In July 2026, two documents landed within days of each other and described the same machine from opposite ends.
The first is a field report from inside Clay, written for GTM Strategist by Alex Lindahl, one of the company’s first GTM Engineers. It catalogues ten places where Clay’s sales organization runs AI in production, and it is unusually specific about what the AI is allowed to finish. The second is a thesis post from Union Square Ventures titled “Obliterate, Don’t Automate.” It is bylined to the firm rather than to an individual partner, and it argues that the licensed intermediary standing between a customer and the expertise they need is the inefficiency worth deleting.
Read together, they are not in conflict about the technology. They are in conflict about whether the human standing at the gate is doing work or blocking it.
What Clay Chose to Keep
Lindahl’s design rule is stated plainly: “you should separate recommendation from execution (AI can nail the strategy, but sellers need the underlying evidence to verify before acting).” Everything in his account follows from that split.
The MSA redline example is the sharpest one. Contracts arrive, an AI pass triages the redlines, and the output routes to counsel. The reported effect: “Legal gets back in four hours instead of two days.” Counsel still reads what reaches them. What changed is the queue, the ordering, and the volume of low-stakes clauses that never needed a lawyer’s attention in the first place.
The Monday morning routine is the same shape at smaller scale. A rep’s signup-briefing prep “used to take him an hour. Now it takes four minutes.” Summed across the week, Lindahl reports “almost a full day back each week.” Clay also scores calls against an internal rubric it calls the 3Ts: Tailor, Teach, Take control. The rubric is applied by machine. The coaching conversation that follows is not.
Lindahl states the boundary in a sentence: “The AI doesn’t close deals. Marcus closes deals.” And later: “None of these tools replaced Marcus. They just gave him his day back.”
That is a governance architecture wearing a productivity story. Recommendation is unconstrained. Execution carries a named human. The named human keeps their day because the machine stopped consuming it, not because the machine took over their signature.
What USV Wants Removed
The USV post reaches for a historical analogy: “The internet collapsed the cost of distribution and stripped the market power from the media companies that controlled it. AI does the same thing to expertise.”
The prescription follows directly. “Instead of renting expertise from a gatekeeper, you can access and deploy it directly. No gatekeeper required.” And, at portfolio scale: “Almost everywhere you look, there’s an opportunity to transfer the balance of power away from a gatekeeper and towards a customer.”
This is not abstract. USV points at Doctronic, whose AI doctor “can even legally write prescriptions, starting in Utah.” A prescription is a consequential write in the strictest sense. It is a legally binding instruction with a named liable party attached, and Utah has now permitted the software to be that party’s proxy.
The post contains no statistics and makes no empirical claim about outcomes. It is a thesis about where market power should move, published by a firm that has already deployed capital behind it.
Both Descriptions Are Accurate
Clay and USV are describing the identical mechanical event.
When Clay’s system triages MSA redlines, it removes most of the lawyer’s contact with the contract. Ninety percent of the clauses never reach a human. From the customer’s perspective, the wait dropped from two days to four hours because the intermediary was largely bypassed. USV would call that obliteration and be correct. Lindahl calls it triage and is also correct. The same code produced both sentences.
The disagreement is about the remaining ten percent, and specifically about what that residue is made of.
USV’s model treats it as unfinished automation. The gatekeeper’s remaining touch is friction that a better model, a clearer regulatory path, or a more aggressive startup will eventually clear away. Utah cleared it for prescriptions. Other states will follow. The trajectory is the argument.
Clay’s model treats the residue as structural. The reason Marcus signs is not that the model cannot draft. It is that when a deal is misrepresented, a customer sues someone, and the someone has to be a person or a firm that can be sued. We have written about where the boundary sits in an AI-native sales org and about who owns an account decision when the pipeline is machine-generated. Both landed on the same place: the last signature is where liability attaches, and liability does not compress the way review time does.
The Question USV Left Open
The obliteration thesis leaves one question unanswered. When the licensed gatekeeper is removed, where does the liability go?
The gatekeeper was never only a bottleneck. A licensed professional is a bundle: expertise, plus a license that can be revoked, plus insurance, plus a body of malpractice law with decades of precedent about what counts as negligence. AI collapses the cost of the first item. It does nothing to the other three.
Doctronic writing prescriptions in Utah means one of four things happened. A supervising physician absorbed the liability and the automation moved the work without moving the risk. Or the platform absorbed it, which makes the platform an insurer with a novel actuarial problem and no loss history. Or the state created a new liability category that has not been tested in court. Or the risk quietly transferred to the patient, who now holds an outcome that no professional is answerable for.
Those are very different worlds. The USV post does not distinguish between them, and the distinction is the entire operating question for anyone deploying this pattern. Harvey’s expansion inside law firms is instructive here: the tooling scaled fast and the partner’s signature did not move, because the signature was never the slow part.
The pattern generalizes past medicine and law. Agent roles are being defined across finance, HR, and support, and every one of those functions has some version of a person whose approval means somebody is answerable. Removing the approval step is easy. Relocating the answerability is the part nobody has specified.
Run the Accountability Inventory This Week
Take your three highest-volume approval steps. For each one, write two sentences.
Sentence one: what does this approver actually contribute? Distinguish knowledge from accountability. If the approver’s contribution is knowledge, an AI pass plus a verifiable evidence trail can carry most of it, exactly as Clay’s redline triage does. Route only the exceptions to the human and reclaim the queue time.
Sentence two: if this decision goes wrong, who gets sued, fined, or fired? Write the name of a person or a legal entity. If you cannot write one, you have not automated a gatekeeper. You have deleted an accountability layer and left the position vacant.
Most teams will find that their approval steps are a mix, and that the mix is not documented anywhere. That is the useful output. The steps that are pure knowledge should be compressed aggressively this quarter. The steps that carry liability need the named party written down before the model touches them, because the fastest way to discover you removed accountability is to need it and go looking.
Clay’s version of this ran the inventory implicitly and kept Marcus. USV’s version is betting that entire categories of Marcus were holding nothing. Both bets will be settled by litigation, not by benchmarks, and the settlement will arrive years after the deployment decisions are made.
This analysis synthesizes 10 Ways Clay’s GTM Engineers Use AI to Accelerate Sales (GTM Strategist, July 2026), Obliterate, Don’t Automate (Union Square Ventures, July 2026).
Victorino Group helps organizations separate knowledge work from accountability work before automating either one. Let’s talk.
All articles on The Thinking Wire are written with the assistance of Anthropic's Opus LLM. Each piece goes through multi-agent research to verify facts and surface contradictions, followed by human review and approval before publication. If you find any inaccurate information or wish to contact our editorial team, please reach out at editorial@victorinollc.com . About The Thinking Wire →
If this resonates, let's talk
We help companies implement AI without losing control.
Schedule a Conversation