SaaSCron: SaaS + AI

Will AI agents replace SaaS? The per-seat debate behind the 2026 software selloff

2026-10-31T07:00:00-04:00

You pay for software by the seat, or you sell it that way, and you keep reading that AI agents will make seats obsolete. If an agent does the work of five people inside a business application, does the vendor still get paid for five logins? This explainer sets out both sides with named, dated sources and leaves the forecast to you. For the subscription model at stake, start with What is SaaS?.

“SaaSpocalypse” is a nickname for the fear that AI agents will hollow out software companies. The research firm Gartner used the word in a July 2026 release, redefining it as “the disaggregation of the legacy SaaS market”. Software Equity Group (SEG), which publishes the SEG SaaS Index of public SaaS companies, described a “sharp early-year selloff” in its report on the second quarter of 2026. For how those valuation multiples are calculated, see SaaS valuation multiples.

Agents and seats in plain terms

Think of a toll road that charges per car. If a company replaces five delivery vans with one larger truck, the same goods travel the road, and the toll collects one fee instead of five. The road’s owner then needs a different way to charge, such as by weight.

Two terms carry the debate:

Per-seat pricing: you pay a fixed amount per named user per month. Revenue grows when the customer hires.
AI agent: software that carries out multi-step tasks across systems on someone’s behalf, not only answering questions.

Gartner names the effect “agentic arbitrage”, which “happens when AI agents complete tasks across multiple systems”.

The case that agents shrink SaaS

Gartner’s release of July 1, 2026 puts a number on the risk. It says up to $234 billion of enterprise application spending is “exposed to agentic arbitrage between now and 2030”, roughly 20% of enterprise application SaaS spending by 2030.

George Brocklehurst, Managing Vice President at Gartner, gives the mechanism: “Agentic systems deliver outcomes directly, bypassing traditional user experience (UX)-heavy applications.” On revenue, he says agentic AI “breaks the link between user growth and revenue growth for many enterprise software vendors.” He adds that “the user interface is no longer a differentiation.”

Valuations fell in the same period. SEG reports that the median EV/TTM revenue multiple of its index fell from 5.7x in the second quarter of 2025 to 3.2x in the second quarter of 2026. The multiple compares a company’s enterprise value with its revenue over the trailing twelve months. SEG says the reset was broad as AI reshapes software, and that the market now reserves premium valuations for companies with strong operating performance and defensible advantages.

The case that SaaS changes shape instead

The same Gartner release argues against reading this as an ending. It says “SaaS will not be destroyed; it will emerge in a different form”, and Brocklehurst calls the shift “less an apocalypse and more of a metamorphosis.” His argument for incumbents rests on data: “Better outcomes from AI require systems that can retain deep institutional memory and customer context over time.” On that argument, vendors that already hold years of customer records start ahead.

Incumbents are also selling agents themselves. Salesforce reported on August 26, 2026 that revenue for its fiscal second quarter of 2027 was $11.3 billion, up 11% year on year. It said Agentforce annual recurring revenue (ARR) “exceeded $1.5 billion, up over 240% Y/Y.” The same release notes that from that quarter, Agentforce ARR also counts Salesforce’s AI offerings, Slackbot and Headless 360, so the figure is not directly comparable with earlier quarters.

Pricing is moving too. Salesforce’s Agentforce pricing page lists $2 per conversation and Flex Credits at $500 per 100,000, noting that “Agentforce actions are 20 Flex Credits”. It also lists a per-user add-on at $125 per user per month. Seats remain on the list, next to meters for the work an agent does.

Example: fewer seats, new meters

Example (hypothetical vendor, real reference rate): a vendor charges $100 per seat per month, and a customer holds 50 seats.

Current bill: 50 x $100 = $5,000 a month.
Agents take over routine work, and the customer renews with 30 seats: 30 x $100 = $3,000.
Revenue lost: $5,000 - $3,000 = $2,000 a month, or $2,000 / $5,000 = 40%.
Suppose the vendor meters agent work at Salesforce’s published Flex Credit rate: $500 / 100,000 = $0.005 per credit, and 20 credits per action is 20 x $0.005 = $0.10 per action.
Actions needed to replace the lost revenue: $2,000 / $0.10 = 20,000 a month, or 20,000 / 30 = about 667 a day.

