One Update Could Wipe Out Your Company in 2027

The cost of running the most powerful AI models may drop in 2027, only for some companies that built their businesses on top of these models to discover that their profits haven't improved. This paradox is the key to understanding the next phase of the Agentic AI market. A company doesn't profit simply because it bought cheaper intelligence, but when it can turn that intelligence into work executed correctly, at a cost lower than the price paid by the customer.

Tech companies themselves have started changing their pricing models in ways that reveal this shift. In Microsoft Copilot Studio, a generative answer costs two Copilot Credits, while an action executed by an agent consumes five credits. Additional costs may pile up on top of that when the agent uses advanced reasoning models or extra workflows. In short, writing an answer is cheaper than actually doing something with it inside the system.

Salesforce goes a step further, selling 100,000 Flex Credits for $500, with a standard action executed by Agentforce consuming 20 credits—about ten cents per action based on published pricing. It also offers pricing at $2 per conversation in certain cases. The economic message is clear: the market is shifting from billing the customer for 'talking to AI' to billing them for the work it performs.

The trend is even clearer with Intercom, which started pricing its agent, Fin, starting at $0.99 per outcome, rather than per message or token count. If it resolves a support issue or executes a specific action successfully, the company earns its fee. If it fails and the customer escalates to a human agent, the outcome isn't billed the same way. Intercom states that Fin resolves an average of 76% of conversations across more than 8,000 customers, with some reaching over 90%. These are company-reported numbers, but they reveal where the business model is moving.

Here lies the real danger for an entire generation of startups.

Many Agentic AI companies emerged because models, no matter how powerful, were unable to operate inside an enterprise on their own. A model needed something to connect it to email, customer databases, accounting software, file repositories, and payment systems. There had to be a software layer that saves task state, sequences steps, calls tools, and retries if something breaks. Developers called this layer orchestration. Just a year ago, many investors saw this as a prime gap for building a company, but that gap is now narrowing rapidly.

In its Agents API, OpenAI manages sessions, orchestration, context compression, and work restoration, while supporting tool calling and connections to MCP servers, as well as running agents within custom execution environments. What used to require a suite of third-party products is gradually being baked into the core infrastructure developers already buy.

Google is taking the same path. Gemini Enterprise Agent Platform brings models, development environments, memory, code execution, and governance into a single platform. It even grants agents independent digital identities through which it enforces permissions, monitors system access, and logs activity. When an agent gains a recognized 'job identity' inside an organization, there is far less a startup can charge for merely managing the connection between the model and client software.

Meanwhile, Anthropic helped turn the Model Context Protocol – MCP into a widespread standard for connecting AI to tools and data. The company stated in January 2026 that the protocol reached 100 million monthly downloads. As connecting to tools becomes standardized and easier, the economic value of a company whose sole secret was knowing how to make a model talk to another software program plummets.

Here, an investor must separate two things that look similar in a demo but are completely different economically. Connecting Claude to Salesforce is a technical feat. But making Claude process a refund request correctly—in compliance with company policy, employee permission tiers, amount limits, and audit requirements—is an entire business, and that is where the profit lies.

If you ask a model to write an apology email, you can read the text and judge it. But if you give it the power to refund a customer's money, modify a financial record, approve a discount, or screen a customer according to KYC procedures, an error moves off the chat screen and directly into the general ledger or compliance file.

Then, the agent needs more than just a good model. It needs to know who holds approval authority, which edge cases demand human intervention, what needs logging, how to roll back an action if an error occurs, and which legal rule applies to each case. This layer is what will create the 'execution economy.'

Take a simple example: if a task execution attempt costs ten cents, and the agent succeeds in 80% of attempts, the direct cost per successful task becomes 12.5 cents—before adding retries, tool calls, employee time spent resolving failed cases, and potential losses from wrong decisions.

The problem escalates quickly when a task consists of sequential steps. If each step has 95% accuracy, that might sound excellent. But for a task made up of ten steps where every single step must succeed, the chance of completing it end-to-end drops to just around 60%. This is a simple mathematical result: (0.95^{10}) equals roughly 0.60.

