AI is driving the marginal cost of competent generation toward zero. That single fact reprices every business, because the value of anything you sell is partly the competence embedded in it — and the part that was competence is becoming free. The work of future-proofing is not defending against a vague threat. It is a specific, scoreable exercise: decompose what you sell into activities, mark which activities are made of the competence AI now commoditizes, and deliberately migrate value toward the things that stay scarce.
Five things stay scarce when generation is free: verification and trust, taste and judgment, proprietary data, distribution, and accountability. Everything in this guide is a way to move your business off the first list and onto the second. I run this audit on my own companies — Kommerce, a cash-on-delivery commerce OS, and Velya, which builds AI agents that qualify leads — and I will use them where they show the mechanism honestly.
Work through it in order. By the end you will have a scored exposure map, a commoditization verdict on your core product, a ranked list of repositioning moves, and a one-page worksheet you can run again next quarter.
The principle: value flows to what AI can't supply
Start from the mechanism, because it tells you exactly what to look for. AI is extraordinary at generation — producing a competent first version of almost anything: code, copy, a legal clause, a lead-qualification script, a plausible research hypothesis. It is far weaker at knowing whether the generated thing is correct, safe, and true in your specific context. Generation had a scaling law. Verification does not. When generation gets cheap and abundant, the binding constraint flips to verification, and value follows the constraint the way water follows a slope. This is why, when generation is free, verification becomes the thing worth paying for — the scarce resource is no longer producing the artifact but certifying it against reality.
The same collapse hits taste from the other side. When everyone can generate a hundred competent options for a penny, the scarce skill is choosing the right one — knowing which of the hundred is actually good, on brand, and correct for this customer. Generation went to zero; selection did not. That is why taste and judgment become the last moat when competence is free: the curation layer is worth more precisely because the raw material underneath it is worthless.
Here is the whole principle in one table. Learn to feel which column any activity belongs in.
| AI commoditizes (value leaving) | AI can't supply (value flowing here) |
|---|---|
| Producing a competent first draft | Verifying it's correct in context |
| Generating many options | Choosing the right one (taste) |
| Applying general knowledge | Proprietary data no one else has |
| Writing the feature | Owning the distribution to the customer |
| Executing a defined task | Holding accountability for the outcome |
None of this is a forecast. It is already priced into what customers will pay: the generatable half of your offering is deflating right now. The audit is how you find your exposure before the market finds it for you.
Phase 1: The exposure audit
Decompose the business, then score each piece. Do not score "the company" — the company is an average that hides everything useful. Score activities.
Step 1 — List your value-creating activities
Write down 8 to 15 discrete activities your business actually gets paid for. Be concrete. Not "we do marketing" but "we write outbound sequences," "we design the brand system," "we run the ad account." Not "we build software" but "we generate the CRUD screens," "we design the data model," "we handle the edge cases in dunning." For a services business, list the deliverables. For a product, list the jobs the product does for the user. If you can't name 8, you haven't decomposed finely enough — the exposure hides inside the lumped-together lines.
Step 2 — Score each activity on two axes
For every activity, assign two scores from 1 to 5.
- Commoditization exposure (C): How much of this activity is competence AI now supplies for cheap? A 5 means a current model does a near-complete version from a prompt. A 1 means the model can't meaningfully touch it.
- Scarcity anchor (S): How much of this activity's value depends on something AI can't supply — proprietary data, verified trust, distribution you own, accountability you carry, or taste that's hard to specify? A 5 means it's almost entirely anchored to one of those. A 1 means it's floating free with nothing underneath it.
The gap between them is the signal. Compute Exposure = C − S for each activity, ranging from −4 (deeply insulated) to +4 (acutely exposed).
| Activity | Commoditization (C) | Scarcity anchor (S) | Exposure (C−S) | Read |
|---|---|---|---|---|
| Draft outbound email copy | 5 | 1 | +4 | Acute — pure generation, no anchor |
| Generate standard app screens | 5 | 2 | +3 | High — model does it, thin lock-in |
| Qualify inbound leads (Velya) | 4 | 4 | 0 | Contested — AI does the talking, but the value is the verified handoff and the clinic relationship |
| COD trust-scoring (Kommerce) | 3 | 5 | −2 | Insulated — anchored to proprietary delivery-outcome data |
| Own the merchant relationship | 1 | 5 | −4 | Deep moat — distribution AI can't intermediate |
Two entries deserve a note, because they show the audit doing real work rather than flattering the founder.
