Which AI tools will people still pay for in 2027?

Why is it suddenly harder to know which AI tools people will keep paying for?

AI spending is still climbing fast, but the number of AI products that deserve their own subscription is already starting to shrink.

That sounds contradictory until we look at what has changed. Consumer demand for AI remains huge. Sensor Tower's latest State of AI report estimated that global in-app spending on generative-AI apps would exceed $4 billion during the first half of 2026, up 36% from the previous six months. Time spent inside these apps was projected to more than double year over year, from 17.2 billion to roughly 36 billion hours.

At the same time, AI has become almost impossible to avoid. RevenueCat's latest dataset found that 27.1% of subscription apps already qualify as AI-powered. The share reaches 61.4% in Photo & Video and 41.1% in Productivity. Sensor Tower separately found more than 200,000 mobile apps mentioning AI in their descriptions.

That abundance changes the buying decision. Two years ago, access to a good generative model could itself feel like a product. Today, ChatGPT, Gemini, Claude, Microsoft 365, Notion, Adobe, GitHub and dozens of other platforms overlap across writing, research, coding, images, files and agents.

The large platforms are also bundling aggressively. Notion currently includes its core AI features, AI Meeting Notes and Enterprise Search inside Business and Enterprise plans. Adobe includes generative credits throughout Creative Cloud. GitHub gives paid Copilot users unlimited code completions while charging for more expensive agent activity. Microsoft is putting Copilot across products companies already use every day.

People will still spend heavily on AI in 2027. The harder question is how many separate subscriptions survive once the same capabilities appear inside products users already pay for.

Are people paying for AI now, and are they actually sticking around?

Yes, people are spending meaningful money on AI today, but weak retention suggests a large share of current subscriptions will not survive unchanged into 2027.

Sensor Tower estimates that generative-AI mobile apps were on track to generate more than $4 billion of in-app purchase revenue in only six months. ChatGPT became the fastest mobile app to reach one billion monthly active users, while Google's latest reported figure puts the Gemini app at 950 million monthly active users.

There is also evidence that spending is spreading beyond ChatGPT. Sensor Tower measured Claude's US mobile revenue per user rising from less than $0.50 to $2.76 between September 2025 and May 2026. ChatGPT's share of the generative-AI mobile audience fell below 50% for the first time during 2026 as Gemini and Claude gained ground.

The interesting part is how pricing has held up. Midjourney still charges $10, $30, $60 and $120 a month. Cursor ranges from $20 to $200 for individual developers. General AI subscriptions commonly sit around the familiar $20 price point, while heavier users can spend far more.

The warning appears when we look at retention. RevenueCat's State of Subscription Apps 2026 covers more than 115,000 subscription apps. AI apps produced a median $30.16 in realized first-year value per payer, 41% above the $21.37 for non-AI apps. They also converted downloads into paid subscriptions slightly better.

After 12 months, however, only 6.1% of monthly AI subscriptions remained active at the median, versus 9.5% for non-AI apps. Annual AI plans retained 21.1% of subscribers, compared with 30.7% outside AI. AI apps also had a higher median refund rate, 4.2% versus 3.5%.

AI subscriptions are also unusually concentrated in monthly billing: 59.8% of AI subscriptions use monthly plans, compared with 26.2% for non-AI apps. Only 24% of AI subscriptions are annual, versus 41.8% outside AI.

So the market has already proved that consumers will pay. What it has not proved is that they will keep paying for the same products once better substitutes arrive.

RevenueCat median	AI apps	Non-AI apps
Download-to-paid conversion	2.4%	2.0%
First-year value per payer	$30.16	$21.37
12-month retention, monthly plan	6.1%	9.5%
12-month retention, annual plan	21.1%	30.7%
Refund rate	4.2%	3.5%

Will people still pay $20 a month for ChatGPT, Claude or Gemini in 2027?

Yes. General AI assistants should remain a huge paid category in 2027, although paying for several of them at once will make less and less sense for most consumers.

The reason is breadth. A ChatGPT, Claude or Gemini subscription can already replace pieces of a writing app, search tool, tutor, research assistant, coding assistant, translator, image tool and file analyzer. General assistants keep swallowing use cases that once belonged to separate products.

Their audiences are also reaching internet-platform scale. OpenAI now says ChatGPT serves more than one billion weekly users. Google reports 950 million monthly Gemini app users, with daily users tripling over the previous year.

