KM4BI

DoorDash Is Testing Lunch by Text—A Clue to How You’ll Buy Software in 2027

Agents can make an application’s interface disappear without erasing its value—and shift lock-in toward data, workflows, and accumulated context.

Published

DoorDash’s text-ordering experiment points to a future in which customers consume a service without opening the application that provides it. Whether someone messages DoorDash directly or invokes it through another assistant, the company wants to remain the party that receives and fulfills the order. Losing the interface does not necessarily mean losing the customer, transaction, or revenue.

When the agent replaces the interface

The same shift is appearing at work. An agent can read a new customer’s contract, review sales notes and email threads, and produce a kickoff plan. It can query a candidate-data provider, apply a recruiter’s criteria, and prepare a shortlist for a hiring manager.

In both cases, the employee gradually stops opening the specialized tool that used to present the information. The underlying data remains valuable, but the interface and part of the information-processing workflow move into the agent. Software whose proposition loosely combines finding, arranging, presenting, and acting on information becomes vulnerable when none of those components is uniquely defensible.

A different kind of lock-in

The agent itself may not be easy to replace. Over time, employees teach it rules that were never properly documented: what a successful candidate looks like, what a particular hiring manager prefers, which security approval must precede a customer kickoff, or how a team actually uses its systems.

That knowledge may live in memory, custom skills, or a personal harness built around one AI provider. If finance or technology leadership later wants to switch vendors for lower token costs or greater optionality, it may fail to see the weeks of work and operational context embedded in individual workflows.

Data, applications, and agents

The video separates the software stack into three layers. The **data layer** contains structured information, customer records, and other inputs agents need. The **application layer** supplies workflows, rules, approvals, and dependable execution over that data. The **agentic layer** interprets context and coordinates the work.

This framework explains why different categories of software face different levels of risk. A small recruiting firm may keep paying for structured candidate data while building its own search workflow with the provider, rather than buying another packaged tool that mainly digests and displays that data. Payroll is different: configured rules, records, approvals, compliance, and reliable execution are foundations few companies want to recreate. An agent may make that system more useful by identifying problems and explaining complexity without replacing the machinery that ensures people are paid correctly.

Connectivity is table stakes

DoorDash’s text experience and limited MCP beta illustrate the need to be reachable from whatever environment the customer chooses. That accessibility is only an entry requirement, however. It does not explain why a customer should keep paying. If two providers offer equally useful information and equally dependable execution, agent connectivity may make them easier to compare and replace.

Durable value must instead come from hard-to-reproduce data, trusted execution, domain-specific workflows, or a combination of layers that produces better outcomes. Large platforms are already trying to stack these advantages: existing distribution, enterprise data, an agent control plane, and specialized intelligence. At the same time, support for multiple models can reassure buyers who do not want to be trapped by one AI vendor.

What buyers, employees, and sellers should do

Business leaders cannot understand this transition by making a high-level software decision and simply placing “AI” over the existing stack. They need to investigate how teams actually work. AI usage resembles a fingerprint: each group combines data, applications, automations, and agents differently. Standardizing without that knowledge can remove productivity that leadership did not know existed.

Employees should explain how they use AI, which data sources matter, and what context would be lost in a migration. Keeping a productive technique secret makes the workflow brittle if the company replaces its provider. Software vendors, meanwhile, need to show that their data is accessible, their workflows can work with multiple agents, and their value will remain relevant when the next model arrives.

By 2027, companies will not merely be buying features or an AI label. They will be buying the ability to produce measurable outcomes through a flexible stack—even when the visible application interface is no longer the center of the relationship.

Source