AI ROI, the return on what you spend on AI, depends less on the price list than on the time a task really saves once checking is counted. This lesson shows how to estimate it and how to measure it.
In this lesson
- The four ways AI is priced, and which fits which use.
- The costs that do not appear on the price list.
- A simple sum for whether handing a task to AI pays.
- How to measure the result with a two-week log, before and after.
How AI is priced
There are four common pricing models. Most businesses end up paying in more than one way.
- Per seat, per month. A business plan of a chat assistant, priced per user, often cheaper per month if paid yearly. This is the usual starting point: predictable, easy to budget, and it comes with the data and admin terms covered in Lesson 2.
- Included in software you already pay for. Accounting, email, help desk, booking and design software increasingly include AI features, sometimes in the existing price, sometimes as an add-on or a higher plan. Check what you already have before buying anything new.
- Usage-based, through an API. If you, a developer or a software supplier builds on a model directly, you pay for what you use, measured in tokens (small pieces of text going in and out). Costs are tiny per request but can grow quickly with volume, long documents or agents that take many steps (Lesson 8).
- Per outcome. Some customer service bots charge per conversation they resolve or per action they complete, so the price rises with use. See, for example, the pricing as listed on our pages for Fin AI Agent and HubSpot Customer Agent.
The current business plans and prices of the main assistants, as each vendor lists them, are in this table. Each row shows the date we read the vendor’s page.
| Maker | Assistant | Business plans | Built into | Agent features | Data use (business, in the maker’s words) | Price as listed | Date read |
|---|---|---|---|---|---|---|---|
| Alibaba (Qwen) | Qwen Chat | Alibaba Cloud Model Studio (pay-as-you-go model API) | Not stated on the pages we read. | Not stated on the pages we read. |
| October 6, 2026 | |
| Anthropic | Claude |
| Anthropic's own add-ins for other makers' software, as listed on the Claude pricing page: Claude in Chrome, Claude for Microsoft 365 and Claude for Microsoft Outlook (Pro, Max, Team and Enterprise). Microsoft also offers Anthropic models inside Microsoft Copilot (see the Microsoft row). | Included in Claude plans: Pro $17/month with annual billing ($200 billed up front) or $20 billed monthly; Max 5x $100/month; Max 20x $200/month. Team: Standard seat $20/seat/month billed annually ($25 monthly); Premium seat $100/seat/month billed annually ($125 monthly). Enterprise: US$20/seat/month billed annually plus usage at API rates. Not included in the Free plan. As listed on October 3, 2026. | October 6, 2026 | ||
| DeepSeek | DeepSeek (chat at chat.deepseek.com) | DeepSeek API (pay per token from a topped-up balance) | DeepSeek Harness | Not stated on the pages we read. |
| October 6, 2026 | |
| Gemini (Gemini app; Gemini Enterprise app for business) |
| Google Workspace: Gemini in Gmail, Docs, Sheets, Meet and more, the Gemini app and Gemini Notebook are listed on the Workspace plans; the privacy hub also covers Gemini in Chrome. |
| October 6, 2026 | |||
| Meta | Meta AI |
| Meta Business Agent starts with the WhatsApp Business app, Instagram Pro, Messenger and Meta Business Suite; the Meta AI assistant runs on AI glasses and the Meta AI app and web, per Meta's connectors page. | Not used to improve our products.Source |
| October 6, 2026 | |
| Microsoft | Microsoft Copilot (formerly Microsoft 365 Copilot); Microsoft Copilot Chat (formerly Microsoft 365 Copilot Chat) |
| Microsoft 365 apps: Copilot in Word, Excel, PowerPoint, Outlook, OneNote and Microsoft Teams; Copilot Mode in Edge; the Microsoft Copilot app. Also Security Copilot, GitHub Copilot and LinkedIn Hiring Assistant (see tool records). | Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMsSource |
| October 6, 2026 | |
| Mistral AI | Mistral Vibe (formerly Le Chat) |
|
| October 6, 2026 | |||
| OpenAI | ChatGPT |
| Listed on chatgpt.com: ChatGPT for PowerPoint and ChatGPT in Slack and Teams; Microsoft lists 'Latest OpenAI models: GPT-5.X' in Microsoft Copilot (see the Microsoft row). | Your data is never used for training our models.Source |
| October 6, 2026 | |
| xAI | Grok |
| Grok on X (X is a separate company, per SpaceXAI); SpaceXAI's own add-ins for Microsoft Word, Excel, PowerPoint and Outlook and an add-on for Google Workspace; Grok models listed in GitHub Copilot, Microsoft Foundry, Amazon Bedrock and Gemini Enterprise Agent Platform (news page). |
| October 6, 2026 |
Prices change often. Always check the vendor’s own page before you buy.
The costs that are not on the price list
The subscription is usually the smallest cost. The others are where pilots succeed or fail.
- Set-up time. Choosing a tool, connecting it to your systems, writing the instructions and the “facts sheet” it works from, and testing it on real cases. For a simple chat assistant this is hours; for a customer-facing bot or an agent connected to your accounts, it can be weeks.