The arithmetic shows why both sides can be right about different customers. A customer whose agents run 667 actions a day pays the vendor what it paid before. A customer whose agents run 100 a day pays $3,000 + 100 x 30 x $0.10 = $3,300, which is 34% less than before.

What to watch, without a forecast

Neither side’s claim can be checked today; the evidence arrives in later filings and price lists. Four signals are public:

Signal	Where to find it	What it would show
Seat counts and net revenue retention	Vendors’ quarterly filings	Whether customers buy fewer seats at renewal
New meters on pricing pages	Vendors’ pricing pages	Whether vendors charge for agent work beyond seats
AI revenue figures and their definitions	Earnings releases, footnotes	Growth, and whether definitions changed
Index multiples	SEG and other index reports	How investors price the change

For how seat-based and usage-based pricing compare in general, see SaaS pricing models.

What to ask at your next renewal

Whichever side turns out closer, your contract is where the debate reaches you. Ask the vendor:

How is agent work billed: by seat, conversation, action or credit?
What is the rate per unit, and is there a monthly cap on usage charges?
If you cut seats because agents do the work, will agent usage cost more than the seats you save?
Are AI features included in your plan, or sold as add-ons with their own meter?
Which of your records does the agent read, and can you export them if you leave?

Common mistakes

Treating a share-price fall as proof. Lower multiples show what investors expect, not what customers have done. Why it happens: market headlines arrive before filings. Instead: read seat counts and retention in the vendors’ own reports.
Comparing AI revenue across quarters without the footnotes. Salesforce widened what counts as Agentforce ARR in the quarter it reported $1.5 billion. Why it happens: the definition sits below the headline figure. Instead: read the definition before comparing growth rates.
Assuming fewer seats means a smaller bill. In the example, a customer running 667 agent actions a day pays the same $5,000 on 30 seats. Why it happens: seat counts are the familiar unit. Instead: model both meters at renewal and ask for caps on usage charges.
Reading “exposed” as “lost”. Gartner’s $234 billion is spending “exposed to agentic arbitrage” up to 2030, not revenue already gone. Why it happens: large numbers travel without their verbs. Instead: quote the forecast with its wording, its date and its horizon.

Sources

Gartner, Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI, July 1, 2026, checked October 2026.
Software Equity Group, Quarterly SaaS report, 2Q26, checked October 2026.
Salesforce, Salesforce Delivers Record Second Quarter Fiscal 2027 Results, August 26, 2026, and Agentforce pricing, checked October 2026.

In three sentences

The case against SaaS is that AI agents work across many applications, so customers need fewer seats; Gartner says up to $234 billion of enterprise application spending is exposed by 2030. The case for SaaS is that vendors holding customer data are selling agents and adding usage meters, as Salesforce’s $1.5 billion Agentforce ARR and $0.10-per-action Flex Credit rate show. Whether a 40% seat cut becomes a loss depends on how much metered agent work replaces it, so watch seat counts, pricing pages and definitions rather than share prices.

AI SaaS examples: 10 products and what their AI actually does

2026-10-29T07:00:00-04:00

Every SaaS vendor now says its product has AI, and your renewal quotes show new AI lines. You need to see, product by product, what the AI does and how you pay for it. Then you can compare like with like, and spot where you are paying twice.

AI SaaS, here, means software delivered as a service, as explained in What is SaaS?, in which a language model or image generator does part of the work. The ten products below are well-known SaaS tools that added AI, not a ranking. For the wider list of SaaS products, see SaaS examples.

AI in SaaS in plain terms

Think of buying a car. Some models come with cruise control as standard. Others sell it in a package for a fixed price. A few charge you by the mile for using it. The feature is similar in all three; what changes is how it shows up on the bill.

AI features follow the same three patterns:

Included: the AI comes with the plan or a higher tier.
Add-on: a fixed fee per user on top of the plan.
Metered: you pay by usage, counted in credits (prepaid units converted into AI actions) or in finished results.

