Your reps are not lazy, and the tools are not the problem. The problem is quieter than that. AI is very good at producing the artifact of preparation without producing the preparation, and the gap between those two things only shows up at the exact moment a deal is won or lost.
A sales leader looking at the surface metrics of an AI-equipped team usually sees improvement. Prep documents are longer. Discovery covers more ground. CRM hygiene is better than it has been in years. Notes get written the same day instead of the following Friday. Follow-up goes out in an hour. Every input you spent three years fighting for is finally there.
Then you look at win rates on competitive deals and nothing moved. Or it moved the wrong way.
This is the diagnostic we run when a team looks strong on paper and keeps losing deals it should win. The curation logic is narrow on purpose. Every sign below is observable in artifacts you already own: recorded calls, CRM notes, forecast commentary, deal reviews. None of them require a survey, a self-assessment, or an awkward conversation about whether someone is using AI "too much." And none of them are evidence that a rep is coasting. They are evidence that a specific kind of cognitive repetition has quietly left the job, and nobody replaced it with anything.
First, the Mechanism: AI Produces the Artifact, Not the Encoding
In 2011, Betsy Sparrow, Jenny Liu, and Daniel Wegner published a set of experiments in Science that gave us the phrase the Google effect. Participants typed statements into a computer. Half were told the file would be saved, half were told it would be erased. The group that expected the information to remain available remembered the statements noticeably worse, and remembered where the statements were stored noticeably better. Memory reorganized itself around access.
That finding is not a story about people getting worse at thinking. It is a story about the brain being efficient. When a fact is reliably retrievable from outside your head, encoding it inside your head stops being worth the metabolic cost. The brain is doing exactly what it should.
Set that next to one of the most replicated findings in memory research. In 1978, Norman Slamecka and Peter Graf demonstrated the generation effect: material a person produces themselves is retained substantially better than the same material read passively. The effort of generating is not a tax on learning. It is the mechanism of learning. Struggling and then producing builds a structure you can retrieve from later. Reading and then approving does not.
Put those two findings together and the shape of the problem appears. When a rep asks a model for the discovery questions, they get the discovery questions. What they do not get is the map of the account those questions would have been built from: the contradiction between what the VP said and what the org chart implies, the reason this buyer's timeline is really about something else, the three hypotheses the rep would have had to hold in mind long enough to choose one. The rep has the artifact. The encoding step got skipped.
On a call that goes according to plan, this costs nothing. The plan is right there. The moment the buyer says something the plan did not anticipate, there is nothing to retrieve.
Jeff Bloomfield, Braintrust's founder and the author of NeuroSelling, has made a version of this point about scripts for years: a seller who memorizes language can perform the language, and a seller who understands why the language works can build new language on the spot. AI has made the first one nearly free. It has also made the second one look optional.
1. The Notes Are Fluent and Contain No Buyer Language
Pull ten recent opportunity records and search them for quotation marks. AI call summaries are genuinely good, and they paraphrase by design, converting what the buyer said into clean neutral business register. What disappears in that translation is the buyer's own phrasing: the odd word choice, the hedge, the thing they said twice.
Those are the highest-value data in the entire call. A buyer's specific language is the retrieval cue the rep needs six weeks later when the deal stalls and they have to reconstruct what actually mattered. "The customer expressed concerns about the implementation timeline" and "she said, we cannot be mid-migration during open enrollment, that would end me" are not the same note. One is a summary. The other is a lever.
2. Discovery Gets Wider and Never Gets Deeper
Count follow-up questions, not questions. Generated discovery guides are lateral by construction, because breadth is what a model optimizes toward when it does not know which thread will matter. A rep working from that list tends to ask, receive an answer, tick the box, and move to the next topic.
The ratio to watch is second and third questions on the same thread, per call. When that number falls while total question count rises, the rep is executing a list rather than following a buyer. Depth is the part of discovery that has to be generated live, in response to something you did not know you would hear. A list cannot supply it, and a rep who has never had to build one has no habit of reaching for it.
3. Objection Handling Is Fast, Confident, and Slightly Off Target
Listen for speed. A rep working from a real internal model usually pauses before a hard objection, because they are choosing between several possible readings of what was just said. A rep retrieving a stored response answers immediately, and answers the category rather than the sentence.
