Glossary
Principal AI and Silent Downgrade
Two terms for the disclosure gap opening between clients and agencies over AI-produced work.
Principal AI
Principal AI is the practice of an agency producing client deliverables with AI models it procures at its own undisclosed cost, then billing that work at labor-based rates — leaving the client unable to see which models were used, what the compute cost, or what margin was earned on the difference.
How it works. The agency buys model capacity and resells its output inside a fee built on hours, seniority and scope. Routing decisions — which model, how much reasoning, how many passes — are made by the agency, in its own economic interest, with no disclosure obligation. Volume and spend commitments to model providers create the same duty-of-care tension that volume commitments create in media.
Where the name comes from. Principal media buying, in which an agency buys inventory as principal and resells it to the client at a margin the client cannot see. The structure is the same. Only the inventory has changed.
Not to be confused with. AI use. Agencies should use these tools and clients should want them to. Nor is it fraud — most of it exists because no standard does. And it is not resolved by an agency saying “we use AI,” which is a capability statement, not a disclosure.
The test. If you cannot answer which model produced this and what it cost to run, you are in a principal AI relationship — whether or not anyone in the room intended one.
Silent Downgrade
Silent downgrade is the undisclosed substitution of cheaper AI capability into client work — a smaller model, a shorter reasoning budget, fewer passes — producing a deliverable that looks unchanged in format and price while containing measurably less thinking.
Why it is structural. Per-token prices for a fixed capability level keep falling, but tokens consumed per task are rising faster. Agentic workflows burn five to thirty times what a simple query does, and Gartner expects the cost of a single agentic workflow to more than quintuple by 2028.1 Set that against a retainer priced before any of it existed, and downgrading becomes the rational response.
Two directions of harm. When compute gets cheaper, the client overpays and the harm is economic. When compute gets more expensive, the client is shortchanged and the harm is qualitative — and leaves no trace. Neither is visible today, which is why the remedy in both cases is disclosure rather than price negotiation.
Not to be confused with. Model choice. Routing simple work to a cheap model is good practice and good stewardship of a client’s budget. The operative word in the term is silent.
The test. Ask what changed in how your work was produced over the last two quarters. If the agency cannot answer and the deliverables look identical, you have no way to rule it out.
Principal AI is the structure. Silent downgrade is the symptom.
One is a compensation question. The other is a quality question. Both are answered by the same disclosure.
How to cite these terms
In proseMercer Island Group’s Steve Boehler defines principal AI as “the practice of an agency producing client deliverables with AI models it procures at its own undisclosed cost, then billing that work at labor-based rates.”
FormalMercer Island Group (Boehler, S.). “Principal AI and Silent Downgrade.” August 2026. migroup.com/glossary/principal-ai.
Licensed CC BY 4.0 — reproduce, quote or adapt freely, including commercially, with attribution to Mercer Island Group.
Note
- Gartner, reported in CIO Dive, “Advances in AI capabilities to outpace cost savings,” August 17, 2026. ciodive.com/news/ai-systems-outpace-cost-savings/828068/ ↩