Scope
Book a demo

Step 1 of 2

Book a demo

Your work email first, then pick a time.

All posts

ResearchNov 18, 20267 min read

How Do AI Agents Choose Which Software to Use?

In 77% of runs where an AI agent picked a vendor, it named the pick before any search result came back. What decides that pick, and can the agent get in?

The next user to try your product may be an AI agent, not a person. Box CEO Aaron Levie expects that to become the norm: “AI is going to be the biggest consumer and user of software going forward,” he told Computerworld. So what happens when that user reaches your product?

Our previous articles explored where agent experience (AX) breaks. In our testing of Parallel (web research APIs), Kapa (AI assistants for technical docs), and Linear (issue tracking), agents finished 99% of tasks, as measured by our automated checks, when explicitly told to use the product. They did so even though they hit a broken docs page or an API error in about one in three runs.

Completion is only the second half. An agent has to choose a product before it can use one, and nothing about that choice is guaranteed. Here is what decided both, and why most of it never reaches your dashboard.

Key findings

  • Agents named their pick before any search result came back in 77% of the runs that picked a vendor.
  • Parallel was the pick in 83% of runs for deep research and came up in 11% for web search, which it also sells.
  • Asked for an issue tracker, Codex recommended Jira in every run and Claude Code recommended Linear in all but one.
  • With the customer’s API key, agents reached the product’s API in 81% to 100% of runs. Where signup needed a person, none of 59 runs finished setup.

First, the AI agent has to think of you

Before an agent can choose a product, it has to surface it as an option. In our runs, which companies surfaced depended on what the agent was asked for.

Asked to recommend a deep-research API, agents picked Parallel in 83% of runs. Asked to add live web search to an app, they built with Tavily in 78% of runs and with Parallel in 6%. Parallel came up at all in only 11%.

Parallel sells a web search API too. But in these runs, agents tied Parallel to deep research. For web search, the company they tied to it was Tavily.

When agents explained a pick, the usual reason was that the vendor’s description matched the words of the request. One said it chose Tavily because the request’s description “is its exact positioning.”

The same thing happens outside developer tools. In a study of 12 apparel brands across five AI engines, lululemon appeared in 90% of answers when the question said “athleisure” and in none when it said “athletic footwear” (Search Engine Land, 2026).

Across our other runs:

Product What agents did
Linear, asked for an issue tracker Picked it first in 68% of runs
Kapa Recommended it or built with it in 44% of runs
Canva Used it in 5% of runs; built the image generation themselves in 57%

In the Canva runs, even when told not to build a design editor, agents mostly chose Polotno, Bannerbear, or Adobe Express.

Building it themselves is common beyond our runs too. In Amplifying’s study of Claude Code, custom code was the most frequent single pick, at 12% of 2,073.

Ask an agent for a specialization and it reaches for the companies it ties to that field, or builds the thing itself. For yours, are you one of them?

You can’t see that choice in your analytics

Your analytics can’t tell you whether an agent thinks of you or considers you as an option, because an agent usually decides before it reaches anything you can measure. For your company, that means three things:

  • The choice comes before the search. Across the 139 runs where an agent picked a vendor without being told which, 77% named it before any search result came back. A search result led to the pick in 9%. Review articles and forum threads came back in the results, and agents almost never opened them. They went to the docs of the vendor they had already chosen.
  • Being passed over leaves no trace. There is no visit, no bounce and no lost deal. Parallel is the pick for deep research and nearly absent on web search, which it also sells, and no funnel report would show that gap.
  • One check isn’t enough. Asked the same question about an issue tracker, Codex recommended Jira in every run and Claude Code recommended Linear in all but one. The answer changes between model versions too: in the same Amplifying study, one tool fell from 79% of picks to none between two versions.

If you haven’t asked an agent for what you sell, you don’t know which requests it ties to you and which it hands to someone else.

Then the agent has to get in: API keys, errors and install paths

Say you are one of the companies an agent reaches for. It has picked you, and that was the first half. Now it has to get in.

Agents are already arriving in numbers. Databricks put the agent-created share of Neon’s databases at more than 80%, up from 30%. Netlify reports more than ten times the daily signups it had before January 2025, when its CEO made agent experience the company’s priority. A JetBrains survey of more than 15,000 professional developers found nine in ten using a coding agent at work at least weekly.

So what happens when one reaches your product? In our runs, three things decided it.

  • Can it start from a key the customer already has? At Parallel, Kapa, and Linear, the customer’s own API key was all the agent needed. With that key in its environment, the agent called the API directly, with no OAuth screen or browser login in the way: we recorded an answered API call in 100%, 88%, and 81% of runs. At another vendor, agents started the setup wizard in 59 runs and finished it in none, because creating an account needed a person. In a real project, each of those is a customer who never arrives. In our database benchmark, walls like that stopped most of the sessions that failed.
  • Does an error say what to do next? The three products weren’t error-free. Each API rejected a request in 22% to 33% of runs, and agents hit a missing docs page in 28% of Parallel runs and 13% of Kapa runs. Every one of those tasks still finished. When Parallel’s API named the field it rejected, the agent fixed the call on the next try.
  • Does every path lead to the current product? At one company, 28% of agent deployments installed v1. The docs described v2, but an old Helm chart repository still listed v1 as the newest version. The company had fixed its docs and left the path the agent took.

An agent that gets stuck doesn’t file a ticket or email sales. It picks another vendor or builds the thing itself. If a customer handed an agent their key today, how far would it get on your product?

What watching an agent shows you

The only way to see this is to watch it happen. The test is simple. An agent gets a task one of your customers would give it, with no product named. You see whether it thinks of you, what it picks instead, and where it stops if it does try you. Change one thing, run it again, and you see whether the change mattered.

Every one of our numbers in this piece came from tests like that, run at Scope.

Sooner or later an agent will be asked for what you sell. What will it answer?

Scope tracks agent sessions so you can see how agents find and use your product and what to improve so more of them succeed.