In brief
Why it matters now
- Every AI-assisted team says it keeps a human in the loop. A human who doesn't understand the domain catches nothing.
- Workslop, polished output which doesn't move the work forward, is a validation failure. It moves the cost from the sender to the recipient, and in presales the recipient is often the buyer.
- The skill most at risk from careless AI use is the one validation depends on: diagnosing what's wrong.
- How people use AI decides whether they learn. Asking AI to explain preserves understanding. Handing it the whole task erodes it.
- Match validation to risk. A claim in a security questionnaire needs a named expert and a source. An internal note needs the sender to have read it.
- Build depth while you work: predict before you prompt, explain back, practise without AI, and turn AI errors into team training.
A human in the loop needs to know what they're looking at
"Human in the loop" has become the standard reassurance for AI-assisted work. It describes where a person sits in the process. It says nothing about whether the person is able to spot an error.
The evidence on this point is consistent. In a pre-registered field experiment with 758 Boston Consulting Group consultants, AI raised output and quality on tasks inside its capabilities. On more complex tasks, consultants using AI were 19 percentage points less likely to produce the right answer. The researchers called this uneven pattern a "jagged frontier": AI is excellent at some tasks and confidently wrong at others of similar difficulty, and the boundary isn't marked.
Confidence makes it harder. In a 2025 randomised trial, experienced developers took 19% longer with AI on real tasks, yet believed it had saved them 20% (METR). How productive people feel and how productive they are diverge, and the same gap applies to how accurate they feel.
For a Sales Engineer, the frontier runs straight through the work. AI drafts discovery summaries, architecture outlines, integration notes, RFP answers, and security questionnaire responses at speed. Most of the drafting is useful. Some of it states things which aren't true, in exactly the same confident tone.
Buyers have noticed. In Gartner's 2026 buyer research, 51% of B2B buyers expect misleading information from generative AI, and 49% expect it from sales reps. Your team earns its credibility one checked claim at a time.
Workslop is a validation failure
Researchers at Stanford and BetterUp gave the problem a name: workslop, AI-generated work which looks finished and doesn't advance the task. In their 2025 survey of 1,150 US desk workers, about 40% had received some in the past month. Each incident took close to 2 hours to deal with, and about half of recipients rated the sender as less capable afterwards.
Workslop doesn't remove effort. It moves the effort to whoever receives the work. Inside a team, the recipient is a colleague who rewrites the proposal section. In presales, the recipient is often the buyer: a security reviewer who finds an answer contradicting the product documentation, or an architect who notices an integration described for an API version which doesn't exist.
Be kind about where workslop comes from. It usually comes from people working hard to keep up, with more tools, wider spans, and the same quarter-end. Blame produces hiding. A clear standard protects the sender as much as the recipient, and the standard works only if the sender understands the work well enough to check it.
Validation depends on the skill careless AI use erodes
In January 2026, Anthropic published a randomised controlled trial with 52 mostly junior software engineers learning an unfamiliar Python library. The AI group averaged 50% versus 67% for the hand-coding group, a significant difference (p = 0.01, Cohen's d = 0.74). The largest gap was on debugging questions. The AI group finished about 2 minutes faster (roughly 23 vs. 25 minutes), but the difference was not statistically significant.
The largest gap sat in debugging: finding what's wrong and why. Debugging is validation.
The pattern repeats outside software. Endoscopists who used AI routinely saw their unassisted detection rate fall from about 28% to about 22% within months (Lancet Gastroenterology & Hepatology, 2025). Among 319 knowledge workers, higher confidence in AI went with less critical thinking about its output (Microsoft Research and Carnegie Mellon, 2025).
The consequence for technical sales is direct. An SE who no longer reasons through an architecture unaided is less likely to spot the flaw in an AI-drafted one. A team where nobody holds the depth has no one left to check the work.