- Checking time. Every task that goes to AI with a person checking still costs that person’s time. This is the cost most often left out, and it is the one that decides whether a task pays.
- Training. Staff need to learn how to brief the tool and how to spot its mistakes. Budget a few hours each at the start, and time to share what works.
- Mistakes. A wrong answer that reaches a customer costs goodwill, refunds or worse. Lesson 9 has a case where a business was held to what its chatbot said.
- Switching and lock-in. Prices, features and terms change. If your process depends on one vendor’s tool, a price rise or a discontinued feature costs you time to move.
- Management. Someone owns the tool: who has access, which data goes in, whether the spot checks happen. That is a small but real part of someone’s job.
Does it pay? A simple AI ROI sum
For any single task, a rough sum tells you whether it is worth trying:
Hours saved a week = (hours the task takes × times a week × share of the work AI takes) − checking time a week
Multiply the hours saved by what an hour of that person’s time costs you, take off the tool’s cost, and you have a rough weekly return.
- 0.5 h × 40 = 20 h a week on the task
- 20 h × 60% = 12 h that AI takes
- 12 h − 3 h = 9 h saved a week, after checking
Illustrative numbers, not a real business. Say replying to routine customer emails takes half an hour each, there are 40 a week, and AI drafts take 60% of the writing off the person. That is 0.5 × 40 × 0.6 = 12 hours. If checking and fixing the drafts takes three hours a week, the saving is 12 − 3 = 9 hours a week.
Three cautions:
- Be honest about the share. “AI takes 60% of the work” is a guess until you measure it. Start low.
- Checking time is real. If checking takes as long as doing, the saving is zero, however good the drafts look.
- Freed time only pays if it is used. Nine hours freed is worth something only if those hours go on work that matters: more customers served, faster replies, the work you never get to. Decide what that is before you start.
The same picture as bars: the time a task takes before, and after, with the checking time that remains in amber.
Illustrative numbers, not a real business.
How we estimate it for a whole job
Each job page on this site has a “What it costs” section that compares the cost of AI doing the job’s AI-capable tasks with the wages of a person, always as ranges. It is a rough guide at the level of a whole occupation, not a quote for your business, and it is explained on the page. For your own team, use the AI team planner. It turns our task ratings into hours freed and money, always as ranges: the task time freed times the median pay for each job, minus that job’s AI cost range (US pay for now). If you were looking for an AI ROI calculator, that is it: there is no separate calculator on this page, and the planner is where the money ranges live.
Measure it: the two-week log
The most useful lesson from the research in Lesson 3 is that people’s sense of time saved is unreliable. In one careful study, experienced developers using AI tools took 19% longer on real tasks, yet afterwards “still believed AI had sped them up by 20%” (METR). So measure.
A simple before-and-after log works for most tasks:
- Pick one task and the person who usually does it.
- Week 1–2, before: each time the task is done, log the date, how long it took, and anything that went wrong. A shared spreadsheet with four columns is enough.
- Week 3–4, with AI: the same log, plus a column for checking and fixing time, and one for mistakes the AI made that were caught (and any that were not).
- Compare: average time per task before and after, including checking; mistakes before and after; and how the person found it.
- Decide: keep, change or stop. If the saving is small and the mistakes are worrying, stop. That is a good result too: you learned it cheaply.
Measure quality as well as time. If replies are faster but customers write back more often to correct them, the task is not working.
What to count beyond hours
- Response time to customers, before and after.
- Error rate: corrections, complaints, refunds.
- Hand-over rate for customer-facing bots: how often it passes to a person, and whether those hand-overs were the right ones (Lesson 9).
- Staff view: whether the people doing the work think it helps. They notice problems before the numbers do.
A note on big promises
You will see claims that AI “saves ten hours a week per employee” or “pays for itself in a month”. Some businesses do see gains like that on the right tasks. But a claim without a comparison group and your own kind of work tells you little (Lesson 3). The two-week log on your own work is worth more than any vendor’s average.
Key terms
- TokenA small piece of text, roughly a short word or part of one, that a model reads and writes. Usage is often priced in tokens.More in the glossary
- APIA way for software to use a model directly, without a chat window. Developers build products on it and usually pay per use.More in the glossary
- Business plan (AI)A paid plan of an AI assistant for organizations, with admin controls and stated terms on how your data is handled.Plans differ. Read the vendor’s own words on data use, retention and where data is stored before you buy.Lesson 2: what to look for · More in the glossary
- AI cost vs a personThe cost of AI doing the hours of work it could handle, next to a person’s wage for those hours. Both are ranges.The AI cost covers the AI service only: not setup, licenses, oversight or the time a person still spends checking.More in the glossary
Every term is explained in the glossary.
Try this
Take the two tasks you circled in Lesson 4. For each, fill in the return sum with your best guesses: hours per task, times a week, share AI might take, and checking time. Write the answer down. Then run the two-week log on one of them, and put the measured numbers beside your guesses. The gap between the two is the most useful thing you will learn in this course.
Sources
- METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”, 10 July 2025. Read 6 October 2026. https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
- Prices: vendors’ own pricing pages, as recorded with their dates in our vendor table and tool pages.