The Product Desk

Where the opening actually is

OpenAI shipped Presence through a high-touch GA run by forward-deployed engineers and systems integrators. No public pricing. No disclosed geographic limits. No integration-cost estimate. Read that as the seam, not the strategy: while OpenAI hand-builds deployments one enterprise at a time, a competitor leading with transparent, self-serve pricing reaches the mid-market before the sales calls even start. Viktor is already doing the alternate motion. It runs natively inside Slack and Teams for 40,000-plus teams and hands out $100 in free credits. That is the self-serve wedge sliding under the white-glove one.

The more durable shift is what buyers now grade agents on. Presence pairs model reasoning with permissions, policies, evaluations, and escalation rules — nearly the same control surface Cursor just shipped for admins, with per-team toggles and model allow/block lists, and the one Anthropic is building into managed projects. Three vendors landed on the same governance controls in a single cycle. When that happens, the surface is the entry ticket, not the differentiator. An enterprise agent PRD without permissions, escalation, and audit logging is not losing on marketing. It is incomplete.

The ROI language that unlocks budget

Two deployments moved agent ROI from the deck to the ledger. Wordsmith, a startup rather than a lab, runs in production at $7.6B Belron, cutting contract drafting from an hour to five minutes and saving $400,000 in a single quarter. Belron operates in 40+ countries with in-house lawyers in only 15 and is now rethinking legal headcount. Viktor's Hampton rollout reached 18 active staff, 12 internal apps, 26 scheduled tasks, and 887 CEO-level threads in 44 days, against a $440K hiring budget that went entirely unspent. In both, the foundation model is table stakes and the workflow fit is the moat. Wordsmith wins because you email an Excel sheet and get a contract back, routing across OpenAI, Anthropic, and Google underneath. The user does not see the model. That is the point.

The demand gap says the same thing. 83% of 1,402 surveyed leaders say they need infrastructure upgrades to move agentic AI from pilot to production. Buyers want this and cannot get there on their own. That is the lane for a product that makes agents deployable and governable, not just capable.

The smart move

Don't imitate Presence. Out-position it. Lead where OpenAI is slow: pricing transparency, self-serve speed, and a specific vertical workflow the labs skip. Then measure the feature against the external benchmarks buyers now carry — a 44-day adoption curve and hard-dollar savings, not a quarter-long pilot with engagement charts. If the business case can't produce a CFO-legible number, the problem is the feature's design, not its marketing.

The agent's model is table stakes; the workflow it fits into is the only part a foundation model can't ship for you.

Why a payments company wants a router

OpenRouter does one deceptively small thing: pick the best and cheapest model for each request. Watch what a buyer pays for that. Stripe would pay roughly $10 billion, up from a $1.3B valuation, with Databricks also bidding. The number tells you the market now treats model orchestration as strategic infrastructure, not a scripting afterthought. And the number carries an uncomfortable implication for anyone routing through a single aggregator. That layer is about to sit inside a payments or data giant with its own pricing incentives and its own plans for your traffic.

The consolidation isn't isolated. Amazon closed its San Francisco AGI Lab and cut its frontier-model staff, doubling down instead on helping customers build on its existing Nova models. That cedes foundational-model leadership to three labs — Google, OpenAI, and Anthropic — and concentrates supply-side pricing power. Sierra acquired agent startup Takeoff, which is what maturing platforms do when they decide to skip a multi-quarter internal agent build rather than ship it.

Compute stays expensive while the layer concentrates

The timing compounds the risk. Google raised FY26 capex to $195B–$205B, posted its first-ever negative free cash flow (-$5.9B), and its CFO said the cloud business is still capacity-constrained even as Cloud revenue grew 82% YoY to $24.8B. First customer TPU deliveries won't generate meaningful revenue until 2027. The AMD–Anthropic deal — up to $5B invested, 2GW of MI450 chips — also starts H1 2027. Relief on inference pricing is a 2027 event. A roadmap pricing in falling token costs over the next 18 months is building on a floor that isn't there yet.

The counter-evidence worth holding

Consolidation doesn't mean the labs eat everyone. Separate the pitch from the evidence: AlphaSense grew ARR 40% to $700M specifically off AI features in a vertical. That is retention and depth, not an engagement chart. It's the comp to raise when someone in a review says 'a foundation model will just do this.' The defensible answer is proprietary data plus workflow depth, not model access.

The smart move

Here is the forcing function for Monday. Treat vendor strategy as a first-class roadmap decision, not a procurement afterthought. Abstract routing behind your own interface so a change of ownership can't dictate your pricing or terms, and keep a tested failover to at least two direct providers. Then confirm your primary and secondary frontier vendors sit among the three remaining labs.

The layers your AI features sit on are being bought and sold — own the abstraction, or inherit someone else's pricing.

Three teams, one anti-pattern, same day

The convergence is the evidence. HeyGen's Companion Mode storyboards and proposes creative directions before it renders a frame. Manus Plan Mode turns a rough idea into an approvable plan before it builds. Gemini for Slides drafts a deck from the user's real Docs and Sheets but leaves every slide fully editable. Three independent teams shipped the same rejection of the one-shot 'magic button' in a single cycle. That's not coincidence — it's a UX standard crystallizing around a specific user pain: I can't trust or control what the AI hands me.

The second half of the pattern is data grounding. Anthropic wired Claude to its proprietary Economic Index so answers cite a live dataset rather than parametric guesses — a data-moat play, not a model play. Put grounding next to checkpoint-gated generation and the message is consistent: the winning AI feature isn't the one that generates fastest, it's the one users can steer and start from their own work.

Why the model itself stopped being the moat

Underneath the UX shift, capability is commoditizing on two fronts. Black Forest Labs' FLUX 3 folds image, video, audio, and even robot-action generation into one architecture — meaning separate per-modality vendor integrations are now a maintenance liability, not a moat. And Microsoft's VS Code Agent Host Protocol runs Copilot, Claude, and Codex as interchangeable harnesses with worktree isolation and risk-based permissions, giving away multi-agent plumbing for free. If your differentiation story is 'we run multiple models' or 'we support many modalities,' the platform just absorbed it.

The cost pressure runs the same direction: Kimi K3 (2.8T parameters) matched GPT-5.6 Terra on a private security benchmark at substantially lower cost. Caveat: sanctions risk on Chinese-origin models is live, so treat open-weight parity as a COGS lever behind a swappable interface, not a default.

The smart move

Move engineering up the stack now. Add plan-to-preview-to-commit gates to any high-cost or high-stakes generation flow — it's the cheapest, highest-signal change on the table. Reframe your top AI bet from 'generate from a blank prompt' to 'transform the artifacts the user already gave you,' and ground one proprietary dataset you already own. Then redirect any 'multi-agent support' investment toward agent reasoning quality or a vertical workflow, because the plumbing is no longer yours to sell.

The winning AI feature isn't the one that generates fastest — it's the one users can control and start from their own work.