Which AI wrapper ideas can still work in 2027?

Can a simple AI wrapper still survive in 2027?

A simple AI wrapper can still make money in 2027, but we would no longer build one with the expectation that the product itself will stay defensible for long.

The problem has become much more obvious lately. OpenAI, Anthropic, Google and Microsoft keep absorbing features that startups once sold separately. Search, file analysis, long context, web browsing, code execution, memory, connectors, scheduled tasks and multi-step agents are increasingly part of the base platforms.

OpenAI pushed this further with its new Agents API, which lets developers run cloud agents through a managed Codex environment. The platform handles long-running sessions, tools, sandboxes and subagents. Anthropic has moved in the same direction with Claude Cowork. Cowork can work across files, use a built-in browser, run scheduled tasks and load plugins containing skills, connectors and subagents. Anthropic even offers plugin categories for jobs such as legal, sales and finance.

These are exactly the kinds of capabilities that many AI startups were packaging themselves two years ago.

Yet the application layer is also producing some extraordinary businesses. Lovable passed a $500 million annualized revenue run rate. ElevenLabs crossed $500 million ARR. Glean reached $300 million ARR. Sierra reached $200 million ARR in nine quarters. Clay reached roughly $100 million ARR after tripling in a year.

That apparent contradiction explains most of the wrapper debate. Building on someone else's model clearly does not doom a company. The dangerous part is having too little left once we strip the model away.

For 2027, we would ask a brutally simple question: if OpenAI, Anthropic or Google improves the underlying model again, does our product become better or become unnecessary?

Are cheaper AI models changing how wrappers need to make money?

Cheaper AI models are destroying the markup on raw intelligence, so the strongest wrappers are moving their pricing toward completed work rather than model usage.

The fall in inference costs has been enormous. Stanford's AI Index calculated that the cost of achieving roughly GPT-3.5-level performance fell by more than 280 times between late 2022 and late 2024. Since then, competition among OpenAI, Anthropic, Google, xAI, Chinese labs and open-model providers has kept pushing price-performance forward.

We can already choose between models with surprisingly similar practical capabilities but radically different prices. That makes it difficult to charge customers a huge premium merely for sending text to a model and returning the answer.

But an application handling a valuable business process can benefit from exactly the same trend.

Salesforce currently charges Agentforce customers $2 per conversation or, under its Flex Credits system, roughly $0.10 per standard action. Those prices are tied to work performed inside a business system rather than the amount of model computation consumed underneath.

Suppose an insurance workflow previously cost $15 in employee time and an AI system can finish it reliably for $2. Whether the vendor's underlying inference expense is $0.40 or $0.08 barely changes the purchasing decision. The customer sees a large saving, while the vendor benefits from falling model costs.

Traditional per-seat pricing also fits automation awkwardly because good agents can reduce the number of humans touching a workflow. That is why we expect more products to charge per task, conversation, resolution, document or transaction, often alongside annual minimum commitments.

The important question is what the customer is actually buying. "Five million tokens" exposes the vendor to infrastructure price competition. "Ten thousand resolved claims" ties the price to something the customer already values.

What the customer buys	Exposure to cheaper models	2027 outlook
Generated text, summaries or answers	Very high	Poor
Generic AI assistance	High	Weak
AI inside a business workflow	Moderate	Good
Completed business actions	Lower	Very good
Regulated or financially important outcomes	Low	Strongest

Will OpenAI and Anthropic eventually copy every successful AI wrapper?

OpenAI and Anthropic will copy far more wrapper features than most founders would like, but they are unlikely to operate every specialized workflow themselves.

The threat is particularly high these days because frontier labs keep climbing the application stack.

OpenAI already provides managed agents, tools, file handling, web access, coding environments and connectors. Anthropic's plugin system can turn Claude into a specialist for sales, finance, legal work and other functions. Cowork can operate a browser and run recurring tasks. Microsoft is embedding agents into its software ecosystem, while Salesforce is pushing Agentforce across sales, service and internal workflows.

A startup offering one generic capability above the model increasingly sits directly in their path.

The situation changes once an application goes several layers deeper into a specific business process. Consider legal software. Generating a contract clause is easy for a frontier model. Running an enterprise legal workflow means dealing with client matters, document-management systems, access permissions, precedents, citations, jurisdiction, audit trails, confidentiality rules and human approval.

Healthcare creates an even larger operational burden. Producing a transcript is one thing. A system used across a hospital network has to fit the EHR, identify the right medical context, create usable documentation, respect access rules, support billing workflows and survive procurement and security reviews.

We would therefore be particularly cautious with anything that looks like a plausible feature inside the next version of ChatGPT or Claude. Products become safer when copying them would require the foundation-model company to take on the messy operating requirements of a specific industry.

