Cognitive Control vs Cognitive Surrender

The story

Steve used to let his suppliers draft the Statement of Work. It arrived in Word, mostly boilerplate, a few specifics dropped in. Legal would mark it up. Procurement would push back on payment terms. It was slow, but the scope was grounded in what the supplier knew how to deliver.

Then Steve discovered generative AI. Instead of waiting for the supplier's draft, he prompted ChatGPT to write the SoW himself. The output was remarkable. It proposed workstreams he had not considered. It included governance structures, acceptance criteria, risk frameworks. It covered angles that made Steve feel like he was finally in control of defining the work.

He was not. He was reading what a language model had invented based on statistical patterns in its training data. And because it was articulate and thorough, it became the starting point for the engagement. The supplier priced against it. The project plan followed it. Six months later, Steve was managing a programme twice the size of the problem he originally needed solved.

This pattern has a name. We call it cognitive surrender.

What the research says

The cognitive science is clear. Microsoft Research surveyed 319 knowledge workers and found that higher confidence in generative AI output correlates directly with less critical thinking by the user. MIT's Media Lab went further, measuring brain activity via EEG: participants who delegated writing to ChatGPT showed the weakest cognitive engagement across all test groups, an effect that persisted even after the AI was switched off. The researchers called it "cognitive debt."

This matters because every time a large language model generates a scope document, it produces a statistically unique output. That is not a flaw. It is how the technology works. Each generation is a fresh probabilistic walk, and the result is a document that shares structural similarities with others but matches none of them precisely. For a buyer drafting their own SoW with AI, this means every engagement starts from a novel description of work that has never been reviewed, never been priced before, and never been tested against delivery reality.

The commercial cost of that novelty is already quantified. WorldCC research puts contract value leakage at 9 to 11% of annual revenue, driven by ambiguous terms, missing clauses, and scope gaps. PMI found that 47% of project failures trace to poor requirements management, with 52% of all projects experiencing scope creep.

These are not post-signature problems. They are creation-stage problems. And generative AI is amplifying them at the exact moment buyers believe they are solving them.

The paradox of unlimited options

There is a deeper mechanism at work. Psychologist Barry Schwartz's research on choice overload demonstrated that when options multiply, decision quality degrades. Baumeister's work on decision fatigue showed that self-regulatory capacity is depletable: the more decisions you force in sequence, the worse the later ones become.

Generative AI is the ultimate choice-overload machine. Every regeneration produces a new option space. Every prompt revision creates another plausible scope. The buyer sits in front of an infinite document factory with no basis for choosing between outputs other than which one reads best. Reading well and scoping accurately are not the same thing.

Cognitive control

The alternative is not to go back to waiting for the supplier's Word document. It is to structure the authoring process so that the buyer defines the work through guided questions rather than freeform generation.

Deployed's Document Assembly platform does this. The buyer answers a structured questionnaire. Conditional logic selects the right clauses from a pre-approved library. The document assembles itself from components that legal has already reviewed, procurement has already priced, and delivery has already validated. Variation happens where variation is justified. Everywhere else, the enterprise's own standards apply.

A structured questionnaire that fixes the boilerplate frees the author to focus on the genuinely novel scope content. Generative AI does the opposite. It spreads attention across boilerplate and substance alike, treating both as equally open to reinvention.

But the real value is upstream. Questionnaire-driven documents produce structured data as a by-product: engagement type, deliverable count, risk tier, jurisdictional flags. This is exactly the substrate that downstream AI tools need to extract obligations, score risk, and build portfolio analytics. Freeform AI drafting produces only prose.

The irony is sharp. Organisations that want AI to improve their contract intelligence are simultaneously using AI to destroy the data quality that intelligence requires.

Where AI actually belongs

None of this means AI has no place in scope documentation. It means AI belongs after the structure, not before it. Once a document exists as structured, classified, audit-trailed data, AI is genuinely useful for obligation extraction, deviation tracking, and negotiation analytics.

Steve did not need a more eloquent SoW. He needed the right questions asked before the first word was written. Generative AI gave him a document that felt like control. Document assembly would have given him actual control.

To learn more, contact us or connect on LinkedIn to find out more.