The buyer says the timeline does not work because of a systems freeze in Q4, and the rep smoothly handles "timing objection" in general terms. Fluency without fit is the tell. It is also the hardest sign to catch, because in the room it sounds like poise, and it only reads as a miss on the recording.
4. The Rep Can Make the Case for the Deal but Not Cite the Evidence
Ask one question in your next deal review: what did they say that makes you believe that? A rep who did the cognitive work answers with a moment, a sentence, a person's reaction, a specific silence. A rep working from generated summaries answers with an inference dressed as an observation: "they are clearly feeling the pain," "the champion is bought in," "budget is not going to be an issue."
This gap is diagnostic because AI summaries are built to deliver conclusions, and conclusions are precisely the thing a seller should be producing themselves. When the conclusion arrives pre-made, nobody notices that the evidence underneath it was never examined. The deal review then inherits the same error, and the forecast inherits it from the deal review.
5. Call Prep Is Thorough and Completely Interchangeable
Take a rep's prep document, delete the company name, and hand it to a colleague working a different account. If it still reads as a perfectly reasonable prep doc, it was never prep. It was research, formatted.
Generated prep reliably produces industry context, plausible priorities, and sensible questions. What it cannot produce is the one thing that is only true of this buyer, this quarter, given what happened on the last call, because that requires holding the account in working memory long enough to notice a contradiction. The test takes ninety seconds and is very difficult to argue with.
6. Forecast Commentary Reads the Same Across Every Deal
Export the commentary field for a rep's entire pipeline and read it in one sitting. When a rep is thinking, the notes are uneven: some deals get three sentences of specific, uncomfortable risk, and others get the equivalent of a shrug. When a rep is generating, the commentary converges toward a house style, roughly the same three clauses in the same order with the nouns swapped.
Uniformity in forecast language is not discipline. It is a signal that the rep is describing the stage the deal sits in rather than the buyer the deal depends on. Slippage tends to follow, because the deal-specific risk was never articulated anywhere a manager could see it and act on it.
7. Reps Ask the Tool Before They Ask Each Other
Watch the team channels. Daniel Wegner's other well-known contribution to memory research is transactive memory: groups store knowledge inside each other, and part of a team's real capability is its shared index of who knows what. A sales team is a transactive memory system, and a good one is worth a great deal.
When "who has been through a security review at a bank this size" goes to a model instead of to the person two rows over, the rep gets a competent generic answer and the team loses a retrieval path it will need again. The observable version is a quieter deal-help channel, fewer hallway questions, and hard-won account knowledge that stops circulating even though nobody has left the company.
8. The Call Goes Quiet the Moment It Leaves the Plan
Find the point in a recording where the buyer introduces something unscripted: a new stakeholder, a budget change, a competitor nobody had mentioned. Then listen to the next thirty seconds.
Reps operating from an internal model bridge. They reflect the new thing back, connect it to something said earlier in the conversation, and ask about it. Reps operating from a generated plan acknowledge and steer back. This is the Google effect arriving in a live conversation: excellent access to the material, no encoded structure to improvise from, so the only available move is to return to the part they can retrieve.
9. Ramp Is Faster Than Ever and the Plateau Arrives Sooner
Look at the shape of the curve, not the first number on it. AI-assisted onboarding genuinely compresses time to first meaningful activity, and that is a real gain worth protecting. The pattern to watch for sits in months four through eight: a cohort that reached competent-looking output quickly and then stopped improving, while the previous cohort's curve was still climbing at the same point.
The generation effect predicts this exactly. The early errors, the clumsy first attempts at a value story, the call where the rep could not answer and had to go find out, are effortful production. Effortful production is what builds the structure that later performance draws on, and it is the same reason training that is never retrieved decays so fast. Remove the struggle and you have not accelerated the curriculum, you have deleted it.