How you use AI decides whether you learn
The Anthropic study holds the most useful finding for anyone leading technical people. Using AI didn't guarantee a lower score. How people used it made the difference.
| How people used AI | Average quiz score |
|---|---|
| Generated code, then asked follow-up questions to understand it | 86% |
| Asked for code along with explanations | 68% |
| Asked only conceptual questions and worked through errors themselves | 65% |
| Handed all the code writing to AI | 39% |
| Started with questions, then handed everything over | 35% |
| Relied on AI to debug and verify the code | 24% |
Source: Anthropic, January 2026. Groups of 2 to 7 people each.
Read the table with care. The authors note this is correlational, not causal, and the sample was small. The quiz measured learning right after the task. The direction still matches decades of learning research: people retain what they work to understand.
The lowest score belongs to the group who used AI to check their work. Delegating validation to the same system which produced the output is the pattern to design out of your team.
The authors' recommendation for organisations: deploy AI tools deliberately, not only aggressively. Productivity gains put at risk the skills needed to validate AI-written code, especially for junior engineers.
Match the validation to the risk
Validating everything to the same depth wastes the time AI saved. Set the depth by what happens if the output is wrong.
| Output | Risk if wrong | Validation standard |
|---|---|---|
| Internal notes and meeting summaries | Low: wasted colleague time | The sender has read it and says what AI did |
| Qualification fields (MEDDPICC and similar) | Forecast error | Every field links to evidence: a quote, an email, or a document |
| Discovery and account research | A wrong first impression with the buyer | Facts checked against primary sources before the call |
| Architecture and integration designs | A failed POC or a failed implementation | A qualified SE reasons through it unaided, then tests it in a lab or POC environment |
| Demo scripts and data | Lost credibility in the room | Rehearsed end to end by the presenter |
| Security, compliance, pricing, and roadmap claims | Contractual, legal, or reputational exposure | A named expert verifies each claim against a current source document before it goes out |
The last row admits no exceptions: nothing AI-generated reaches a customer without a named person checking it.
Build depth while you work
Validation and development use the same habits. Each one below catches errors today and builds the knowledge to catch them tomorrow.
- Predict before you prompt. Write your own answer, outline, or architecture sketch first, then compare it with the AI's. The differences show you what to check and what to learn.
- Ask why, then explain it back. Ask AI to explain its reasoning. Then explain the result in your own words to a colleague or the customer's architect. If you're unable to explain it, you haven't validated it.
- Practise without AI on purpose. Hold whiteboard architecture sessions, discovery role-plays, and objection drills with no tools. In Microsoft's 2026 Work Trend Index, advanced AI users deliberately did some work without AI 43% of the time, against 30% of other users.
- Run error hunts. Once a fortnight, give the team an AI draft with seeded or real errors (an RFP answer, an integration design, a POC plan) and find them together. Workslop becomes training material.
- Teach back. Each week, one person teaches the team something they learned while checking AI output. Knowledge moves from one person's validation to the whole team.
- Redesign the junior route. Research, preparation, and first drafts were how new SEs learned, and AI now does them first. Replace the lost practice with shadowing, live practice, and side-by-side review of AI drafts against expert work.
What leaders measure
Usage figures show adoption. They don't show whether your team is getting better or worse at the work. Track these each quarter instead:
- Errors caught before the customer, and errors the customer found first.
- Rework on proposals, RFP answers, and POC plans.
- Capability without AI, through periodic discovery role-plays and whiteboard sessions scored against a rubric.
- Time to competence for new hires, now they have fewer routine tasks to learn on.
Managers set the bar. In the same Microsoft study, managers of advanced AI users set explicit quality standards for AI-assisted work 83% of the time, against 57% for others. Our research report, "You, your team, and AI", sets out the wider management picture behind these numbers.
The expert in the loop
AI makes a strong team faster. It makes a team without depth confidently wrong, faster. The difference is the person checking the work, and whether they still understand it well enough to check.
Look at the last customer-facing answer your team sent. Who checked it, and would they have been able to write it themselves?