Dear B2B CEO,
Last quarter you lost a deal to a firm you consider worse than yours.
The prospect was qualified. They never took a meeting. They went somewhere else, and when you asked, sales told you it went on price.
You filed that under excuses, which is where it has belonged for most of your career. Any sales leader can say “we lost on price,” and most of the time it means “we lost, and I would rather not examine why.”
The excuse is turning into a description. The reason is more specific, and more fixable, than AI coming for your industry.
What You Actually Sell
There is a name in economics for the kind of thing you sell.
In 1973, Michael Darby and Edi Karni published a paper about markets where fraud is a structural feature rather than a moral failure. They needed a word for goods the buyer cannot fully judge even after receiving them. They called them credence goods.
Their examples were car repair, surgery, and legal advice. You pay, something happens, and you have no reliable way to know whether it was the right something, or whether you needed as much of it as you got.
Now think about what your client can actually check after you deliver.
They can tell whether it works. They can tell whether it shipped on time, whether it holds up under load, and whether their people use it. All of that is visible, and if you get it wrong you hear about it inside a month.
Here is what they cannot check. Whether the architecture you chose was the right one. Whether the fourteen hundred hours you billed needed to be fourteen hundred. Whether the custom integration you scoped was necessary, or whether something off the shelf would have carried them for three years. Whether the senior people who came to the pitch did the work. What the maintenance will cost them in year four.
They can hire someone to audit it. That is real, it happens, and it costs a fraction of the alternative, which would be paying a second firm to build the same thing a different way and comparing the two.
An audit will tell them whether the code is sound, whether it is documented, and whether the security holds. Worth doing, and too few clients do it.
It will not tell them whether the fourteen hundred hours were fourteen hundred hours, because effort does not survive into the artifact. It will not tell them whether the scope was necessary, because answering that means reconstructing a decision made a year ago from information nobody wrote down. And when the auditor says they would have built it differently, your client is holding one expert’s opinion set against another’s, which is not the same as an answer.
Underneath that sits the part you cannot spend your way out of.
The audit is a credence good too. Your client cannot check the checker.
And the auditor who would like the remediation work has a reason to find something.
Darby and Karni built this into the definition. A good stays a credence good when checking quality is impossible, or costs more than the answer is worth.
So your work is verifiable on the surface and unverifiable underneath. The surface is where your client forms opinions. The layer beneath holds most of the money and nearly all of the judgment.
And the smaller your client is, the further that goes. A buyer with a strong CTO can commission an audit and read it properly. A buyer without one is choosing who to trust, which describes most of the mid-market.
The same split runs through a systems integration, a managed service, and a strategy engagement. If you sell a product rather than work, your credence layer is implementation, support, security posture, and your roadmap. Everything the buyer cannot trial.
This Was Never Your Problem
Here’s the thing. You have probably read that as a description of a difficulty. It is a description of your margin.
Because that lower layer cannot be checked, buyers in your market have always substituted something else. Reputation. Referrals. Whether the person across the table seemed like someone who had done this before. Whether three people they trust said your name without being asked.
Twenty years of relationship selling in B2B services was always a rational response to an information problem.
And it protected you. That substitute took a decade to build and no competitor could copy it in a quarter, which is a better moat than most product companies will ever have.
Reputation was doing a job. It was the mechanism that kept that unverifiable layer honest, because a firm that over-scoped or under-delivered got found out eventually and stopped getting referred.
Hold on to that word. Mechanism.
There Are Two Judges Now
Somewhere between your prospect deciding they have a problem and a human being deciding to sign, a machine now sits.
It reads your website. It drafts the requirements. It builds the comparison table. Often it writes the RFP itself: when Deloitte asked chief procurement officers what they use generative AI for, generating RFPs and RFQs came second on the list, named by 42% of them.
That machine cannot see your work.
What Happens When the Machine Buys
The limits first, so you can weigh this properly.
In March, two economists at Göttingen, Alexander Erlei and Lukas Meub, ran the first study I have seen that puts language models on both sides of a credence goods market. Six hundred simulated one-shot markets, four sellers against four buyers, plus repeated rounds of sixteen.
It is a simulation. No real money, no real vendors. It tested a single model, GPT-5.1, so nothing in it is a property of “AI” in general. The goods are abstract rather than software projects. The human comparison comes from a separate 2011 experiment instead of a control group. It is a preprint, and it has not been peer reviewed.
