Agent Atlas

Workplace

AI at Work

Buying seats does not make a company better at its work. Training does. The studies show the same tool can save a lot of time on one task and slow people down on the next. This page covers what the evidence says, what a company plan changes, and how to roll it out.

Training kitCopilot at Work →Microsoft Copilot app by app, prompts for eight roles, building agents, a 90-day plan and how to measure it.

The evidence, both directions

Factual
Controlled studies: big gains on some tasks, worse results on others.

453 professionals, writing tasks (Science)

+40% less time per task

Consultants, tasks inside the frontier (Organization Science)

+25.1% faster

5,172 support agents (arXiv)

+15% more issues resolved per hour

One coding task with an AI pair programmer (arXiv)

+55.8% faster

Consultants, a task outside the frontier (Organization Science)

−19% less likely to be correct

Experienced developers, their own repositories (METR)

−19% slower

One study calls it the jagged technological frontier. Inside it, consultants with AI finished 12.2% more tasks, 25.1% faster, and their work was better. On one task picked to sit outside it, they were 19% less likely to get it right (Organization Science). Each bar measures something different, so compare directions, not lengths.

What a company plan buys

Factual
Seven workplace products: price, what happens to your data, and the controls.
ProductPriceYour dataControls
ChatGPT Business$20 a seat a month billed yearly, $25 monthlyNot used for training by defaultSAML SSO and MFA. Connects to Google Workspace, Slack, GitHub, Microsoft 365 (OpenAI; OpenAI; OpenAI)
ChatGPT EnterpriseAsk salesNot used for training by default. Custom retentionSCIM, encryption keys you manage, role-based access, user analytics (OpenAI; OpenAI)
Claude Enterprise$20 a seat a month billed yearly, plus usage at API rates. 20 seats minimumNot used for training by default. Custom retention, logged in audit logsSingle sign-on (Anthropic; Anthropic; Anthropic)
Microsoft 365 Copilot$30 a user a month, paid yearly, on top of a Microsoft 365 planPrompts, responses and Graph data not used to train foundation modelsAudit logs come with the organization’s auditing. Connectors index data or fetch it live over MCP (Microsoft; Microsoft; Microsoft; Microsoft)
Gemini in Google WorkspaceBusiness Standard lists $16.80 and $14 a user a monthNot used to train Gemini or for adsGemini cannot open Workspace content you have no permission to see (Google; Google)
GleanNot publishedIndexed into your own isolated tenantConnectors copy each source’s permissions (Glean)
Salesforce Agentforce$2 a conversation, or credits, or per userAnswers questions and takes actions in SalesforcePriced by what the agent does, not only by seat (Salesforce; Salesforce)

Prices as printed on 15 November 2026.

What companies say it did

Factual
Klarna and Moderna, in their own announcements.

$40M

Klarna’s estimate of the profit its support assistant would add in 2024. In its first month it had 2.3 million conversations, two-thirds of customer service chats, the work of 700 full-time agents (Klarna).

The company’s own first-month estimate, not an audited result.

Everyone

Moderna gave every employee a purpose-built AI assistant in 2024 and trains people on the job (Moderna). Its AI Academy for staff at every level started in 2021(Moderna).

Training came first.

How adoption is organized

Simplified
Six steps, from who owns it to how you measure it.

Each step comes from a standard, a regulator or a study. The order is ours.

  1. Give it an owner

    Successful organizations put AI policy under central ownership, such as an AI Center of Excellence that sets the common boundaries. (Microsoft)

  2. Sort the uses

    Individual work, like writing help and meeting preparation, is a different project from automation that changes how the business runs. (Microsoft)

  3. Pick the plan for the data

    Company plans change the terms. Anthropic’s consumer terms, which keep data five years if you allow training, do not apply to Claude for Work or the API. (Anthropic)

  4. Govern to a standard

    NIST’s AI Risk Management Framework is voluntary, with a profile for generative AI. ISO/IEC 42001 sets requirements for an AI management system. (NIST; NIST; ISO)

  5. Train everyone

    In the EU it is the law: providers and deployers of AI systems must support the AI literacy of their staff. (European Union)

  6. Measure each task

    The same tool sped some tasks up and slowed others down. Measure the task, not the tool. (Organization Science; METR)

Do the seats pay for themselves?

Estimate
Your team, your hours, your measured time saved. Seats against hours back.

Hours back a month0

Worth$0

Seats cost$0

Net a month$0

Arithmetic: people × hours × 4.33 weeks × time saved, priced at the hourly cost, minus seats. The presets are the study results above (Science; arXiv; METR). A lab result is not your office: run a pilot, measure the task, and put your own number here. Claude Enterprise also bills usage on top of the seat.

Courses that already exist

Factual
Free and paid AI training from the labs, Google, Microsoft and one company’s own academy.
  • Anthropic

    AI Fluency

    The 4Ds: Delegation, Description, Discernment and Diligence. (Anthropic)

  • Anthropic

    AI Fluency for educators

    For faculty and education leaders, built on the 4Ds, one of three courses made with educators. (Anthropic; Anthropic)

  • OpenAI

    OpenAI Academy

    Using AI at work, in school and at home, with guides on how teams use ChatGPT by role. (OpenAI)

  • Google

    Google AI Essentials

    Using generative AI to speed up daily tasks and develop ideas. (Google)

  • Google

    Google AI Professional Certificate

    Eight courses of about an hour each, for any role, no experience needed. (Google)

  • Microsoft Learn

    AI fluency learning path

    From foundational concepts to more advanced topics. (Microsoft)

  • Moderna, in-house

    AI Academy

    Built in 2021 for its 2,400 team members, at every level. (Moderna)

A first workshop

One hour for a team that has just been given AI. Ours, built on the evidence above.
  1. 10 min Where it helps and where it hurts: the jagged frontier, with the team’s own tasks sorted in or out.
  2. 10 min The company’s plan: what happens to the data, who can see what, and why personal accounts are on different terms.
  3. 20 min The 4Ds on a real task from each person’s week: delegate it, describe it, judge the result, own it.
  4. 10 min One personal use and one process to automate, written down with an owner.
  5. 10 min Time the task before and after, and put the number in the calculator.

The 4Ds are Anthropic’s (Anthropic). For how to write the task itself, see the Prompt Workshop.