The Nine Signs at a Glance
Use this as a scoring sheet on a single rep's last five deals, not as a team-wide audit. Three or more signs present in one rep's artifacts is a coaching conversation. Three or more signs present across most of the team is a system problem, and the system is yours.
| Sign | Where to Look | What It Usually Means | Repetition to Restore |
|---|---|---|---|
| 1. Fluent notes, no buyer language | CRM opportunity records | Buyer phrasing was paraphrased away before it was encoded | Rep writes three verbatim quotes before the AI summary is attached |
| 2. Discovery wider, never deeper | Call recordings | The rep is executing a question list rather than following a thread | Track follow-up questions per thread, coach to two minimum |
| 3. Fast, off-target objection handling | Call recordings, minute 20 onward | Stored response retrieved by category, not by content | Live practice on the specific objection, said in the buyer's words |
| 4. The case without the evidence | Deal reviews | Conclusions arrived pre-made and were never tested | Every claim in a review must be sourced to a quote |
| 5. Thorough, interchangeable prep | Prep documents | Research was formatted, not reasoned through | One written hypothesis per call that only fits this account |
| 6. Identical forecast commentary | Forecast notes, whole pipeline | The rep is describing the stage, not the buyer | Name one deal-specific risk in the rep's own sentence |
| 7. Tool before teammate | Team channels, deal desk volume | The team's shared memory has stopped being used | Route account-history questions to a named person first |
| 8. Silence off-plan | Call recordings, unscripted moments | Access without an internal structure to improvise from | Unscripted roleplay where the buyer deliberately goes sideways |
| 9. Fast ramp, early plateau | Cohort performance, months four to eight | The effortful errors that build skill were engineered out | Reintroduce produce-first reps before AI assistance is allowed |
How to Use This List Without Turning It Into a Witch Hunt
The failure mode of a diagnostic like this is a manager who reads it on a Sunday and spends Monday looking for someone to blame. That will cost you more than the original problem. Three rules make it useful instead.
Score the artifacts, not the people
Every sign here is a property of a document or a recording, which means it can be evaluated without anyone defending themselves. Read five opportunity records. Listen to two calls. Export one forecast. You will know within an hour whether this is a problem on your team, and you will know it from evidence rather than impression. That also keeps the conversation honest in the other direction: a rep who has all nine artifacts clean is using AI extremely well, and should be saying so out loud in your next team meeting.
Put generation before the tool, not instead of it
This is the whole fix, and it is smaller than it sounds. The generation effect does not require that a rep do the work alone. It requires that they produce something first. Two minutes of writing three discovery hypotheses before opening the AI research brief. One sentence naming the buyer's real risk before reading the generated summary. A verbal answer to an objection before pulling the recommended framing.
The rep still uses the tool, and the tool still improves the output. What changes is that there is now something in the rep's head for the tool's output to attach to. Compare-your-answer is a far stronger encoding event than read-the-answer, and it costs about ninety seconds per call.
Coach the retrieval, not the output
Most sales coaching reviews the artifact: the deck, the email, the call plan. When AI is in the workflow, artifact quality is no longer a reliable signal of rep capability, because a strong artifact and a weak schema now look identical from the outside. So change what you inspect. Ask reps to explain, without notes, why the account is going to buy and what would have to be true for it to fall apart. That is a retrieval test, and retrieval is where skill either exists or does not.
This is also why practice has to be unscripted to be worth anything now. Rehearsing a known objection with a known answer builds very little, because the answer is already retrievable. Practice earns its cost when the buyer goes somewhere the rep did not prepare for, which is the only condition under which the rep has to generate.
What This Is Really About
None of this is an argument for using less AI. The reps on your team who lean on these tools are usually the ambitious ones, and the efficiency they are getting is real. Prep that used to take forty minutes takes six. That time exists now, and it should be spent on something.
The mistake was assuming it would automatically get spent on thinking. It does not, because nothing in the workflow asks for thinking anymore. The repetitions that used to be forced by the job, sitting with an account until a hypothesis formed, fumbling an objection badly enough to remember it, writing a call plan from a blank page, were never on anyone's development plan. They were just the friction of the work. When the friction went away, the development went with it, silently, and the artifacts kept looking fine.
The teams handling this well are not restricting the tools. They are putting a small, deliberate act of generation in front of every place a tool now produces an answer, and they are coaching retrieval instead of output. That is a design decision, and it has to be made on purpose, because the default is drift.
If you are seeing three or more of these signs in your own team's calls and CRM, that is worth a conversation. We help sales organizations rebuild the specific repetitions that AI removed, without giving up any of the speed. Start a conversation with Braintrust and we will walk your artifacts with you.
