution that Basis checks its accounting agents against written behavior specs drawn from long practice in tax work: a second agent, acting as a judge, reads each run's record and asks whether a behavior's trigger occurred and, if so, whether the behavior was followed. In a hypothetical he gave, an agent that looks at its PowerPoint edits before delivering them would catch some formatting errors but add latency and cost.
One person's account on a podcast, checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described.
Simile co-founder and chief executive Joon Sung Park said on Latent Space that Simile rests on a study he led of about 1,000 Americans, sampled to represent the US, who gave two hours of interviews and other data and came back two weeks later for surveys, economic games, a personality test and published experiments; their digital twins reproduced their behavior and attitudes 85% as accurately as the people reproduced their own answers.
One person's account on a podcast of a study he led; checked against the transcript, not the audio. The study itself wasn't opened for this page.
Disclosure: The speaker led the study and runs the company that sells the product built on it.
Hint co-founder and chief technology officer Kyle Rush said on The Cognitive Revolution that when Hint trialed an AI agent that calls home-service contractors for a homeowner, it once called a generator technician 17 times in a row until he picked up, so he left a job site thinking it was an emergency, and then asked for a model number it didn't need; he said the technology needs to evolve.
One person's account on a podcast of one incident in a trial; checked against the transcript, not the audio.
Disclosure: The speaker's company makes the agent described.
Pavankumar Reddy Muddireddy, who leads audio research at Mistral AI, said on Machine Learning Street Talk that customers his team spoke with complain that speech recognition is far from solved where they deploy it, even in customer service, that it makes many mistakes, and that they add a lot of scaffolding to handle the corner cases that come up over millions of sessions with a voice agent.
One person's report on a podcast of what customers told his team; checked against the transcript, not the audio.
Disclosure: The speaker's company sells speech models, and the episode was made in partnership with it.
Parallel co-founder and chief executive Parag Agrawal said on The Twenty Minute VC that Parallel sells web search to AI agents, with a fast mode for voice agents and an advanced one for costly background agents, and that its customers' agents mostly do knowledge work, such as coding, law, productivity and insurance underwriting; he said coding invokes web search in about 5% of prompts, because most of that work reads a customer's own code and data.
One person's estimate on a podcast; checked against the transcript, not the audio.
Disclosure: The speaker's company sells the product described.
Alibaba.com president Kuo Zhang said on The Artificial Intelligence Show that Accio Work turns Alibaba.com's sourcing agent into a workspace for small sellers with stores on several platforms, whose agents give one view of the business and publish a listing once across Shopify, Amazon, TikTok Shop and others; he said the tokens its users consume rose sixfold in the four months after its April launch.
One person's account on a podcast; checked against the transcript, not the audio. Tokens consumed measure use, not sales.
Disclosure: The speaker's company sells the product described.
The usual ways of sorting AI products ask where the AI sits in front of a person: inside a tool, in a panel beside it, as an add-on seat or as an agent that finishes a job. That list misses products whose user is another company's agent. MCP servers let outside agents read and write a company's data, web search is sold to agents rather than to people, as Parallel's account above describes, and some databases are now created by agents more often than by people. This is the project's reading of the cards below and the accounts above, not a measured share of any market: nobody yet counts how much software is bought for agents to use.
1 million
Atlassian said monthly active users of its MCP server and Teamwork Graph CLI, through which agents and other tools reach its Teamwork Graph from outside Atlassian's own apps, more than doubled in the quarter ended June 30, 2026, to more than 1 million, with MCP calls up more than 400% from the prior quarter; 98% of MCP users were also active in Jira's own interface in the same month.
Atlassian's own count, with no base. It counts users of its MCP server and command-line tool, not agents or the work they did.
Disclosure: Atlassian's own figures about its own product.
10,000
Anthropic said in December 2025 that there were more than 10,000 active public MCP servers and 97M+ monthly SDK downloads across Python and TypeScript, and that ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code had adopted MCP.
Anthropic's own count, with no method given. It counts public servers, not how often agents call them.
Disclosure: Anthropic created MCP, and Claude helped compile this research.
55%
In a16z's October 2026 survey of 100 CISOs at companies with more than 1,000 employees, counting limited as well as broad production use, 92% report employee generative AI tools in production, 83% coding assistants or agents, and 55% MCP servers or other agent-to-application integrations.
Counts limited as well as broad production use, as the survey does. These are security chiefs at large companies, and the report doesn't say how they were chosen.
Disclosure: Andreessen Horowitz, which ran the survey, invests in AI companies, including Armadin and Neo, which its report names.
80%
Databricks said in May 2025 that over 80% of the databases provisioned on Neon were created by AI agents rather than humans, up from 30% when Neon became generally available a year earlier.
One platform built for agents, described by the company that was buying it.
70%
Supabase said in October 2026 that it adds more than 1 million users and 4 million databases a month, with 70% of new databases created by agents or AI-driven tools.
Supabase's own figures, from a funding announcement about its own platform.
No product on this page publishes, for itself, the three things a buyer most needs: how many customers keep using it, how much of its work is accepted as it stands, and what a person still checks before the work counts. The resolution rates use three different tests for resolved, and none reports how many customers came back with the same problem. The coding figures count code generated, merged or reviewed by an agent, not changes that held up. Salesforce's weekly help-site counts come closest, because they show what customers abandoned and what went to a person. The podcast accounts describe the checks sellers built, a judge agent that reads each run, a second agent that watches the first and pull requests that wait for a person's review, and none gives a rate for them.
The AI at Work Benchmark
Score your organization's practices, measure your people with an anonymous pulse, and see both against the published evidence.