I am telling you all of that before the findings because the conclusion is uncomfortable enough that you should get the chance to discount it yourself.
Here is what they found.
The machine buyers fixated on price and could not read what a markup meant. The authors write that they focus “narrowly on price levels rather than understanding strategic incentives embedded in markups, making them vulnerable to exploitation.” Prices in the agent markets ran well below the human benchmark, and the buyers piled onto whichever seller was cheapest rather than spreading across the market.
Then the finding that should hold your attention.
The machine buyers did not punish sellers who cheated them. They kept going back.
The paper says it plainly: these buyers “do not exploit the potential to discipline identifiable experts, and instead continuously approach experts despite economic losses once prices are low enough.”
Reputation did not fix it. In human versions of this experiment, reputation reliably reduces overcharging. With the agents, the effect was context-dependent, with no consistent benefit. Verifiability, the other classic remedy, provided “little benefit.”
That is the mechanism failing. The thing that has kept your market honest, and kept you paid, is a buyer who remembers being cheated. This one does not.
The First Way to Lose
So there is an obvious move, and an industry is already selling it to you.
Shape everything for the machine. Write the pages it likes. Fit the criteria. Get the certifications listed, the numbers formatted, the words in the right places. Be legible, whether or not you are any good.
And it works. That is the part worth sitting with. The research says the doorman does not punish you for playing it. You get through.
Then a human being walks into the room.
Gartner surveyed 645 B2B buyers in late 2025. Forty-five percent used generative AI in a recent purchase. Sixty-nine percent said they prefer to validate what the AI told them with a sales rep. Fifty-one percent said they are more likely to run into misleading information from generative AI than from a salesperson, which is a striking thing for buyers to admit about a tool they keep using anyway.
Your buyers do not trust the machine either. They use it to narrow the field, and then they check.
So you arrive in a room with someone who has bought this six times and can tell inside twenty minutes whether you have done the work. Everything you shaped for the filter is now in front of a judge who can see.
You lose there. And nobody in your company attributes that loss to the door, because the funnel looked healthy. Sales gets asked why conversion is soft.
The Second Way to Lose
Fine, you think. Then be real.
Do excellent work, keep your word, let the results speak, and let referrals arrive the way they always have. That is the honest answer, it is what I spend most of my time arguing for, and it is what actually wins the room.
On its own, it is also a way to disappear.
Reputation lives in people’s heads. The doorman cannot read heads.
It reads what is written down.
So the firm with fifteen years of quietly excellent delivery, whose proof exists as goodwill inside a few hundred people, does not make the shortlist the machine assembled from what it could parse. You never reach the room where you would have won.
That firm cannot diagnose itself either. Inbound conversations thin out over eighteen months, and the explanation becomes brand awareness, or the market, or the year.
It was a door.
Now go back to the deal you lost last quarter, the one you filed under excuses. You do not know which of these two things happened to you. That is the actual problem.
One Record, Two Judges
Both firms are half right, which is why both are stuck.
You cannot pass the door by being good, because the door cannot see good. You cannot pass the room by being legible, because the room can see everything else.
The useful part is that fixing this does not take two separate efforts.
What the machine needs is specifics it can parse. What the human needs is proof you have done this before. Those are the same material.
Write down what is already true about how you work. What you built, for what kind of company, under what constraint. What it cost and how long it took. What went wrong, because something did, and what you did about it. Why you chose that approach over the two you rejected.
Adjectives will not carry it. Adjectives are what your competitors have. Every firm in your category claims to be senior, pragmatic, and partner-led, which means the machine reads all of you as identical and falls back on the one number that differs.
Specifics are what a machine can check and what a buyer cannot fake having read.
Why Most Firms Do Not Have This
Because it is nobody’s job in the quarter it would need to start.
A record like that exists only if someone captured it while the work was happening, project after project, for years. It cannot be assembled in the six weeks after you notice you need it. By then the details are gone, and so are half the people who knew them.
That is what structure is for, and it is the least glamorous part of what I do. It is the plain discipline of capturing what is true, on a cadence, so that in three years you hold something no competitor can produce on demand.
Random acts of marketing produce a case study when somebody remembers. Consistency produces a body of evidence.
What a Machine Can Check
Name the three things that make you worth more than the cheapest firm on the list. Not adjectives. Three specific things you do that they do not.
Then go and find where a machine could verify each one without ever speaking to you.
If you cannot find them, it cannot either. And it will do the only thing left to it, which is to compare the one number that is always legible.
Yours sincerely,
