Information Asymmetry

Why does this matter to founders?#

It shows up as a price, not as a complaint. An unknown vendor can be objectively better and still receive a lower willingness to pay, because the buyer cannot verify security, reliability, or implementation quality. That gap is part of your acquisition economics — see Willingness to Pay for how to measure the discount rather than argue with it.

Marketplaces attract the wrong supply if pooled pricing persists. When good and bad providers receive the same price but good provision costs more, quality supply leaves. Ratings help only when identities persist, outcomes are recorded against real transactions, and manipulation is policed. This is the operating problem behind trust and safety in any marketplace.

Pricing is a screening instrument, not only a monetisation one. Plan limits, deposits, annual commitments, trials, and guarantees change which customers select an offer and how they behave afterwards. That is the same machinery as price fences, aimed at type rather than at willingness to pay.

Fundraising is an asymmetric-information process by construction. You know more about pipeline quality, technical debt, concentration, and retention risk. Investors respond with diligence, staged financing, milestones, and governance — screening, in other words. Reconciled cohort data and consistent metric definitions reduce the uncertainty discount; inconsistent ones widen it. See Startup Valuation Methods and Term Sheets.

How do you design a mechanism?#

1. Separate hidden type from hidden action#

Before exchange: quality, risk, intent, and ability to pay. After exchange: effort, care, usage intensity, and honest reporting. Adverse selection is fixed by verification and self-selection; moral hazard is fixed by incentives, monitoring, and staging. Applying the wrong one is the most common design error.

2. Quantify the pooling loss, then iterate#

pooled value = Σ (probability of type × value of type)

Compute which types still find participation worthwhile at that pooled price, remove the ones that do not, and recompute the pool. The iteration is the lemons problem. If it converges on a market containing only your worst supply, you have located the real constraint.

3. Make the separating condition explicit#

For a signal to separate, the high type must gain from sending it and the low type must not:

benefit_high − cost_high(signal) > benefit_high without signal

benefit_low − cost_low(signal) ≤ benefit_low without signal

Both lines must hold. If you cannot write the second one truthfully, you have a marketing asset, not a signal.

4. Match verification intensity to the risk at stake#

Samples, trials, reference calls, security reports, penetration tests, usage evidence, third-party audits, identity checks, milestone payments, escrow, warranties, and outcome monitoring form a ladder of increasing cost and increasing informativeness. Buying more verification than the decision warrants is a real cost that shows up in cycle time and services margin.

5. Close the learning loop#

Record outcomes, allow correction and appeal, and update reputation. Punish manipulation, not negative feedback — suppressing bad reviews is both an information failure and, in the US, a violation of the FTC's Consumer Reviews and Testimonials Rule.

Worked example: when does an audit pay for itself?#

A buyer faces two equally likely vendor types. A reliable implementation is worth $20,000; an unreliable one is worth $8,000. A reliable vendor's delivery cost is $12,000 and it will not sell below $16,000; an unreliable vendor will accept $7,000.

Step 1 — The pooled market unravels#

pooled value = 0.5 × $20,000 + 0.5 × $8,000 = $14,000

At a ceiling of $14,000 the reliable vendor cannot clear its $16,000 floor and exits. Buyers infer the remaining pool is unreliable and willingness to pay falls toward $8,000. Quality investment earns nothing. This is Akerlof's mechanism, in four lines.

Step 2 — Add an independent audit#

An audit costs $1,000 and passes 90% of reliable vendors and 10% of unreliable ones. After a pass, by Bayes' rule:

P(reliable | pass) = (0.5 × 0.9) / [(0.5 × 0.9) + (0.5 × 0.1)] = 0.45 / 0.50 = 90%

buyer's expected value after a pass = 0.9 × $20,000 + 0.1 × $8,000 = $18,800

Step 3 — Check that the mechanism is worth running for both sides#

Both sides gain, so the mechanism holds. The audit created $4,000 of price uplift against the $14,000 pooled ceiling, of which $3,000 is net of its cost.

Step 4 — Stress-test the audit, because this is where it breaks#

Suppose the audit is weaker: it passes 80% of reliable vendors and 30% of unreliable ones.

P(reliable | pass) = 0.40 / 0.55 = 72.7%

expected value after a pass = 0.727 × $20,000 + 0.273 × $8,000 = $16,727

The reliable vendor now needs $16,000 + $1,000 = $17,000. The buyer will pay at most $16,727. The mechanism fails — reliable vendors still exit, and the market still unravels, despite an audit existing and being widely used.

The lesson is uncomfortable and specific: the value of verification lives almost entirely in its false-positive rate. A certification that almost everyone passes conveys almost nothing, however expensive it is to obtain. Between the two scenarios nothing changed except audit quality, and the market flipped from working to failing.

What the arithmetic hides. It assumes the buyer knows the prior probabilities and the audit's error rates, that value is a point estimate rather than a distribution, and that a single verification event settles the question. It also addresses selection only — a passed audit says nothing about whether the vendor will still be trying in month nine. That is moral hazard, and it needs a different instrument: milestone payments, a monitored pilot, a warranty, or an outcome-linked component.

For a SaaS founder, the "audit" is usually a scoped proof of value with pre-agreed success criteria. Its economic value is the price or conversion it unlocks, minus its delivery cost and the sales delay it creates — not the number of pilots run. That cost is real: Snowflake's professional-services line ran at a (31)% GAAP gross margin in fiscal 2026.

What are the common mistakes?#

1. Treating disclosure as verification. A 60-page security document that nobody has the expertise to evaluate transfers no information.
2. Calling brand a signal. A signal must be differentially costly or valuable by type. Spend that a low-quality competitor could match equally is not separating.
3. Running reputation without transaction verification. Ratings not tied to a completed, identity-bound transaction are cheap to manufacture — and, in the US, incentivising positive sentiment or suppressing negative reviews is separately unlawful.
4. Fixing moral hazard with better screening. Once the contract is signed, the problem is effort, not type. Add milestones, monitoring, or an outcome-linked component instead of another reference call.
5. Over-verifying. Every additional check lengthens the cycle and costs margin. Match verification to the loss at stake, not to internal anxiety.

When does the mechanism break?#

• The signal becomes cheap to imitate. Certifications degrade into box-ticking once the pass rate approaches 100%. Track the false-positive rate, not the adoption rate.
• The verifier has a conflict. An auditor paid by the audited party, or a rating system monetised by the rated party, produces predictable drift.
• The guarantee is not collectable. A warranty from a company without the capital or the legal exposure to honour it is not a bond.
• Identities reset. Reputation only works when a bad history cannot be discarded by re-registering.
• The screen excludes good participants. Deposits, commitments, and scoring models systematically filter out cash-constrained, new, privacy-sensitive, or under-represented participants. Track the selection effect and provide an appeals path — otherwise the screen quietly shrinks your addressable supply.
• Transparency creates a new exposure. Sharing security architecture, customer lists, or model details to build trust can leak competitively or legally sensitive information. Share the minimum needed for the decision, aggregated and access-controlled.