Velya's lead-qualification conversation scores a 4 on commoditization — a frontier model can hold that conversation. If Velya were only the conversation, it would be acutely exposed. What pulls its scarcity anchor up to a 4 is everything around the conversation: the verified handoff into the clinic's calendar, the accountability for a qualified booking actually showing up, the accumulated data on which lead patterns convert for this clinic. Net exposure lands near zero — which is the honest read. The product is neither safe nor doomed; it lives or dies on whether we keep deepening the anchor faster than the conversation commoditizes.
Kommerce's trust-scoring scores lower on commoditization not because the algorithm is exotic — plenty of models can rank a risk score — but because the input is a proprietary dataset: which addresses, phone patterns, and order shapes actually resulted in a successful cash-on-delivery handoff across markets where that data barely exists publicly. AI commoditizes the modeling. It does not hand you the data. That is the difference between an exposed activity and an anchored one.
Step 3 — Find your eroding moat
Now aggregate. Sort activities by Exposure descending. Everything at +2 or higher is competence you are currently charging for that the market is about to stop paying full price for. Then ask the harder question about the anchored ones: is the anchor actually holding, or is it a story? Many "moats" are just capability leads wearing a costume. Network effects in particular get over-claimed — most of what founders call a network-effects moat isn't one, and if your scarcity score is resting on an assumed network effect, discount it hard until you can show the retention curve that proves it.
Output of Phase 1: a ranked list of exposed activities, and an honest read on which anchors are real. Keep it. Phase 4 acts on it.
Phase 2: The commoditization test for your core product
The activity audit tells you where value is leaking. This test tells you whether your central bet is a capability or a position — the single most important distinction in the whole exercise.
- A capability is something the product does: it summarizes, it drafts, it classifies, it generates. Capabilities are made of competence. The model is coming for all of them, and the model provider ships them to everyone at once.
- A position is something the product occupies: trust the customer extends to you, data only you hold, a workflow you sit inside as the system of record, distribution you own to the buyer, accountability you carry when it goes wrong. Positions are not made of competence, so cheaper competence doesn't erode them.
Run your core product through five questions. Each "yes" to a capability question is exposure; each "yes" to a position question is insulation.
- Capability: If a frontier model added my core feature as a native button next month, is my primary value gone? If yes, you are selling a capability.
- Position — data: Do I hold data, generated by my own operation, that a competitor with the same model still could not reproduce? If yes, you have a data position.
- Position — workflow: Am I the system of record — does the customer's daily work, approvals, and history live inside me, so leaving means losing state? If yes, you have a workflow position.
- Position — distribution: Do I own the path to the customer such that the model provider would have to go through me to reach them? If yes, you have a distribution position.
- Position — accountability: When the output is wrong, does the customer hold me responsible — and pay me partly for carrying that risk? If yes, you have an accountability position.
The pricing pressure behind question 1 is not speculative. The per-unit cost of inference is falling roughly ten-fold a year, and that collapse is about to break every pricing model built on charging for the capability itself. If your revenue is a margin on top of a capability whose underlying cost is deflating an order of magnitude annually, that margin is being competed to zero on a schedule. Positions don't have that problem, because you're not charging for the tokens — you're charging for the trust, the data, the lock-in, or the liability.
Verdict rule: if you answered yes to question 1 and no to questions 2 through 5, you are a wrapper with a capability lead and no position. That is survivable — a capability lead buys you time and revenue — but only if you spend the lead buying a position before it closes. Do not mistake the lead for the moat. That confusion is how fast-growing companies walk confidently off a cliff.
Phase 3: The repositioning moves
Every exposed activity and every capability-only product needs to migrate toward a position. There are five moves. Pick based on which anchor you can most credibly build, not which sounds best.
Move 1 — Climb to the verification and judgment layer
Stop selling the generation; sell the certification that the generation is right. If you write code, become the team that guarantees it's correct, secure, and shippable. If you generate marketing, own the judgment of what's on-brand and what converts. The durable version is verification tied to consequences — the kind with no short machine-checkable certificate, where being wrong costs real money and someone has to answer for it. That is exactly the verification that stays scarce longest, for the reasons laid out in the case that verification is the new bottleneck. Concretely for Velya: the moat isn't the qualifying conversation, it's standing behind the qualified booking — carrying the accountability that the lead the clinic paid for is real.