Scale alone does not guarantee subscriptions, especially because the free tiers are getting better. OpenAI has even added advertising as another way to monetize free ChatGPT usage, with the company recently saying ChatGPT Ads reached a $1 billion annualized revenue run rate.

This makes the paid proposition clearer. Heavy users will pay for capacity, stronger models, long-running agents, richer context, privacy features and the ability to make the assistant work across more of their lives. Casual users will increasingly have excellent free options.

Model leadership also changes too quickly to create durable loyalty by itself. A subscriber who pays because Claude writes better today can move when Gemini or ChatGPT gets ahead. Personal history, projects, files and recurring workflows create much more friction.

We therefore expect one premium general assistant to remain a normal purchase for heavy users. Paying for three similar ones should stay mostly a power-user habit.

Will developers still pay for AI coding tools in 2027?

Yes, with unusually high confidence. AI coding has already become too useful, too frequent and too economically valuable to collapse into a free feature.

Anthropic disclosed that Claude Code reached a $1 billion run-rate only six months after becoming publicly available. Microsoft now reports 50 million GitHub Copilot users, and GitHub Copilot revenue accelerated by more than 60% quarter over quarter in Microsoft's latest reported results.

Independent developer research makes the change look even bigger. JetBrains surveyed more than 15,000 professional developers and found that 90% were using AI coding agents at work at least weekly during May-July 2026, while 68% were using them daily.

Claude Code's workplace adoption rose from 18% in January to 39% by May-July. Codex climbed from 3% to 16% over the same period. GitHub Copilot fell from 29% to 21% and Cursor slipped from 18% to 12%, which shows how quickly individual leaders can change even while the category itself gets stronger.

Developers are also delegating substantial amounts of actual coding. Among people who named Claude Code as their main AI coding tool, JetBrains found 32% were generating more than 80% of their code with agents. The figure reached 42% among Codex users.

Current pricing confirms that vendors see heavy coding as something worth charging for. Cursor sells individual plans at $20, $60 and $200 per month. GitHub currently offers a free tier, a $10 Pro plan and a $39 Pro+ plan, then meters more demanding AI usage through credits.

A developer costing a company thousands of dollars a month does not need an AI agent to save many hours before a $20, $40 or even $200 bill looks trivial. That basic arithmetic gives coding tools far more room to charge than an AI app that helps someone rewrite two emails a week.

Individual winners will keep changing. Paid AI coding itself looks extremely durable.

Will companies keep paying for Microsoft 365 Copilot and enterprise AI search?

Yes, although companies are increasingly paying for AI that understands their own organization rather than for a generic chatbot attached to every employee.

Microsoft provides the strongest evidence. Paid Microsoft 365 Copilot seats rose from 15 million to more than 20 million and then above 30 million over three consecutive reported quarters. Net seat additions more than doubled quarter over quarter in the latest period.

These are increasingly large deployments. Accenture reached more than 740,000 Copilot seats. EY subsequently deployed Microsoft's E7 offering to 400,000 employees. Bayer, Johnson & Johnson, Mercedes and Roche have each committed to at least 90,000 seats.

Microsoft's latest product and pricing decisions are just as interesting as the seat count. The company says it is moving beyond pure per-seat pricing toward per-seat-plus-consumption. Thousands of customers are already paying usage charges for its Cowork product.

That fits what we see elsewhere. Notion includes Agent, Meeting Notes and Enterprise Search in its Business and Enterprise packages, then charges credits when Custom Agents run scheduled or triggered work. Glean has grown from $100 million to more than $300 million ARR in only 15 months while selling enterprise context, search and agents. More than 85% of Glean customers deploy the product across at least five departments.

Company context is becoming the valuable layer. Microsoft says its Work IQ system spans more than 17 exabytes across documents, communications, meetings and other organizational information. Glean connects knowledge across many business applications while preserving permissions. Notion's advantage is that work already lives inside the workspace.

A generic model can get dramatically smarter without knowing which employee can see a confidential document, which customer record is authoritative or what happened in last week's private meeting. That gives enterprise context real staying power.

Current enterprise signal	Latest reported scale
Microsoft 365 Copilot	More than 30M paid seats
GitHub Copilot	50M users
Glean	More than $300M ARR
Glean deployment depth	85%+ of customers across 5+ departments
Microsoft Work IQ context	More than 17 exabytes

Can paid AI search survive when ChatGPT and Google already search the web?