Why are some AI wrappers reaching hundreds of millions in revenue anyway?

Some AI wrappers are growing incredibly fast because customers are paying for a complete job rather than for access to the underlying model.

The numbers have become difficult to dismiss.

Menlo Ventures estimated that enterprises spent $19 billion on generative-AI applications in 2025, out of $37 billion spent across the broader enterprise generative-AI market. It also counted at least 50 AI products above $100 million in ARR.

Lovable is an unusually clear example. The company passed a $500 million annualized revenue run rate after reporting $400 million only a few months earlier. It now says its platform hosts around 60 million projects attracting roughly 900 million monthly visits.

Glean provides a very different example. It reached $300 million ARR only 15 months after passing $100 million. Its customers are buying access to company knowledge, permissions, people, applications and workflows through one context layer.

ElevenLabs crossed $500 million ARR after ending the previous year around $350 million. According to the company, a growing part of that demand comes from enterprises deploying voice agents for customer support, sales, hiring and marketing.

These businesses look very different on the surface, but in each case the model sits inside a larger system involving workflow, context, deployment or distribution. That is why the label "wrapper" is becoming less useful on its own.

Are vertical AI wrappers the safest bet for 2027?

Vertical AI looks like the strongest wrapper category for 2027 when the software can take over expensive work inside a messy industry.

Enterprise spending already points in that direction. Menlo Ventures estimated vertical generative-AI applications at $3.5 billion in 2025, nearly three times the previous year's level. Healthcare alone represented roughly $1.5 billion, or 43% of the category.

Healthcare is especially revealing because adoption has moved well beyond experimental chatbots. Menlo estimated the ambient-scribe market alone at roughly $600 million. Abridge says its technology has been deployed across more than 150 enterprise health systems and was expected to support around 50 million conversations during 2025.

Abridge had already reached about $117 million in contracted ARR in early 2025. More interestingly, the company kept expanding beyond transcription into areas such as medical coding. Its early integration with Epic also placed Abridge directly inside the workflow clinicians already use.

Legal AI shows a similar progression. Harvey passed $100 million ARR, then kept growing as large law firms expanded deployments. The underlying models are available elsewhere, yet large legal organizations still pay for software built around matters, documents, permissions, security and legal workflows.

Insurance, accounting, tax, procurement and logistics share many of these characteristics. Their processes mix messy documents, rules, approvals, legacy software and expensive human labor.

The best opportunities are narrower than "AI for insurance" or "AI for dentists." We would look for one recurring job with a real budget attached to it and enough operational complexity that a general assistant cannot finish it reliably.

AI wrapper opportunity	What makes it harder to copy	Our 2027 view
Healthcare administration	EHRs, compliance, medical context, billing workflows	Excellent
Legal workflow automation	Matters, permissions, confidentiality, jurisdiction	Excellent
Insurance claims and underwriting	Legacy systems, rules, documents, financial exposure	Excellent
Accounting and tax operations	Records, audit trails, changing rules	Strong
Procurement and logistics	Fragmented systems, approvals, transactions	Strong
Generic industry chatbot	Very little beyond prompting	Weak

Can AI customer-service wrappers still win?

AI customer service can still support huge businesses, although a startup that merely answers support questions has arrived too late.

Sierra shows how far the category has moved. The company reached $200 million ARR in nine quarters and said more than 40% of the Fortune 50 were customers. Its agents handle tasks ranging from mortgage refinancing to insurance claims and retail returns.

Those examples involve actions. Answering "Where is my refund?" is easy. Authenticating the customer, checking the order, deciding whether the refund is allowed, executing it and updating the relevant systems is much harder.

Salesforce is already attacking the same market from inside the CRM. Its current Agentforce pricing spans per-action, per-conversation and per-user options. Existing help-desk vendors are doing the same.

A new customer-service startup therefore needs a narrower entry point or a much deeper product. Financial services, travel disruptions, insurance claims, healthcare administration and complex B2B support all look more attractive than another generic website chatbot.

Can AI sales wrappers still work?

AI sales wrappers can work very well when they decide who to contact, gather the right information and trigger the workflow; another tool for writing personalized emails has very little room left.

Clay is a useful example because the language model handles only part of the job. Clay connects to hundreds of enrichment sources, can waterfall between providers, research prospects with AI and push the resulting data into systems such as Salesforce and HubSpot.

OpenAI itself used Clay to raise inbound lead-enrichment coverage from roughly 40% to 80%. Anthropic has also used Clay for lead enrichment and routing.

Clay reached approximately $100 million ARR after tripling in a year. That growth came during a period when writing a personalized sales email became almost free.

