Asymmetric Adaptation: Seven Tests for an AI-Native Opportunity — and Why “They Haven’t Added a Chatbot” Isn’t One

Someone brings you an idea. The pitch is sharp. The market is large. The demo works. And then comes the competitive slide: the incumbent has not added a chatbot yet.

That line feels like an opening. It is usually a distraction.

A chatbot is a feature. An opportunity is a place where a company can respond to a worthwhile customer difference more cheaply than a competent incumbent can change its packages, pricing, suppliers and systems — and keep improving through each engagement. If your selection method cannot tell those two apart, you will back ordinary services wearing AI vocabulary, and you will miss the openings that matter.

The question is not “have they added AI?” It is “what would an otherwise competent incumbent have to change to match this experience?”

Asymmetric adaptation — keep the analogy honest

There is a cybersecurity picture that travels into business if you keep only the useful part. A static firewall encodes known patterns. A dynamic attacker probes, varies, and looks for the place the fixed rule cannot cheaply follow. In a business setting, the AI-native challenger is the party that can reconsider its approach for this customer situation; the software-centred incumbent is often the party that must make the situation fit an established package.

Keep that transferable insight — asymmetric adaptation — and drop the claim that defenders cannot adapt. Real defences already blend learned scoring with static rules. Cloudflare, for example, pairs machine-learning request scoring with managed signature rules and recommends using both together: the signatures catch known attacks with low false positives; the model catches variations that exact patterns miss.¹² The business parallel is the same: incumbents will adopt models. Your thesis must survive that — not depend on them remaining forever static.

So the competitive variable is not feature parity. It is the cost and time required to respond to a worthwhile difference.

Imagine an incumbent facing an unusual but valuable need:

“That is outside our package. We would need an exception, specialist involvement, different supplier arrangements and a system change.”

The challenger’s intended response is:

“Let us understand the objective, inspect what we know and have available, and assemble an appropriate delivery.”

The attacker need not reproduce the incumbent’s entire feature catalogue. It competes precisely where the standard package fits badly.

The incumbent does not have to be stupid

This is the part most “AI disruption” pitches skip.

A competent incumbent can still have pricing, organisation, suppliers and technology optimised for repeatable packages. Serving greater variation may undermine the economics it has built. The opening is structural, not moral. You are not betting that their product team is asleep. You are asking what they would have to unwind — packages, price books, supplier contracts, system assumptions — to match an experience that adapts per customer.

“They haven’t added a chatbot” answers none of that. Adding a chat window leaves the package economics untouched. Seema Amble’s September 2026 a16z essay makes the same warning from the other direction: incumbents are already moving past “slap on a chatbot” toward agents that take action inside systems of record, while general-purpose agents can sit as a front door. A specialised company still has to do a cross-system job better — through context, tools, evaluation and a learning loop — or the window closes.³

Hypothetical worked market (labelled as such). Take a packaged corporate-events platform: published formats, fixed tiers, supplier catalogue, registration workflow. A customer wants a small gathering built around a specific relationship objective — beginners included without embarrassment, sponsors contributing introductions rather than logos, follow-through that continues after the day. Matching that experience is not a chatbot ticket. It is a change to what the company sells, how it prices exceptions, which suppliers it can compose, and whether its systems treat variation as a bug or as the work. That is the competent-incumbent counterfactual. If the answer is “they just need an LLM in the help desk,” you do not have an opening.

Where to hunt: suppressed differences

The territory is not “every market AI can touch.” It is narrower and more useful:

Look for markets where valuable differences between customers are being suppressed by standardised offerings, expensive expert attention or rigid operating systems.

Why those three? Standard packages force dissimilar customers into the same answer. Expensive experts make fitted attention a luxury good. Rigid operating systems — including the software-in-the-middle pattern I have argued elsewhere, where permanent machinery freezes yesterday’s assumptions into tomorrow’s exceptions — make worthwhile variation expensive to serve. One paragraph is enough here: the architecture explains why incumbents stay package-shaped; this piece is about where that creates an attackable opening.

Five shapes sit in that territory. They are not predetermined products. They are places to investigate:

The attractive zone is not unlimited uniqueness. It is:

Different enough that understanding and adaptation create value; similar enough that knowledge and delivery capabilities transfer.

That is how a company delivers distinctive experiences without becoming an endless series of unrelated consulting jobs. And “experience” is not limited to leisure. In a business setting, a remarkable experience might be less uncertainty, fewer handoffs, better preparedness, a resolved problem, or confidence that someone is holding the whole responsibility. As I have argued elsewhere, experience is the unit of value — not code, features or completed agent tasks. This piece takes that as given and asks a different question: where does that unit create a company worth selecting?