Ten products and their AI

Every fact comes from the vendor’s own page, checked on October 17, 2026. AI plans change often; re-check before you buy.

Product	What the AI does	How it is priced
Microsoft Copilot	Assistant in Microsoft 365 apps; work-based chat over data the user’s account can access	Add-on license for eligible Microsoft 365 plans (Microsoft 365 E7 includes it); web-based Copilot Chat included at no extra cost
Google Workspace	Gemini in Gmail on Business Starter; Gemini in Gmail, Docs, Sheets, Slides, Vids, Drive, Meet and Chat on Standard and Plus	Included in the plan; access widens by tier
Slack	Conversation summaries, daily recaps and AI search on every plan; AI workflow generation on Business+	Included in the per-user plan; higher Slackbot limits on Business+ invoice plans need Flex Credits bought through Salesforce
Notion	Notion Agent for multi-step tasks, AI Meeting Notes, Enterprise Search (beta)	Trial on Free and Plus; included on Business at $20 per member a month; Custom Agents at $10 per 1,000 Notion credits
Atlassian Rovo	Search, chat and agents across Jira, Confluence and connected tools	Included with Standard, Premium and Enterprise Cloud plans; Jira and Confluence get 25, 70 or 150 credits per user a month; $0.01 per extra credit
GitHub Copilot	Code completion, chat, and an agent that works on a task and opens a pull request	Free, Pro $10, Pro+ $39 and Max $100 a month; chat and agent use drawn from AI Credits at $0.01 each
Figma	Figma Make turns a prompt into a prototype or web app; AI image editing	AI credits per seat: 3,000 a month on a Professional Full seat, 3,500 on Organization, 4,250 on Enterprise
Intercom	Fin AI Agent answers customers and completes workflows	$0.99 per outcome, on top of seats at $29, $85 or $132 a month
Salesforce Agentforce	AI agents inside Salesforce’s sales, service and other apps	Flex Credits at $500 per 100,000 (20 credits per standard action); add-ons at $125 per user a month; or $2 per conversation
HubSpot	Customer, prospecting, data, content and revenue agents, plus an agent builder	Agent Hub included in Starter, Professional and Enterprise; many agents run on HubSpot Credits

What the list shows

Suites put AI in the seat. Google, Slack and Notion’s Business plan include their AI in the per-user price, with more of it in higher tiers. The cost is hidden in the upgrade.

Tools for makers meter with credits. GitHub, Figma and Atlassian give each seat an allowance and price what goes beyond it. Figma warns that the number of credits a request uses “varies by request”.

Support and CRM agents charge for work done. Intercom bills per outcome and Salesforce per action or conversation. These agents act for your customers, so their volume follows your customer base, not your headcount.

The words change as fast as the prices. On October 17, 2026, Microsoft’s licensing page said the Microsoft 365 Copilot app “is now called Microsoft Copilot”. On the same day, GitHub’s plans page metered usage in AI Credits, and GitHub Docs had moved the older “premium requests” billing, at $0.04 a request, to pages marked legacy. Read the plan page and the docs on the same day, and keep the date.

Worked example: how far an included allowance goes

Example: a labeled hypothetical using Atlassian’s published numbers. Your company has 25 people on Jira Standard. Each gets 25 Rovo credits a month, and Atlassian pools all allowances in an organization into one shared pool. A basic AI interaction costs 10 credits.

Monthly pool: 25 users x 25 credits = 625 credits.
Basic interactions covered: 625 / 10 = 62.5, so 62 interactions for the whole company, or about 2.5 per person a month.
Suppose the team uses 300 basic interactions: 300 x 10 = 3,000 credits.
Credits over the allowance: 3,000 - 625 = 2,375. At $0.01 each: 2,375 x $0.01 = $23.75 a month.

Two readings follow. The included allowance covers very little regular use, so “included” does not mean “unlimited”. But the overage is small next to the seat price, as long as extra usage stays turned on. With it turned off, Atlassian pauses the credit-consuming features until the monthly reset, and your team loses the feature mid-month instead.