Yes, but we are much less confident about standalone AI search than about coding or enterprise AI.

The basic consumer feature is getting squeezed from every direction. ChatGPT searches the web inside a general assistant. Google's search product increasingly answers questions with Gemini-powered AI. Claude can research across the web. Browsers themselves are becoming agentic.

A person who occasionally wants an answer with sources has less reason each year to buy a separate search subscription.

Professional research is different. Someone investigating 50 companies, comparing filings, searching private documents, revisiting previous work and building a report can save enough time to justify a specialist product. Premium databases and proprietary information make that case stronger again.

That is why the category keeps stretching beyond search. Products such as Perplexity increasingly emphasize browsers, agents and research workflows rather than a simple answer box. Glean started in enterprise search and expanded into assistants and agents. General assistants are doing the same from the opposite direction.

By 2027, paid AI search probably survives mainly where "search" is only the first step in a larger research workflow.

Which simple AI tools are most likely to become free or disappear?

Generic AI writing, summarization, file chat, transcription and basic image generation face the hardest 2027 pricing environment because larger platforms can reproduce these features cheaply and distribute them instantly.

Look at writing first. Gmail, Google Docs, Microsoft Word, Notion, ChatGPT, Claude and Gemini can all rewrite or draft text. A standalone tool whose main feature is "make this paragraph sound better" has to justify another subscription against alternatives that users already have open.

Meeting notes are heading in the same direction. Notion includes AI Meeting Notes in its Business and Enterprise plans. Microsoft can summarize inside Teams. Google can do the same around Meet and Workspace. Dedicated meeting tools can still work, but they need to turn conversations into something more valuable: CRM updates, sales coaching, institutional memory, compliance records or automated follow-up.

Simple file chat has an even weaker moat. General assistants can read PDFs, spreadsheets, presentations and documents directly. The old proposition of uploading a PDF and asking questions has become a standard AI interaction.

Basic image generation is also becoming plentiful. Adobe now gives users unlimited access to selected standard image generation inside some paid plans while reserving credits for more demanding features. General AI assistants can produce images directly. Midjourney still has strong paid plans, but image quality alone becomes a dangerous advantage when competing models leapfrog one another every few months.

Transcription follows the same path. Accurate speech-to-text used to be the whole product. Today, the commercially interesting part begins after transcription: understanding the conversation, updating another system, identifying risk, generating a medical note or performing an action.

If a user can recreate most of the product by opening a general assistant and typing one prompt, charging separately will be difficult.

Will lawyers still pay for specialized legal AI when general models get better?

Yes. Legal AI looks increasingly like one of the strongest specialist categories because lawyers need much more than a clever answer.

The latest adoption data is hard to dismiss. Thomson Reuters found that 41% of law firms and 47% of corporate legal departments now use generative AI, up from 28% and 23% respectively in the previous year's survey.

Harvey has also moved well past experimental use. The company says 80% of the Am Law 100 uses its platform, alongside five Fortune 10 companies. Harvey users have built more than 25,000 custom workflows, and its agents now receive more than 400,000 legal queries per day.

Those numbers help explain why general-model improvements have not wiped out the specialist layer. Legal teams want work grounded in trusted material, connected to their own documents and repeatable across firm-specific workflows.

Thomson Reuters' latest professional research makes the trust requirement unusually clear. Among surveyed professionals, 96% said AI tools need to protect confidential information, 94% said outputs should be grounded in authoritative content and 90% wanted reasoning that could be explained and defended.

Those requirements favor products connected to Westlaw, Practical Law, internal matter data, document-management systems, firm templates and established security controls.

Harvey's newest financing provides another fresh indication of how investors and customers see the category. The company recently raised $550 million at a $15.5 billion valuation, only months after raising at $11 billion.

Valuations can get ahead of reality, so we would never use that number as proof of product value on its own. The combination of 80% Am Law 100 penetration, hundreds of thousands of daily agentic queries and 25,000-plus customer-built workflows is much stronger evidence.

Legal AI should still command serious budgets in 2027.

Will hospitals keep paying for AI medical scribes?

Yes, with high confidence. Ambient clinical AI is one of the clearest cases where the product saves enough unpleasant, expensive work to support long-term spending.

Microsoft says its healthcare AI is on pace to automate more than 100 million patient encounters during the calendar year, including 28 million in its latest quarter, twice the volume from a year earlier. Mass General Brigham has rolled Dragon Copilot out to more than 4,000 providers after a study found a 21% reduction in burnout.