Here, our whole understanding of technical progress changes. A model improving from 95% to 97% accuracy per step pushes the overall success rate for a ten-step task from about 60% to nearly 74%. A few percentage points in per-step reliability can therefore translate into a massive gain in actual output.

This is why tracking token prices is not enough.

A model's price could drop by half while the final task cost remains high because the agent keeps retrying, pausing for approvals, or routing edge cases to a human worker. Thus, intelligence might look extremely cheap on an API bill, while the product itself remains low-margin.

Google is treating this as an operational infrastructure problem. In May 2026, it launched Project Agent Executor because tasks running over hours or days become fragile, requiring work resumption after failures or human approval delays. In other words, one of the world's largest computing giants reached the same conclusion: the challenge is no longer making the agent 'think', but ensuring it reaches the finish line.

This point also explains Gartner's pessimism regarding parts of the market. The research firm predicted that over 40% of Agentic AI projects will be canceled by the end of 2027 due to spiraling costs, unclear business value, or inadequate risk controls. In a more recent forecast, it noted that 40% of enterprises might reduce agent autonomy or pause them altogether due to post-deployment governance issues.

These numbers do not imply a total crash of the agent market. More likely, they point to a brutal shakeout separating different types of companies.

The most vulnerable category consists of companies that buy API calls from OpenAI, Anthropic, or Google, add an interface, some prompts, and connectors, and mark up the price. The underlying provider is constantly advancing behind them, building memory, tool integration, execution, and governance directly into its platform. With every new update, a feature for which the intermediary charged separately could simply disappear.

Conversely, a company selling an economic outcome stands in a very different position. A collection agent can charge based on successfully recovered funds. A customer service agent can be priced per resolved issue. A procurement agent's fee can be tied to the volume of orders correctly processed according to policy.

The concept of outcome-based pricing has become a core part of industry discourse. Sierra explicitly states that customers should pay for results rather than usage, while Intercom built a major portion of Fin's pricing around this very principle.

This model radically shifts incentives. When a company charges per million tokens, its revenue increases the more the client uses the model. But when it charges a dollar per completed task, it profits by executing that task with the fewest possible retries, tokens, and human interventions.

Can efficiency itself become the product?

In 2027, a company using an average model might outperform one using the most powerful model on the market. If the former completes 90% of workflows without human intervention while the latter stalls at 70%, the economics of the first company will usually win out—even if its underlying model scores lower on intelligence benchmarks.

This also redefines competitive moats. Great prompts, polished UI, or a set of connectors will no longer suffice. Value will lie in deep domain expertise of a specific process: how an insurance claim is handled, how a loan application is vetted, how an invoice is closed, how a tax exception is resolved, and exactly when an agent must step back and hand off to a human.

This knowledge compounds through live execution. Every edge case the system encounters enriches its decision library. Every human intervention can later be turned into a rule, test case, or new workflow branch. This builds the proprietary data layer that a general model cannot easily replicate on its own.

Therefore, 2027 could see a clear divide. Horizontal startups built solely around model orchestration and tool integration will see their value erode, turning into add-ons, consulting services, or acquisition targets. In contrast, value will accrue to companies that own an end-to-end business process and master its unit economics, error modes, and edge cases.

Investors looking to profit from AI will need to evaluate companies by new metrics.

Revenue growth, customer retention, and gross margins will remain important. But a new metric will become far more telling for Agentic AI companies: the true cost per successful outcome.

This figure must include model and tool costs, retries, compute overhead, human intervention, and error remediation. Only then do the true unit economics of the company emerge.

If this cost drops as models improve, the company owns an engine that transforms technological leaps directly into expanding profit margins.

However, if model prices decline year after year while the cost to complete a task stays unchanged, the product may impress in a demo, but the investor is actually looking at a cheap interface sitting atop an expensive execution engine.