Checklist: before you pay for an AI feature

Find the unit: seat, tier, credit, action, conversation or outcome.
Find the conversion: how many credits one task uses.
Check whether allowances pool across the company or stay with each user.
Check what happens at the limit: a pause, or an overage bill.
Check the data terms by tier. Notion lists zero data retention with its AI model providers only on Enterprise.
Check whether a tool you already pay for includes the same AI. Google Workspace Standard already includes Gemini in Docs and Gmail.
Write down the date you checked each price.

For how one of these suites is priced overall, see Microsoft 365 as SaaS and Google Workspace as SaaS.

Common mistakes

Comparing seat prices and ignoring credits. Two tools at the same seat price can differ widely once usage starts. Why it happens: the seat price is the headline on the pricing page. Instead: estimate a month of real use and price it in each tool’s unit.
Reading “included” as “unlimited”. In the example, an included allowance covers about 2.5 basic interactions per person a month. Why it happens: “included” is the word on the plan card. Instead: find the allowance and the limit behavior before you roll the feature out.
Paying twice for the same AI. A team buys a separate writing assistant while its suite already includes one. Why it happens: different teams buy different tools. Instead: list the AI each existing tool includes before approving a new one.
Relying on an old price or plan name. Microsoft renamed its Copilot app, and GitHub moved Copilot from premium requests to AI Credits. Why it happens: last quarter’s comparison sheet still circulates. Instead: date every price and re-check it before each purchase or renewal.

Sources

All checked October 17, 2026:

Microsoft Learn, License options for Microsoft Copilot.
Google Workspace, Pricing; Slack, Pricing; Notion, Pricing.
Atlassian, Rovo pricing; GitHub, Copilot plans and Copilot requests (legacy); Figma, Pricing.
Intercom, Pricing; Salesforce, Agentforce pricing; HubSpot, AI agents.

In three sentences

Ten well-known SaaS products ship AI in three ways: included in the plan (Google, Slack, Notion), as a per-user add-on (Microsoft, Salesforce), or metered by credits and outcomes (GitHub, Figma, Intercom). In the example, 25 Jira Standard users share 625 Rovo credits, enough for about 62 basic AI interactions a month, with overage at $0.01 a credit. Compare AI features by their unit, allowance and limit behavior, check what your existing tools already include, and date every price you rely on.

AI pricing in SaaS: credits, usage and outcome-based models compared

2026-10-27T12:00:00-04:00

Your team has shipped an AI feature, and finance wants to know how to charge for it. A flat seat price ignores that every AI request costs money to run. A pure meter scares buyers who want a predictable bill. You need to know the models vendors use, and what each one does to your customer’s invoice and to your margin.

AI pricing is how a SaaS vendor charges for features that run on machine-learning models: assistants, agents and generators. It sits on top of the subscription basics in What is SaaS?. The general price structures (seat, tier, usage, freemium) are covered in SaaS pricing models. This guide covers only the AI layer.

AI pricing in plain terms

Think of it like a phone plan. Some plans include calls in the monthly fee. Others give you a bundle of minutes and charge for extra. A few charge only when the call connects. The network costs the carrier money every minute either way. The plan only decides who carries that cost.

AI features work the same way. Each request to a model has a running cost, called inference. A credit is a prepaid unit that the vendor converts into actions at a published rate. An outcome is a finished result, such as a resolved support conversation, that the vendor defines and counts.

The five models vendors use

Every example below comes from the vendor’s own page, checked on October 17, 2026. AI prices change often, so check the date before you rely on one.

Model	What you pay for	Public example	Who carries the usage risk
Bundled in the seat	The per-user price, AI included	Slack lists AI conversation summaries, recaps and AI search inside its per-user plans, with “Advanced AI” on Business+ (Slack)	Vendor
Per-user add-on	A fixed monthly fee per person on top of a base plan	Salesforce lists Agentforce add-ons at $125 per user per month (Salesforce); Microsoft sells Copilot as an add-on to eligible Microsoft 365 plans (Microsoft Learn)	Vendor, for heavy users
Allowance plus credits	A monthly allowance, then a price per extra credit	Atlassian Rovo includes 25 credits per user a month on Jira and Confluence Standard, then $0.01 per credit (Atlassian); GitHub Copilot Pro at $10 a month lists $15 of AI Credits ($10 base plus $5 flex) at $0.01 each (GitHub)	Shared
Per action	Each step an agent takes	Salesforce Flex Credits: $500 per 100,000 credits, 20 credits per standard action, which works out to $0.10 an action	Customer
Per outcome	Each result the vendor counts as done	Intercom prices Fin at $0.99 per outcome (Intercom)	Customer pays per result; vendor absorbs failed attempts