Abridge says more than 300 enterprise health systems representing over 250 million patients have adopted its context-aware clinical decision-support capability. Whatever we make of vendor-reported reach, deployments are clearly moving beyond small physician trials.

Independent research helps quantify the value. A study in pediatric hematology-oncology examined 11,544 outpatient encounters. Among repeat users, shorter appointments using Microsoft's DAX Copilot required a median 11 minutes of active documentation compared with 24 minutes without it. Adjusted analysis found roughly 30% lower documentation time.

Another Stanford study examined 8,740 emergency-department encounters, including 976 where an ambient scribe was used. That is useful because it measures voluntary use in an actual hospital rather than a polished demo.

The remaining problem is accuracy. A study of 1,003 UK general practitioners found that 80% of current ambient-scribe users reported spending less time on documentation and 70% reported lower cognitive load. Yet 32% said errors occurred often or always, and 14% reported encountering errors with potentially significant-to-critical implications.

That error rate means hospitals still need verification, governance and careful deployment. It also gives specialized clinical systems more room to justify their price through EHR integration, audit trails and safety controls.

Will companies really keep paying for AI customer-service agents?

Yes. Customer-service AI has one of the clearest paths to durable payment because companies can count what the agent resolves.

Intercom's Fin currently starts at $0.99 for an outcome. More than 12,000 teams use Fin, and Intercom says the average resolution rate has reached 76%.

Zendesk has independently arrived at almost the same economic model. Its AI agents are billed through "automated resolutions," meaning customer requests completed without escalation to a human. The company introduced new resolution tiers in 2026 to reflect different levels of value produced by the AI.

Salesforce takes a related approach. Agentforce can charge through Flex Credits, where a standard action costs about $0.10, or through a $2-per-conversation model.

Three major customer-service vendors moving toward actions, resolutions and outcomes tells us much more than another chatbot launch.

Imagine a company with 100,000 support conversations. If an agent successfully handles 60,000 conversations at roughly $1 each, the AI bill comes to about $60,000. The company can compare that directly with the cost of staffing those interactions, including salaries, management, hiring, training and 24-hour coverage.

The important caveat is measurement. Vendors and customers will fight over what qualifies as a resolution, especially when a user returns later or needs a human anyway. Outcome pricing only works when the outcome is genuinely measurable.

Still, customer service sits very close to the ideal 2027 AI business: repeated work, high volume, clear incumbent labor cost and an output the customer can count.

Will people still pay for AI image and video generators?

Yes, although we expect much more pricing pressure in images than in high-end video and professional creative workflows.

Generic image generation is already becoming abundant. Adobe bundles standard generation into Creative Cloud and Firefly plans. General assistants can create images. Open and proprietary models keep improving. A consumer who simply wants ten good-looking pictures has plenty of options.

Yet creators continue to pay for products such as Midjourney, which currently charges $10 to $120 per month. Adobe also sells Firefly tiers with increasing credit pools, while premium generation such as video consumes those credits. Heavy use still costs real compute.

Video makes the payment case stronger. Producing good video remains more computationally demanding, and professional users care about consistency, editing, control and throughput. A tool that saves a production team days of work can support much higher pricing than one that generates a single social-media image.

The workflow around the model matters too. Adobe sits inside Photoshop, Illustrator, Express and the rest of Creative Cloud. Canva combines AI with templates, brand assets, collaboration and publishing. A creator may switch image models while staying inside the same production environment.

Professional creative AI should remain paid. Standalone image-generation subscriptions face much tougher competition.

Will AI voice and avatar tools still deserve separate subscriptions?

Yes, especially when companies use AI voice and video to replace recurring production or customer-service costs.

ElevenLabs gives us one of the strongest recent examples. The company ended 2025 with $350 million ARR and crossed $500 million within the first four months of 2026. ElevenLabs says much of that growth is coming from companies deploying voice agents in support, sales, hiring and marketing.

That increase represents roughly $150 million of additional ARR in four months, or more than 40% growth from the year-end base. The scale makes it difficult to explain the business as people paying to make funny voices.

Synthesia shows a similar shift in enterprise video. The company says contracts worth more than $100,000 tripled over a 12-month period and net revenue retention exceeded 140%. It also says its product is used by 90% of the Fortune 100.

