The Product Desk

The Distribution Shift in Numbers

Three independent data points converged this week to confirm a structural change in how AI products reach enterprise users. Perplexity's Computer agent has grown enterprise Slack users 17% per week since its April 2026 launch — roughly 3x per month. Viktor, a Poland-based startup, had its best week ever (400+ new customer signups) directly after Claude Tag launched. Neither company pays Salesforce or Microsoft a dollar for that distribution.

Satya Nadella told shareholders the most exciting AI developments are plugins in Word and Excel — reading this as a decision: Microsoft values platform gravity over Copilot revenue exclusivity. Teams has several times more users than Slack. The channel is free. Microsoft will tolerate agents that compete with Copilot inside its own workspace.

The Anthropic Trap You Must Navigate

Here's where the story gets uncomfortable. A major software CEO privately drew the Facebook Like button analogy — comparing Anthropic's integrations to a data intelligence play, not a distribution partnership. Supporting evidence: In Fall 2025, an Anthropic official publicly stated they used Claude to build a Slack-like workplace chat app. Figma dropped collaboration talks because Anthropic is building design tools. Salesforce employees internally compare Claude Tag to 'letting a fox into the hen house.'

Anthropic integrates, learns the workflow, then ships the competing product. The question isn't whether to be on the platform — it's what survives when Claude can orchestrate 80% of the workflows you own today.

The 2x2 That Decides Your Strategy

Run this diagnostic before your sprint plan. Axis 1: Can Claude or Perplexity already do the core task your product charges for? Axis 2: When it does that task, does the output need revision against something durable — proprietary data, source-of-truth integration, a verification step only your system can perform?

• They can do it + output is usable: You're being disintermediated, not distributed. The agent channel is where you get replaced.
• They can do it + needs revision: That revision step IS your product. Ship it as a Slack agent immediately — it's worth more than your standalone app.
• They can't do it: Your moat holds, but ship the agent anyway for distribution before the gap closes.

Why 90 Days Is Your Window

Perplexity's agent already creates GitHub tickets and queries Snowflake databases. Claude Tag asks customers to connect it to their apps and databases with near-zero friction. AWS's $1B Forward Deployed Engineer investment confirms enterprise AI implementation remains brutally complex — which means the winning agents are winning on setup simplicity, not raw capability. The cost of entry is zero. The cost of waiting is that a general-purpose agent learns your user's workflow this quarter and automates it next quarter.

The Buyer Your Page Can't Serve

An AI agent visited a SaaS pricing page this week. It rendered the JavaScript, waited for the modal, parsed the comparison table, and gave up. The buyer got a wrong answer — not because the pricing was bad, but because the page was built for a human with a cursor, and the thing reading it was not.

This is happening at scale. ChatGPT preferentially cites third-party sources (Reddit, G2, review sites) over brand websites. It deduplicates by domain — each domain gets one shot. It favors plain HTML over JS-rendered content. If your critical product information lives behind JavaScript rendering, modals, or interactive components, AI agents literally cannot see it.

The /pricing.md Pattern

Several companies have quietly begun shipping /pricing.md files — plain markdown at a static URL with plan names, features, prices, overage costs, and billing terms. No authentication, no rendering required. Some host these without linking them anywhere, betting agents will find them the way crawlers found robots.txt. For B2B SaaS, this scopes to hours of work and determines whether an AI agent can evaluate your product at all.

Shopify's Bet: The Global Catalog for Agent Buyers

Shopify is building something more ambitious: an LLM-powered clustering layer across billions of products from millions of merchants, organized into a single searchable intelligence layer explicitly designed so AI agents can map products to buyer preferences. This isn't a feature — it's a platform architecture decision betting that agents, not humans, will be the primary 'shoppers' of the future.

Products that are opaque to agents — complex UIs with no structured APIs, unstructured data models, workflows requiring human visual interpretation — will lose distribution as agent-mediated workflows become mainstream. This is the SEO of the agent era.

The Immediate Playbook

The fix splits into two timeframes. This week: ship a /pricing.md file and ensure critical product pages render in plain HTML. This quarter: add an 'agent persona' to your user research framework, audit your information architecture for machine navigability, and prototype structured APIs that let agents complete tasks in your product.

One additional conversion signal worth noting: adding reassurance copy inside the primary CTA button (not beside it) produced a 29% conversion lift in testing. At the decision moment, attention narrows to the button — copy outside it is invisible. Test 'Start free trial — no credit card needed' as button text against your current CTA this week.

What the Bill Actually Requires

A product lead skimmed the headline, filed it under "consumer protection," and went back to the sprint board. That's the pitch. Here's what it does to a roadmap. Sen. Warner released the AI AGENT Act on June 29, 2026. Strip the framing and the bill imposes four concrete constraints on AI agent providers:

1. FTC registration — agents taking consequential actions become registrable entities
2. Human-operator linkage — every agent action must trace to a responsible human, with attribution running action → agent → operator
3. Consent and revocation controls — users can stop any agent action at any point in the chain
4. Third-party certification — not FTC-enforced directly, but validated by certification bodies, which creates a new compliance market

Three of these are surface features you can bolt on after launch. Human-operator linkage is not. It has to live in the data model from the first action an agent takes. Skip it now and you rewrite history later.

The Cost Gap: Build Now vs. Retrofit Later

The bill also names bot-to-bot interactions as a risk vector, so Agent A calling Agent B calling a tool needs governance checkpoints. And the part buyers will feel first: large platforms must give users the right to choose at least one compliant agent provider. That is an interoperability mandate. This is the DMA for AI agents.

Building human-operator linkage and revocation into agents that already shipped is not a sprint. It's a migration, plus a backfill of records you never captured, plus a certification pass on top. Building the same constraints before launch is a design decision that costs days.

The Research That Changes Your UX Copy

Boston University's Emma Wiles measured what happens when you call an agent a "coworker." Managers reviewing work attributed to an agentic "AI employee" caught 18% fewer errors than when the identical output was labeled a "chatbot." Microsoft, OpenAI, Anthropic, and Google are all marketing agents as digital colleagues. The research says that positioning makes the humans reviewing the work worse at reviewing it. Ford's public reversal — rehiring 350 engineers after AI quality control failed, saving hundreds of millions — is the enterprise version of the same finding.

How This Plays Out in Procurement

The FTC registry is a soft gate in law and a hard gate in deals. Enterprise procurement will treat FTC registration the way it treats SOC 2: table stakes, checked before the demo. Okta's GA of AI agent governance for FedRAMP/HIPAA environments — agents as first-class identities with least-privilege tokens, audit logging, SIEM streaming, and kill switches — now sets the floor buyers will ask for. So the forcing function, before the next agentic action ships: can you answer three questions on demand — which agent, whose authority, which version?