The rows run from most predictable for the buyer to most closely tied to use. Salesforce shows that one vendor can sell several at once: the same page lists per-user add-ons, Flex Credits and a $2 per conversation option.

What each model means on an invoice

Bundled and per-user prices give your customer a fixed bill. The vendor pays for every heavy user out of the same fee. That works while average use stays low.

Credits split the risk. The included allowance covers normal use, and the meter starts above it. Two details decide how this feels in practice. First, whether allowances pool: Atlassian combines an organization’s allowances “into a single shared pool”. Second, what happens at the limit: with extra usage turned off, Atlassian pauses credit-consuming features until the monthly reset. GitHub lets Business and Enterprise admins decide whether paid usage beyond the allowance is allowed; if it is not, Copilot pauses until the next cycle.

Per-action pricing bills the agent’s work, not its success. An agent that takes eight steps to do a three-step job costs more than twice as much.

Per-outcome pricing bills results, so the definition of a result is the contract. Intercom counts an outcome when the customer confirms the issue is resolved, when they do not ask for more help after Fin responds, or when Fin completes a workflow, including handoffs. It charges once per conversation. Its option for teams on another help desk adds a minimum monthly commitment, “e.g. 50 outcomes”.

Worked example: one product priced three ways

Example: a labeled hypothetical. A help-desk vendor sells an AI agent to a customer with 20 support agents. The AI handles 3,000 conversations a month and resolves 1,800 of them, or 60%. Each conversation takes 5 model actions on average, and each action costs the vendor $0.04 in inference.

Vendor’s inference cost: 3,000 x 5 x $0.04 = $600 a month, in every option.
Option A, per-user add-on at $30: 20 x $30 = $600. AI margin: $600 - $600 = $0.
Option B, per action at $0.10: 3,000 x 5 = 15,000 actions x $0.10 = $1,500. Margin: $1,500 - $600 = $900, or 60%.
Option C, per outcome at $0.90: 1,800 x $0.90 = $1,620. Margin: $1,620 - $600 = $1,020, or 63%.

Now change one input at a time.

Scenario	A: per user	B: per action	C: per outcome	Vendor cost
Base: 3,000 conversations, 60% resolved	$600	$1,500	$1,620	$600
Volume doubles: 6,000 conversations, 60% resolved	$600	$3,000	3,600 x $0.90 = $3,240	$1,200
Quality drops: 3,000 conversations, 30% resolved	$600	$1,500	900 x $0.90 = $810	$600

Read the rows. When volume doubles, option A loses $600 a month ($600 - $1,200), because the seat price does not move with use. When quality drops, option B still bills $1,500, while option C falls to $810 and leaves the vendor $210. For the buyer, the cost per resolved conversation in the base case is $600 / 1,800 = $0.33 under A, $1,500 / 1,800 = $0.83 under B and $0.90 under C.

So the seat model protects the buyer’s budget and exposes the vendor. Per-action pricing exposes the buyer to an inefficient agent. Per-outcome pricing ties the vendor’s revenue to the agent’s quality, but only as far as the outcome definition is honest.

Questions to ask before you buy or set an AI price

What exactly is one unit: a credit, an action, a conversation or an outcome? Get the definition in writing.
How many credits does one task use? Salesforce charges 20 credits for a standard action and 30 for a voice action.
Do included allowances pool across users, or does each person keep their own?
At the limit, does the feature pause or bill overage? Who decides?
Is there a minimum commitment?
Can the conversion rate change? Figma’s pricing page says “Included AI credits are subject to change”, and Salesforce marks its page as subject to change.

