AI Radar

This is the first time using AI in your work becomes an administrative obligation rather than a choice. For an agency it means the question "was this made with AI?" is no longer a conversation with the client but a field in the deliverable. And because the marking is machine-readable, a client can check it afterwards without asking you.

Do an inventory this month: which deliverables are fully generated, which AI-assisted, which made by hand. Put the marking in the production process, not in the small print of a proposal — a label that first appears on the invoice arrives too late.

The workhorse tier is where an agency’s margin lives — that is where the volume runs. A doubling on a benchmark that mimics office processes, at under a dollar per million input tokens, shifts which tasks you still do by hand at all. The catch is in the small print: budgeting on $0.75 today is budgeting on a price that disappears in four months.

Move bulk work here — extraction, classification, first drafts — but run your business case on the 2027 rate. Anything that only works at the introductory price does not work.

Until now the agent sat beside the stack: you gave it access to your systems. Here the customer context sits inside the conversation, with the CRM’s governance wrapped around it. That moves the question from "which agent do we use?" to "where does the data that makes it useful live?" — and the answer decides who owns the customer relationship. Benioff did not brush off the SaaSpocalypse question in the same week by accident.

For your largest clients, check where the customer context actually sits and who is allowed to open it up. An agent that lives in your client’s CRM is not a tool you choose but a channel you plug into.

The interesting price cut is not on tokens but on memory. An agent that re-reads the same brief, brand guide and client file thirty times pays its bill exactly there — and that line item got four times cheaper without the list price moving. If you are not using caching, none of that 45% reaches you.

Check whether your long-running jobs have prompt caching switched on, and whether your fixed context — brand guidelines, tone, client facts — sits at the front of the prompt where it can be cached. That is the cheapest optimisation available this month.

There are now two frontier models at exactly the same list price: Astra and Fable 5.1, both $10/$50. So the choice is no longer about budget but about the task — and about what you have to be able to explain. Astra is faster and stronger at computer work; it is also the model whose reasoning you can least retell. For work going to a client or a regulator, that has become a real criterion.

Re-test your two heaviest workflows on both Astra and Fable 5.1 this month, and compare on outcome, not on benchmark. Make human review mandatory on anything outbound: with a model whose reasoning you cannot read, the review is your only audit trail.

Those two numbers are each other’s mirror image, and together they are the whole story: traffic falls, value per visitor rises. Anyone still judging AI visibility on sessions sees a falling line and draws the wrong conclusion. Anyone judging it on citations and conversion sees a channel that is smaller and better at once.

Measure citations, not just visits. Record each quarter which questions you are cited for in ChatGPT, Perplexity, Gemini and Google’s AI overviews, from which source, and put the conversion of that traffic beside it. That is one report, and it answers the only question your client is asking.