The floor moved; the ceiling did not

Automated scoring did not compress this market evenly. It removed the price from one specific thing - a plausible-sounding general assessment delivered quickly - and left everything above it roughly where it was.

That is a more useful description than "AI took the low end", because it tells you exactly which sentences in your own deliverable are now worth nothing.

The part that is gone

A structured, competently written, non-specific assessment is available instantly and free. It reads well, it is internally consistent, and it is produced in less time than it takes you to open the brief.

So any component of your offer that consists of applying a general framework to a submission without saying anything that required you specifically to have looked at it is no longer a billable component. It can still be in the deliverable - buyers like structure - but it cannot be the reason for the price. This is now measured rather than anecdotal. After ChatGPT's release, Hui, Reshef and Zhou found writing freelancers on a large online marketplace lost 2% of monthly jobs and 5.2% of monthly earnings, and higher-quality freelancers were not spared. Demirci, Hannane and Zhu's 2024 CESifo working paper found a 21% drop in job posts for automation-prone writing and coding work within eight months, with the remaining jobs more complex and better paid.

The technical account of what automated scoring reliably does and does not do belongs to the property that covers it: the how-it-works explanation is the reference, and its companion material on where those systems are and are not accurate matters more to an earner than it sounds, because the accuracy boundary is the boundary of your remaining market.

The part that did not move

Judgement about context. Comparison against things only a person has seen. Willingness to say the awkward thing, and to be accountable for having said it. The simple fact that a named human being looked, which is the entire product in a surprising share of bookings.

None of this got cheaper, and demand for it did not fall. If anything the reverse: a buyer who has already run the free version and found it unsatisfying arrives with a sharper question and less price sensitivity than a buyer who has not. Most buyers now arrive that way, having tried one of the free tools first, which changes the opening of the conversation considerably.

What this does to a rate card

The effect is a hollowing rather than a decline. The cheap generic tier is not worth offering, the middle is thinner than it was, and the top is unaffected.

The practical consequence is that an entry-level tier priced to be accessible is now mostly a way of losing money slowly, because it competes against zero. Why the cheapest tier attracts the hardest buyers was true before this; it is now true with a smaller upside attached.

What to change in the deliverable

Move the generic material to the top and make it short. Buyers do want the framework, and it costs you almost nothing to produce, so it should occupy the first fifth of the document rather than the middle three-fifths.

Spend the recovered time on observations that could only have come from having looked at this one submission. Three of those is a memorable deliverable; ten paragraphs of correct generality is a forgettable one, and now also a replaceable one.

Say what you could not assess. An honest limitation is something automated output structurally does not produce, because it has no model of its own competence in the way you do, and buyers notice the difference immediately.

The direction of travel

Nobody sensible should offer you a number for how this progresses, and this site is not going to invent one. The observable direction is that the boundary between "a model does this fine" and "this needs a person" has moved upward once and will move again, and it moves fastest through anything that is pattern-matching against a fixed rubric.

That suggests a defensive posture that is not defensive at all: keep the part of your offer that is a judgement call growing as a share of the whole, and treat any component you could fully specify as a checklist as something you are renting rather than owning. Objective measurement is a clear example - it is precise, specifiable and therefore not a place to build a differentiator, and the conventions for it are documented on the measurement property rather than being anyone's competitive advantage.

What this means for pricing decisions now

Do not lower prices in response to automation. The buyers you would keep by doing so are the ones who will use the free option next month anyway, and the buyers you actually have are paying for the thing that did not get cheaper.

Do reprice the components. If a tier of your rate card is mostly the automatable part, retire it rather than discounting it - the reasoning for retiring rather than cutting is worked through in the piece on rate cards you can defend.

And do watch your own numbers rather than the discourse. Falling conversion at a steady price is a real signal; an article claiming the trade is over is not. Tracking take-home per format will tell you which parts of your offer are eroding months before you could infer it from anything else.

For the platform-side picture of how automated and human assessment sit next to each other as options a buyer chooses between, Rate Cock's comparison of the two is the current source, and it is written for the person doing the choosing rather than the person being chosen.