Move 2 — Deepen proprietary data and relationships
Instrument your operation so that running it produces data no competitor can buy or scrape, then feed that data back into the product so it gets better the more you're used. Kommerce does this structurally: every delivery attempt, successful or failed, sharpens the trust score, and that outcome data doesn't exist in a public corpus anyone can prompt against. Ask: what does my business observe that no one outside it can see? Capture it deliberately. Relationships count too — the merchant who trusts my settlement, the clinic that trusts my handoff — because a model can generate a pitch but can't generate a relationship someone already has with you.
Move 3 — Own distribution the model can't intermediate
The model provider reaches capabilities to everyone simultaneously; it does not reach your specific customer unless it goes through the channel you own. Build the direct relationship, the community, the physical or trust-based last mile. In emerging markets this is concrete — the cash-on-delivery network, the local agent relationships, the settlement rails are distribution no foundation model intermediates. Before you lean on this, be ruthless about whether the "distribution moat" is real retention or an assumed one, using the same skepticism I'd apply to any network-effects claim.
Move 4 — Build the taste and curation layer
When generation is free and infinite, the product becomes the selection: which of the thousand generated options ships. Make your judgment the productized layer — opinionated defaults, curated outputs, a point of view the customer is buying because they can't specify it themselves. This is the taste-as-moat play made operational: you are not competing on whether you can generate, you are competing on whether you can choose, and choosing well is the scarce skill you're monetizing.
Move 5 — Turn AI into leverage on your own cost structure
The threat is also the tool. Every activity you scored high on commoditization is an activity you can now do near-free internally. Use it. Collapse your own generation costs, redeploy the freed capacity into the position-building moves above, and let AI do the drafting while your scarce human attention moves entirely to verification and judgment. The founder who only plays defense against AI loses to the one who uses it to fund the offense.
Prioritize the moves like this: for each exposed activity, pick the one move with the highest anchor-per-effort, and sequence data and distribution moves first — they compound and are hardest for a competitor to copy, whereas a verification or taste layer can be started fast but is easier to imitate.
Phase 4: What not to do
Four failure modes, each of which has killed companies that saw the wave coming and still drowned.
- Don't bet the company on the scaling trend continuing forever. More compute has bought reliable capability gains, but scaling is an empirical regularity, not a theory of intelligence — it can bend or plateau, and it makes no promise about which specific capability arrives next or when. Build for the direction (generation keeps getting cheaper) without staking survival on a specific capability landing on your roadmap's schedule. That distinction, between the robust direction and the fragile specific bet, is the difference between a strategy and a gamble.
- Don't confuse a capability lead for a moat. Being first on top of a new model is real revenue and zero defensibility. The provider can absorb the feature; a competitor can rebuild the wrapper over a weekend. A lead is time to build a position. It is not the position.
- Don't defend the commoditizing activity. Pouring effort into being 15% better at a task the model does for free is a losing race against a curve dropping ten-fold a year. Cede it, adopt the AI yourself, and move up the stack.
- Don't skip re-running the audit. The denominator moves. An activity that scores insulated today flips to exposed the week a model ships the capability it depended on being scarce. A one-time audit is a snapshot of a moving target.
The one-page worksheet
Run this now, then re-run it quarterly and after every major model release.
Exposure scoring — for each of your 8–15 activities:
- Name the activity concretely (a deliverable or a job, not a department).
- Commoditization C (1–5): how completely does a current model do this from a prompt?
- Scarcity anchor S (1–5): how much of the value rests on data, trust, distribution, accountability, or taste AI can't supply?
- Exposure = C − S. Flag everything ≥ +2.
- For each anchored activity, name the anchor in one sentence — if you can't, it isn't real.
Core-product verdict:
- If the model shipped your core feature as a native button next month, is your primary value gone? (Yes = capability, not position.)
- Which of the four positions do you hold — data, workflow, distribution, accountability? Name the evidence for each.
- Is your revenue a margin on a deflating capability cost? (Yes = repriced on a schedule.)
Action list — ranked:
- For each flagged activity, write the single repositioning move (1–5) with the best anchor-per-effort.
- Sequence data and distribution moves first; start the verification and taste layers in parallel.
- Pick your own top commoditizing activity and adopt AI internally this week to fund the rest.
Start Monday: decompose the business into its activities and score just the top five by revenue. If any of your three biggest revenue lines scores +2 or higher with no named anchor underneath it, that is the number to fix this quarter — before the market reprices it for you.