Generating "Hi Pierre, I saw your company just..." is trivial these days. Determining which company is worth contacting, finding the correct buyer, reconciling incomplete databases, understanding what changed at the account and feeding the result back into the CRM still requires a real workflow.

That leaves room for account research, enrichment, qualification, territory planning, pipeline inspection and closed-loop sales automation. Standalone AI outreach writers look much weaker.

Can AI coding wrappers survive Claude Code and OpenAI Codex?

AI coding wrappers can still become enormous, but generating code by itself is quickly becoming the least interesting part of the product.

Coding gives us one of the strongest tests of the wrapper thesis because model companies have attacked this market aggressively. Anthropic has Claude Code. OpenAI has Codex and now runs cloud agents through its managed Codex harness. Google and others are pushing their own coding systems.

At the same time, independent application companies continue to grow.

Menlo estimated enterprise spending on AI coding products at roughly $4 billion in 2025, more than half of all departmental generative-AI spending. Lovable recently passed $500 million in annualized revenue. Replit has also reported explosive growth as it moves from an online development environment toward autonomous software creation.

Lovable is particularly revealing because many of its users are not professional programmers. They describe an application and expect Lovable to help produce something they can actually use. The company increasingly handles the surrounding environment too: projects, backend services, hosting and deployment.

A thin coding extension that forwards repository context to a frontier model has a difficult future. A product that turns a business requirement into working software, tests it, deploys it and keeps it running owns much more of the outcome.

Can AI voice wrappers still work in 2027?

AI voice wrappers can still grow extremely fast, especially when voice is attached to a transaction or business workflow.

ElevenLabs provides the clearest scale data. The company finished 2025 at roughly $350 million ARR and surpassed $500 million during the first four months of 2026. That means it added more than $150 million of ARR in roughly four months.

The source of that growth is becoming more interesting than the speech model itself. ElevenLabs says enterprise customers are deploying voice agents across support, sales, hiring and marketing. The company has also expanded well beyond basic text-to-speech into conversational agents, dubbing, creative tools and other audio products.

Raw voice generation is becoming crowded. OpenAI, Google, ElevenLabs, Deepgram and other vendors can all produce increasingly natural speech.

A dental-office agent offers a better business if it can answer calls, identify patients, schedule appointments, handle cancellations and interact with the practice-management system. The same logic applies to restaurants, home services, automotive dealerships, insurance, healthcare and financial services.

For 2027, workflow-specific voice agents look considerably stronger than standalone voice generation.

Do data, integrations and distribution really protect an AI wrapper?

Data, integrations and distribution can make an AI wrapper much harder to replace when they put the product inside a workflow competitors cannot easily recreate.

Glean is one of the better examples. The company has built hundreds of connections into enterprise applications and keeps permissions attached to the information it indexes. A user can search or run AI across company knowledge without casually exposing documents they were never allowed to see.

Glean reached $300 million ARR after tripling from $100 million in 15 months, while its Fortune 500 customer count nearly doubled year over year.

The harder part lies in everything around the search. Enterprise information changes constantly. Access rights vary by employee. One fact may live in Slack, another in Salesforce and the authoritative version in an internal database. Agents also need to know which source they are allowed to act on.

Clay has a related advantage in go-to-market data. Abridge has it through health-system workflows and EHR integrations. Harvey builds around legal documents, permissions and matters.

Founders should still be careful with the phrase "proprietary data." Saving chat histories does not create much protection. Putting customer PDFs in a vector database barely qualifies either. A stronger data advantage appears when every completed workflow produces information that improves future decisions.

Distribution adds another layer. Microsoft can put AI in front of Microsoft 365 users. Salesforce starts from the CRM. Google controls Workspace. OpenAI can distribute applications through ChatGPT, while Anthropic can surface specialist tools through plugins and connectors.

For an independent startup, the place where the workflow happens becomes extremely valuable. An accountant would rather see AI inside the ledger than visit another chatbot. A salesperson wants research inside the CRM. A clinician wants documentation to appear in the medical record.

The best products combine context, integrations and distribution around the same moment of work.

Can services and human review make an AI wrapper stronger?

Human-heavy deployment can make an AI wrapper stronger when each deployment teaches the company how to automate the next one.

This is happening across enterprise AI today. Sierra uses forward-deployed engineers. Harvey has invested heavily in implementation. Google Cloud and Accenture announced an initiative involving up to 1,000 forward-deployed engineers for Gemini Enterprise deployments.

The reason is fairly mundane. Real companies contain terrible data, unofficial processes, old software, undocumented exceptions and employees who disagree about how the workflow actually works.

A team may initially need to sit with the customer, map the process, connect systems, build evaluations, define approval rules and watch where the agent fails. If those lessons become reusable software, the deployment work strengthens the product.