The business case is straightforward when a company regularly creates onboarding, training or multilingual communications. Recording a presenter, hiring a studio, editing footage and reproducing the same material across many languages is expensive. Updating an AI-generated training video can be far cheaper.

Voice agents have a similar denominator: human calls.

We expect basic text-to-speech and simple avatar generation to become common features. The durable paid products will handle voice identity, permissions, low-latency conversations, phone infrastructure, localization, consistent avatars, enterprise governance and production at scale.

A good synthetic voice alone will stop feeling scarce. Running thousands of useful conversations with that voice is a much bigger business.

Will outcome-based pricing replace the normal monthly AI subscription?

Partly. AI pricing is already moving toward a mix of subscriptions and usage because one flat seat price makes less sense once agents start doing very different amounts of work.

The change has become visible across several unrelated products.

Microsoft says it is moving Microsoft 365 Copilot beyond per-seat pricing toward a combination of seats and consumption. Its Dynamics business is moving in the same direction, and usage-based customer-service credits rose fourfold quarter over quarter in the latest reported period.

GitHub changed Copilot from its older premium-request system to usage-based AI Credits. Paid plans still include an allowance, while expensive models and long agent sessions consume more of it.

Notion includes core AI inside Business and Enterprise plans but charges Notion credits for Custom Agents that run continuously or respond to triggers. Intercom prices Fin by successful outcomes. Zendesk bills automated resolutions. Salesforce charges for Agentforce actions or conversations.

These companies are converging because two employees with the same software seat can generate completely different AI costs. One may use a few summaries a week while another runs long autonomous tasks.

Customers also want spending to follow usage more closely. Paying hundreds of fixed AI seats that employees barely use is difficult to defend.

Monthly subscriptions should survive, but expensive AI work will increasingly be metered on top.

Which AI tools will people still pay for in 2027?

People will still spend heavily on AI in 2027, but the strongest money will flow toward a small number of general assistants and specialist tools that perform expensive work inside real workflows.

The clearest survivors are AI coding agents, enterprise copilots with company context, legal AI, clinical AI scribes, customer-service agents and professional voice or video systems. In each case, we can already see either very large paid adoption, unusually fast revenue growth, measurable labor savings or a direct link between usage and valuable output.

General assistants such as ChatGPT, Claude and Gemini should also remain huge subscription businesses. Their problem is competition with one another, so many consumers will probably choose a primary ecosystem instead of subscribing to every leading model.

AI research tools can survive among professionals, although free search from Google and general assistants makes the consumer case weaker.

Creative AI should split. Serious video, voice and production workflows still have enough compute cost and economic value to charge. Basic image generation gets harder to sell by itself.

Generic AI writing, summarization, transcription, PDF chat and simple wrappers look much more vulnerable. Their core outputs have already become standard features elsewhere, and users have almost no reason to tolerate another subscription when switching takes seconds.

The RevenueCat data gives us a useful warning here. AI apps currently generate 41% more first-year value per payer than non-AI apps while retaining materially fewer subscribers after 12 months. Lots of AI products can make somebody reach for a credit card once. Far fewer have proved that they deserve to stay on it.

By 2027, intelligence will be available almost everywhere. The products that keep getting paid will usually have one of four things: valuable context, a workflow users already depend on, genuinely scarce capacity or a measurable amount of human work removed.

AI tool category	2027 payment outlook	Why people will still pay
AI coding agents	Very strong	Direct developer productivity and intensive daily use
Clinical AI scribes	Very strong	Documented time savings and deep clinical integration
Legal AI	Very strong	Trusted sources, private data, workflow and accountability
Customer-service agents	Very strong	Resolutions can be measured against human labor
Enterprise copilots/search	Strong	Company context, permissions and connected systems
General AI assistants	Strong	Broad utility, personal context and premium capacity
AI voice agents	Strong	Replaces recurring calls and production work
Professional AI video	Strong	Expensive compute plus production savings
Professional AI research	Moderate to strong	Valuable for deep, repeatable research workflows
AI design suites	Moderate to strong	Workflow and editing protect them from model switching
Meeting intelligence	Moderate	Stronger when connected to CRM and downstream actions
Standalone image generators	Moderate	Users pay today, but model switching is easy
Standalone writing assistants	Weakening	Writing is bundled almost everywhere
Basic transcription tools	Weakening	Speech-to-text is becoming a standard feature
Simple PDF/chat tools	Weak	General assistants already handle files well
Generic AI wrappers	Very weak	Little differentiation and almost no switching cost