Who Signs Their Name to This?

Who Signs Their Name to This?

Status
Published
Author
Created
Aug 6, 2026
Tags
Accountability
Agentic AI
Consulting Ethics
Text
Agents can draft the work now. Only a human can still own it.
A partner at a consulting firm signs her name to a deliverable she didn't write. An agent drafted the market sizing, ran the sensitivity analysis, and flagged the two competitors most likely to retaliate on price. She read it, adjusted a paragraph, and sent it to the client under her signature. Nobody in that exchange thinks anything unusual happened. That's the part worth sitting with.
For a hundred years, the signature on a deliverable has meant something specific: this person is answerable for what's inside. Not that they typed every word, but that they stood behind the reasoning enough to put their name where a client could find it if something went wrong. Agentic AI hasn't broken that convention yet. It's just made it possible to sign things you understand a little less each time, produced by a process you can no longer fully trace.
Ask "who owns this?" about a piece of AI-assisted work and the honest answer, most of the time, is: nobody has actually decided. Grant Thornton's 2026 AI Impact Survey of nearly 1,000 business leaders found that governance and compliance gaps — not workforce readiness, not the technology itself — are now the top reason cited for AI initiatives underperforming. Not because the models are unreliable. Because the org chart never caught up to the workflow. Roughly four out of five organizations running agentic pilots are doing so without a named owner sitting above the agent's output, deciding what "good enough to ship" means and absorbing the consequence when it isn't.
That's the accountability vacuum, and it's a specific kind of failure: not a bad decision, but the absence of anyone positioned to make one. When an agent drafts a client memo, runs a due diligence checklist, or proposes a restructuring plan, the work still has to land somewhere — a person still has to decide whether it's true, whether it's appropriate, whether it should go out the door. Skip that step and you haven't removed the decision. You've just made it silently, by default, usually in favor of "ship it."
Consulting has actually solved this problem before, just for a different technology. When Excel models got sophisticated enough that no single reviewer could trace every formula, firms didn't respond by trusting the spreadsheet more. They built review layers, sign-off protocols, second-pair-of-eyes norms — structural insurance against the gap between "the model produced this" and "someone verified this." The lesson wasn't about spreadsheets. It was that as a tool's output outpaces a person's ability to fully audit it, you don't get to relax the ownership question. You have to make it more explicit, not less.
Agentic AI needs the same move, scaled to the fact that agents now touch strategy, not just arithmetic. The reframe isn't "who used the AI." It's "who is the named business owner of this outcome, before the work starts" — someone with the authority to override the agent, the obligation to review what it produces, and the standing to answer for it if it's wrong. That person doesn't need to have written the analysis. They need to have decided to stand behind it, deliberately, in advance — not by default, three drafts later, when it's already halfway to the client.
This matters more in consulting than almost anywhere else, because the product being sold has always been judgment, not output. A client doesn't pay a firm for a document; they pay for the claim that someone qualified has taken responsibility for what's in it. If agents can now produce the document, the fee only still makes sense if a human is doing something an agent can't: deciding what's true, what's risky, and what the client actually needs to hear versus what they asked for. Strip that layer out and the engagement isn't advisory work anymore. It's a subscription to a language model with better formatting.
So the uncomfortable version of the question isn't "did the AI do a good job." It's whether there's a person in the room who would still say yes, out loud, if asked to defend this in front of the client six months from now, after something's gone wrong. If that person doesn't exist yet on your team, you haven't adopted agentic AI. You've just quietly outsourced the one thing that was never supposed to be outsourceable.