For how these meters fit into renewals and proration, see SaaS subscription.

Common mistakes

Bundling AI into the seat without measuring use. In the example, doubling volume turns a break-even add-on into a $600 monthly loss. Why it happens: seat prices are easy to sell and easy to forecast. Instead: meter AI use per account internally, even if you do not bill for it.
Comparing credit prices without the conversion rate. One cent per credit means nothing until you know how many credits a task uses. Why it happens: the credit price is on the pricing page, the conversion is in the docs. Instead: convert every offer to a cost per finished task.
Accepting an outcome definition you have not read. Intercom’s definition counts a conversation where the customer does not ask for more help, and a completed handoff. Why it happens: “pay only for results” sounds safe. Instead: sample billed outcomes each month and check that they were resolved.
Ignoring what happens at the limit. A paused agent in the middle of a busy week can cost more than the overage. Why it happens: limits only matter once you hit them. Instead: decide in advance whether a pause or an overage bill hurts more, and set the budget control to match.

Sources

Intercom, Pricing, checked October 17, 2026.
Salesforce, Agentforce pricing, checked October 17, 2026.
Atlassian, Rovo pricing, checked October 17, 2026.
GitHub, Copilot plans, checked October 17, 2026.
Slack, Pricing; Figma, Pricing, checked October 17, 2026.
Microsoft Learn, License options for Microsoft Copilot, checked October 17, 2026.

In three sentences

SaaS vendors charge for AI in five ways: bundled in the seat, as a per-user add-on, as an allowance plus credits, per action, or per outcome. In the example, the same agent earns $600, $1,500 or $1,620 a month depending on the model, and doubling usage turns the seat add-on into a $600 loss. Before you buy or set a price, convert every offer to a cost per finished task and read the unit definition and the limit behavior in writing.

AI copilots in SaaS apps: what they are, how they work and what they cannot do

2026-10-26T07:00:00-04:00

The apps you already pay for now have a sidebar that drafts emails, summarizes meetings and answers questions about your files. You need to know what it can see, how far you can trust it and whether the extra seat price is worth paying. An AI copilot is an assistant built into a software product that uses a language model to help a person do a task in that product. The person reviews the result and decides what happens next.

That last sentence is the line between a copilot and an agent, and it decides most of the risk. This guide covers how copilots work, what they typically do, what they cannot do, and how to check the seat cost before you buy. For the basics of the software they sit in, see What is SaaS?.

Copilots in plain terms

Think of a co-driver in a rally car. The co-driver reads the notes, calls the next bend and warns about the ditch. The driver keeps both hands on the wheel. If the co-driver misreads a note, the driver can still brake.

A copilot is the co-driver. It suggests, drafts and summarizes; you accept, edit or ignore. An agent, by contrast, takes the wheel for a stretch. Microsoft’s AI agent shared responsibility model describes an agent as one that “invokes tools, calls APIs, writes data, and triggers workflows without a human approving each step”.

GitHub’s description of GitHub Copilot shows both modes in one product. For assistive features, “You review and apply each suggestion.” For its agentic features, where Copilot works through several steps toward a goal, “You remain responsible for reviewing and approving.”

How a copilot works

Microsoft’s data, privacy and security page for its Microsoft 365 assistant (updated October 8, 2026) names three parts, a pattern most copilots share:

A large language model that reads and writes text.
Your organization’s content, “such as emails, chats, and documents that you have permission to access”.
The app you are working in, such as Word or PowerPoint.

The page also records a rename: Microsoft 365 Copilot is now called Microsoft Copilot, and some licenses may keep the old name during the transition.

The second part is the one to understand. Before the model answers, the copilot retrieves content that matches your request and adds it to the prompt. This is called grounding. Microsoft states that its copilot “only surfaces organizational data to which individual users have at least view permissions”. It also states that prompts, responses and data accessed through Microsoft Graph “aren’t used to train foundation LLMs”. Other vendors make their own commitments, so read each vendor’s terms rather than assuming the same.

What copilots typically do

Microsoft’s Copilot overview (updated October 14, 2026) lists in-app features. Together with GitHub’s, they show the usual pattern: draft, summarize, analyze, explain.