The economics become unattractive when every new customer requires another permanent consulting project. We would want implementation effort per customer to decline over time even while contract sizes rise.

Are generic AI productivity wrappers still worth building?

Generic AI productivity wrappers look increasingly weak because ChatGPT, Claude, Gemini and Microsoft Copilot now cover too much of the same ground.

Menlo estimated horizontal generative-AI application spending at $8.4 billion in 2025. Around 86% was captured by general-purpose copilots.

Those assistants have expanded aggressively since then.

Claude Cowork can work with files, browse websites, schedule recurring work and use plugins. Anthropic currently offers plugins that can search across company email, chat, cloud storage, wikis, project-management tools, CRM systems and ticketing platforms.

OpenAI's agents can also work across tools and files and perform longer jobs. Microsoft and Google have the additional advantage of already living inside office software.

That leaves much less room for a new destination app whose pitch is "AI for documents," "AI for research," "AI for email" or "AI for productivity."

There are still exceptions. A horizontal product can win through a unique context layer, a strong collaborative workflow, unusual distribution or a deeply loved interface. But a generic assistant needs a much better reason to exist today than it did two years ago.

Can tiny AI wrappers still be good businesses?

Small AI wrappers can still be excellent businesses even when they would never survive the standards applied to a billion-dollar venture-backed company.

A founder selling a $100-a-month tool to 2,000 specialized businesses has a $2.4 million annual revenue company. If the operation is lean and customer retention is good, that can be a wonderful outcome.

Narrow markets can help because the founder can understand the customer's language, acquire users through specialized channels, answer support personally and adapt faster than a large software company cares to.

Property managers, specialist medical practices, freight brokers, small law firms, laboratories, independent insurance agencies and tradespeople still contain ugly manual processes.

A small wrapper does not need to prove that OpenAI could never copy it. It needs enough workflow, distribution and customer loyalty to remain worth paying for.

Which AI wrapper ideas should founders avoid in 2027?

We would avoid AI wrappers whose entire advantage can be reproduced by opening a frontier assistant, attaching the same file and writing one reasonably good instruction.

Generic writing tools sit high on that list. So do simple summarizers, PDF chat, basic research agents, prompt marketplaces, generic email assistants and undifferentiated meeting summaries.

The warning has expanded to basic agent builders too. OpenAI's new Agents API already provides managed agent execution, sandboxes, tools, long-running sessions and subagents. Anthropic packages skills, connectors and subagents directly into plugins. Salesforce and Microsoft are doing the same inside their enterprise ecosystems.

Model weaknesses are another dangerous foundation for a company. Products built around "the model cannot browse," "the context window is too short," "the model cannot remember previous work" or "the model cannot use tools" have repeatedly watched those limitations disappear.

We would also be skeptical of products with impressive demos but no obvious recurring workflow. AI makes demos unusually easy to build. The harder test is whether somebody comes back next week and lets the software do the job again.

AI wrapper idea	Our 2027 view
Generic AI writer	Avoid
PDF chatbot	Avoid
Prompt marketplace	Avoid
Generic research assistant	Avoid
Basic email-writing assistant	Avoid
Undifferentiated meeting summarizer	Avoid
Thin multi-agent builder	Avoid
Vertical agent completing real work	Attractive
Deeply integrated operational agent	Very attractive
Outcome-based AI workflow	Very attractive

So which AI wrapper ideas can still work in 2027?

AI wrappers can absolutely still work in 2027, but we would put almost all of our attention on products that own the work surrounding the model.

Regulated vertical agents look strongest. Healthcare, legal, insurance, tax and financial operations combine large labor budgets with workflows that general assistants struggle to complete safely on their own.

Operational agents come next. Customer service, sales operations, procurement, logistics and back-office workflows become interesting when the software can actually change records, make decisions, move information between systems and finish tasks.

AI-native creation platforms still have room as well. Lovable's current growth shows that even aggressive competition from coding models does not eliminate a product that moves users much closer to a finished application.

Voice also remains attractive when attached to a real workflow. The opportunity is much better in handling appointments, customer problems, claims or transactions than in generating realistic speech by itself.

Enterprise context is another strong layer. Glean's rise to $300 million ARR shows how valuable permissions-aware organizational knowledge can become when AI agents need reliable context before acting.

And there will be thousands of smaller wrappers serving narrow markets profitably. Many will never build deep technical moats, but strong distribution and close workflow fit can still create durable cash-generating businesses.

Our favorite test for a 2027 AI wrapper is simple: remove the foundation model from the company description and look at what remains.

If the answer is a difficult workflow, customer relationships, integrations, proprietary operational data, distribution and responsibility for a valuable outcome, there may be a serious company there.

If almost nothing remains beyond the prompt and the interface, we would keep looking.