Three weeks into a project, the delivery lead’s inbox fills up with client questions. What AI tools did you use? Who reviewed the output? What changed between the first draft and the version we saw? Will the process be consistent going forward?
The work is good and the client liked the first round. When the revision cycle opens, the answers aren’t ready, and everyone on the agency side assumes someone else owns them. Approvals stall on a question nobody was assigned to answer.
The AI didn’t cause that but a missing decision did. Nobody agreed up front who reviews AI-assisted work, who signs off, and who tells the client. Those decisions don’t show up in a standard project plan, so they get skipped, and they surface at the worst possible moment: in front of the client.
In our last post on AI governance, we covered the policy layer: what agencies commit to at the organizational level. A policy that states AI-assisted output gets reviewed before client delivery only holds if someone owns the review. RACI hasn’t stopped working in the AI era. It has four new questions to answer, and the teams that answer them at kickoff are the ones whose approvals don’t stall.
Four questions a standard RACI skips
The Responsible / Accountable / Consulted / Informed model works fine when people do all the work: someone drafts, someone approves, someone gets briefed. AI doesn’t break that logic. It adds tasks the model was never asked to cover, because those tasks didn’t exist when most teams wrote their project templates.
Who owns quality review? When an AI draft hits a senior strategist’s inbox, who is responsible for catching the off-brand phrasing: the strategist, or the writer who prompted the tool? On a human-only team, the writer owns the draft. On an AI-assisted team, the writer is editing the AI’s first draft, and the line between drafting and reviewing blurs. If both assume the other owns it, nothing gets caught.
Who tells the client AI was used? If the answer is “whoever happens to be on the call when they ask,” it will be different on every project. Disclosure that’s consistent reassures a client; improvised disclosure makes them nervous.
Who flags the work before AI touches sensitive content? Brand voice, regulated copy, anything involving client data: these need a human decision before an AI tool is involved. Most agencies have informal rules. Few have named the person who enforces it.
Who is accountable for output that misses the brief? When an AI-generated asset doesn’t land, the next conversation is about whose name is attached. On well-run teams, the kickoff doc already names who answers it
The framework still works. It just needs four more rows, one for each question.
The four rows to add
| Accountability area | R — Responsible | A — Accountable | C — Consulted | I — Informed |
|---|---|---|---|---|
| 1. AI output review | Team member who prompted the tool | Senior practitioner; sign-off logged before work ships | Delivery lead, for client-facing deliverables | Project manager, for timeline impact |
| 2. Client communication on AI use | Delivery lead, at kickoff | Account lead; owns disclosure and any mid-project update | Senior practitioner, on what’s accurate to say | Full project team |
| 3. Quality gate before delivery | Senior practitioner doing the review | Delivery lead; signs off that the review happened | SMEs as needed: legal for indemnity, brand lead for voice | Account lead |
| 4. Sensitive content escalation | Team member about to use AI; raises the flag before starting | Delivery lead; approves or escalates | Account lead; client compliance contact for regulated work | Senior practitioner who will review the output |
Some suggestions on how to make this run:
Log the sign-off in your project tracker. The agency, not the freelancer, answers to the client for AI-assisted output, so make the senior practitioner’s accountability explicit in any freelancer agreement.
Cover four points in the kickoff client conversation: What AI is for on this engagement, what gets flagged for human review, who communicates changes to the client, and what happens if scope changes. The project manager documents this in the kickoff notes. When the client asks later, the answer is on file.
One line belongs in every project plan: AI-assisted output reviewed by [named role] before client delivery. That line is the difference between “we’re careful” and a process a procurement team will accept.
Clients approve faster when they know what AI touched
Tell a client at kickoff what AI is for and who reviewed it, and there’s nothing for them to question when they see it. Let them find out mid-project that AI touched something nobody mentioned, and they’ll start second-guessing everything else you’ve sent. What changed is whether they heard it from you first.
Peter Kang, co-founder of Barrel Holdings and a recurring voice in the AgencyHabits research, argues that the agencies that scale well are the ones with systems for delivering impactful work, rather than relying on instinct and strong relationships. A kickoff RACI is the smallest version of that system: a one-page record of who owns what, settled before the work starts.
On VIP and Pressable engagements, enterprise clients audit how the work gets made, not just what gets delivered. With those clients, “who reviewed the AI-assisted work on this account?” is a question someone will ask. If the review was logged, you can name who and when on the spot.
Where to start
The companion AI-workflow RACI template covers all four questions above. Take it into your next kickoff, adapt it for the engagement in about 20 minutes, and settle ownership before the first deliverable is due.
This piece is about how AI governance gets put into practice. In The Enterprise Shift: Why AI-Era Clients Are Buying Differently, we looked at how procurement behavior has changed for enterprise WordPress buyers. The accountability framework in this post is one way agencies can make those expectations operational.
This content is provided by Automattic for general informational purposes only and is not legal, financial, or professional advice. Any tools, methods, templates, certifications, or processes referenced are shared “as is” and without warranties. Agencies should independently evaluate any approach for fit, security, compliance, and client obligations.
Automattic 