App	What the copilot does (vendor’s description, checked October 2026)	Who acts on the result
Word	Draft, rewrite and summarize documents	You
Excel	Analyze data, create formulas and visuals	You
Outlook	Draft emails and summarize threads	You send
Teams	Summarize and transcribe meetings, capture action items	You follow up
GitHub	Inline code suggestions; answers about files and repositories	You review and apply
GitHub, agentic features	Works through several steps, edits files, prepares pull requests	You review and approve

The last row is where a copilot product starts to behave like an agent. The approval step is still yours, but more happens before it.

What copilots cannot do

Guarantee accuracy. Microsoft’s own page says the responses that generative AI produces “aren’t guaranteed to be 100% factual”, and that users should use their judgment before sending output to others. A summary can drop the one sentence that mattered. A formula can look right and reference the wrong column.

Fix your permissions. A copilot that respects permissions also respects bad ones. If a salary spreadsheet is shared with the whole company, the copilot can surface it to anyone who asks the right question. Microsoft points admins to SharePoint Advanced Management, which “helps you reduce oversharing”, for this reason. The copilot did not create the exposure; it made it easy to find.

Know what it was not given. A copilot sees what it retrieves for that request. Content in a system it is not connected to does not exist for it.

Act on its own, by design. That is a feature. When you need actions taken without review, you are buying an agent, with a different risk profile.

The seat cost, worked through

Copilots are usually sold as a per-user add-on. Prices change often, so check the vendor’s page on the day you buy.

Example: GitHub lists Copilot Business at $19 per granted seat per month and Copilot Enterprise at $39 (checked October 17, 2026). A team of 40 developers, with a hypothetical loaded cost of $75 an hour per developer:

Business, yearly: 40 x $19 x 12 = $9,120.
Enterprise, yearly: 40 x $39 x 12 = $18,720.
Break-even on Business: $19 / $75 = 0.25 hours, about 15 minutes saved per developer per month.
Break-even on Enterprise: $39 / $75 = 0.52 hours, about 31 minutes per developer per month.

The seat price is a floor, not a cap. GitHub’s plans documentation says each Business seat adds 1,900 AI Credits to a shared pool. Use beyond the pool is billed at $0.01 a credit if your admins allow it. The break-even is low, which is why the real question is usage, not price. If only 25 of the 40 developers use it, the Business cost per active user is $19 x 40 / 25 = $30.40 a month. Measure active use for a month or two before you buy seats for everyone. For how seat-based licenses work in general, see SaaS licensing.

Common mistakes

Rolling out before cleaning up sharing. The copilot surfaces whatever each person can already open. Why it happens: oversharing was invisible until search got this good. Instead: review sharing on sensitive sites and files before switching the copilot on.
Treating output as fact. Drafts and summaries can be wrong in ways that read well. Why it happens: the text is fluent and confident. Instead: check anything that leaves your team, and keep the human review step in every process.
Buying seats for everyone on day one. Unused seats raise the real cost per active user. Why it happens: the vendor’s bundle price looks simpler. Instead: start with a pilot group and use the vendor’s usage reports to decide on the rest.
Calling a copilot an agent, or the reverse. The approvals and permissions you need differ. Why it happens: vendors use both words for the same product. Instead: ask who acts on the output and whether a person approves each action.

For how Microsoft’s and Google’s suites package their tools, see Microsoft 365 as SaaS and Google Workspace as SaaS.

Sources

Microsoft Learn, Data, privacy and security for Microsoft Copilot and What is Microsoft Copilot?, checked October 17, 2026.
Microsoft Learn, AI agent shared responsibility model, checked October 17, 2026.
GitHub Docs, What is GitHub Copilot? and GitHub Copilot plans, checked October 17, 2026.

In three sentences

An AI copilot is an assistant inside a SaaS app that drafts, summarizes and analyzes using your data, while a person reviews and acts on the result. It sees what you already have permission to open, so oversharing becomes visible, and its output is not guaranteed to be accurate. Seat prices such as GitHub’s $19 a month break even at minutes of saved work, so measure who actually uses it before you buy seats for everyone.