The AI Proposal Scenario That Should Worry Your Sales Team

A sales rep asks an AI agent to pull pricing for a customer. The agent pulls from locked-down pricing systems, internal messages and old quotes, and to the agent the answer looks right. Then a second agent takes that data and writes the proposal for the rep, and nobody checks either handoff. The piece frames this as a scenario. What ends up in that proposal is the part that should worry you.

A pricing question, two AI agents, and a proposal nobody could trace

Here is the scene the piece lays out. It is a scenario, not a named company, and the piece uses it to show why AI governance needs to evolve. A sales rep asks an AI agent to pull pricing for a customer. The agent goes and gets it, pulling from locked-down pricing systems, internal messages and old quotes. To the agent, the answer looks right.

So a second agent takes that data and writes the proposal for the rep. And nobody checks either handoff.

In the scenario, the proposal turns out to hold unapproved, confidential and sensitive information. When that gets discovered, the real problem shows up: nobody can say where the bad data got in or how it moved across systems, files and agents.

Notice what broke. My read: not the model, the trail. And a missing trail is the one problem you cannot fix after the mistake is already made.

Why this is landing on sales floors right now

More companies are putting agentic AI into everyday work, and they keep running into the same wall. Their data is now constantly pulled in, copied, summed up, reshaped, reused and sent on again in new forms.

Plain version: your customer's information does not sit still anymore. A rep types a note. An assistant summarizes it. A quoting tool copies it. A follow-up email carries a piece of it somewhere else. By Friday the same number lives in five places, and some of them are wrong.

The piece says the old defenses do not cover this. Defenses built around apps, systems, networks or devices alone no longer get the job done. What teams need instead is data lineage, which means a working map of how data moves, so they can manage who owns what, who is accountable, and who can be trusted across human, agent and mixed workflows.

That is enterprise language for a small-business problem. You are not buying a data lineage platform for a four-person sales floor. You need to answer one question: where did this number come from? Hold onto it. It comes back in a minute.

A firewall guards a door. It cannot tell you which door a copied number walked through three weeks ago, and that is why the piece keeps landing on lineage instead of on another product you buy. My read: the map matters more than the lock.

The line in the piece that should stop you cold

One line in that piece does more work than the rest of it combined: "AI security isn't a tool problem; it's a data visibility problem."

Think about that as an owner. You can buy all the software you want. It will not fix a business where nobody knows where the numbers live.

Now the disclosure, because you deserve it. The piece comes from the custom content arm of MIT Technology Review, which is called Insights, not from its newsroom. It also states that humans researched and wrote it, and any AI tools were kept to production work under human oversight. My read: put the messenger aside for a second. Salespeople have been passing half-built quotes to each other forever. AI just makes the handoff faster and a lot quieter.

Picture it in a six-person insurance agency

A producer is chasing a commercial auto renewal. He asks an AI assistant to pull together everything the agency has on the account. The assistant builds him a clean one-page summary and drafts the renewal proposal from it. He reads page one, likes it, sends it.

Two days later the client asks about a line in that proposal. It is a line that was never supposed to leave the agency's files.

Now the producer has to reverse-engineer his own proposal. Which system fed that line? The carrier portal? A note a CSR typed after a call in March? A renewal email from last year? He cannot say, so he cannot explain it, and now he is guessing in front of a client.

My read: the danger in an agency is not an AI that writes something wrong. It is an AI that writes something right, pulled from a place it should never have looked, then hands it to you looking polished. Nobody loses an account over a typo. People lose accounts over a document they cannot defend. And if you cannot trace it, you cannot defend it.

The same thing plays out on a lot, only faster.

Picture it on a golf cart and trailer lot

A dealership with four salespeople. A rep gets a text at 8 p.m. from a customer who wants a number on a used four-passenger cart plus an equipment trailer. He does not want to bother the owner on a Saturday night, so he asks an AI tool to write the quote up from what is already in the system. It gives him a number. He texts it back.

Monday morning he cannot rebuild that quote. Tuesday the customer walks in, points at the text, and wants the same deal on a second unit. The rep now has two choices. Honor a number he cannot explain, or walk it back with a customer standing in front of him. Both cost him money, and one costs him the referral.

My read: on a lot, the killer is speed without a trail. The rep who answers the 8 p.m. text first usually wins the deal. The trouble is that he has to live with what he sent, and if he cannot show you how he got there, you are the one eating the difference.

Both stories point at the same thing the piece called out. Not a tool problem. A visibility problem. So here is what I would do about it this week.

How to use this in your business this week

1. Trace one real deal by hand. Pick something you quoted in the last ten days. Open every place it lives: your phone, the text thread, email, the CRM note, the quote sheet. Write down where each number in that quote came from. Give it 20 minutes. If you run out of time before you run out of places, you just found the piece's problem inside your own building.

1. Ask every rep to name the AI tools they used last week. Not the ones you bought. Every one, including whatever is free on their phone. Put the list on a whiteboard and leave it up. The list usually gets longer than owners expect.

1. Write down the three categories that never leave your building. In an agency that might be internal pricing rules, commission notes and underwriter comments. On a lot it might be cost, holdback and trade-in numbers. Three is plenty. Post it where the phones get answered.

1. Add the check that is missing. The piece's scene turns ugly because nobody checks either handoff. Someone has to be the last set of eyes on anything a customer sees. Assign one named person per deal. "The team" is not a person.

1. Feed your AI tools one approved price sheet. If a rep is drafting quotes with an assistant, that assistant should work from a single current sheet, not a folder of old ones. That is my take, not the piece's, but it is the cheapest control you will ever put in place.

1. Use the source question. Three prompts you can adapt today. For an insurance agency, tell the tool to draft a renewal proposal for a commercial auto client using only the pricing you paste in, to write MISSING and stop when a number is absent, and to stay out of every other file, email and note. For a golf cart and trailer dealership, ask for a follow-up text to a customer who got a quote on a used four-passenger cart and an equipment trailer, using only the price you supply, kept under 40 words and ending on one question. And for any team, ask the tool to list every source it used to build the quote and to flag any number that did not come from the price sheet you handed it.

That last one is the poor man's version of what the piece is asking big companies to build. It takes ten seconds and it puts a trail on the page.

1. Run a two-minute Monday audit. Pick one quote that left the building last week. Ask the rep to show you the trail. Two minutes, every Monday. Do it five weeks and your people will start keeping the trail before you ask, which is the whole point.

Mistakes to avoid

• Banning AI and calling it a policy. The tools are already on your reps' phones, and a ban you cannot enforce just moves the activity somewhere you cannot see it. My read: control the inputs, not the tools.

• Letting one rep be the test case. When one guy quietly experiments with quotes nobody else can read, you have built the piece's exact scene with one agent instead of two.

• Trusting the output because it looks clean. That is what happens in the piece — to the agent, the answer looked right. Pretty is not the same as approved.

• Assuming your CRM holds the whole story. If half your deals are settled over personal texts, your record has holes in it, and you will find the holes at the worst possible moment. That is a follow-up problem, and follow-up problems are where deals quietly die.