Two routes into formation

Market-first: find the unmet experience, then assemble the assets — expertise, relationships, authorised access, reusable judgement — required to deliver it.

Asset-first: inspect the knowledge, capabilities and authorised relationships you already hold, then ask which valuable experience this particular combination makes possible. A company can emerge from a domain expert, a warm customer route, reusable IP and an underserved outcome before anyone writes a product specification.

The guardrail is the same either way: the network supplies reachable possibilities; customer evidence decides which deserve a company. Asset-first without that discipline recreates the original mistake at another level — forcing a business nobody wants because the assets were convenient.

My co-founder’s line lands here with more precision than it first appears: companies, not products. The target is not another application. It is a company whose ability to understand, deliver and adapt is materially better than the alternatives.

What “AI-native” means out there — and why that is not enough

“AI-native” is being used to describe several different things. Some mean AI-centred software. Some mean an organisation run by people and agents. Others mean a business that delivers the work customers used to buy as a service. Those are related. They are not interchangeable. Each source below is useful for one job in the argument — this is not a literature review.

a16z, several strands. In 2023, Alex Immerman and Jamie Sullivan argued that AI-native consumer services could provide individually tailored expertise and adapt over time — expanding access, willingness to pay and retention, not merely cutting the cost of an existing digital service.⁴ In February 2025, Kimberly Tan described building an AI-native full-stack firm that competes with existing service providers on speed, cost and quality, embedding AI and human judgement in the operation — and noted that some industries prefer buying outcomes rather than products, so delivering the service with AI inside can be the right commercial shape.⁵ Anish Acharya’s January 2026 notes take a different emphasis: as coding gets cheaper, every team should become a software team, and strong AI applications combine models, specialised interfaces and a large feature surface that is now inexpensive to build — with the hard problem shifting from how to build to what to build.⁶ That software-first strand is real. It is not the opportunity shape this piece favours. Amble’s September 2026 essay, already cited, supplies the learning-system and counterfactual discipline.

Sequoia. March 2026’s Services: The New Software draws the commercial line cleanly: “A copilot sells the tool. An autopilot sells the work.”⁷ If you sell the tool, you race the model. If you sell the work, model improvement can make the service faster and cheaper. Useful for: who owns completion?

Bessemer. AI-native services thinking across the venture literature converges on a similar commercial claim: the customer can buy a finished result delivered through software and people, and absence of people is not by itself evidence of superior economics. A live Bessemer Atlas piece on durable monetisation sharpens the pricing test: durable AI pricing ties revenue to something a customer can independently verify, not to a claim — soft ROI that leaves the outcome in the customer’s hands compresses willingness to pay compared with closing the loop end-to-end.⁸ Useful for: verified outcomes and surplus, not staffing cosplay.

IBM. Cole Stryker’s February 2026 explainer treats AI-native as foundational — AI as a core component rather than a bolt-on — and applies it to startups and to established firms that comprehensively re-orient. The remove-the-AI test is explicit: if the AI were removed, the product would not merely degrade; it would cease to be useful at all.⁹ Useful for: structural dependence. Birth date is not the test.

Microsoft. The 2025 Work Trend Index describes the Frontier Firm: an organisation built around on-demand intelligence, human–agent teams, and a progression from assistance toward agents running processes under human direction.¹⁰ Useful for: examine the whole operating model, not only whether the customer-facing UI contains AI.

Consulting framings of the “agentic organisation.” Parallel public writing describes redesign across business model, operating model, governance, workforce and technology — future-back rather than bolting agents onto today’s departments. Useful as a shape-statement only here: start from the future service and business, then determine the production system. Do not recreate the org chart as a cast of agents.

Every / Dan Shipper. Agent-native architecture treats features as outcomes an agent pursues with tools, composed through instructions — including requests developers did not explicitly anticipate. The diagnostic question is exact: “To change behavior, do you edit prompts or refactor code?”¹¹ Useful for: adaptable delivery mechanism. Not, on its own, a definition of the company’s customer promise or economics.

Definition I would keep, short:

An AI-native company is designed around AI as a core production capability, materially changing what it can deliver, how it operates, or the economics of serving its customers.

Preferred opportunity shape, stated separately: companies that turn abundant intelligence and proprietary knowledge into better customer outcomes — adapting delivery to the situation, retaining useful learning, and increasing capability without proportionally increasing cost, complexity or founder dependence.

Seven tests

These are evaluation additions, not a rival brand of investment thesis. They give existing criteria — changed economics, customer desire, robustness as models improve, compounding assets — more precise meaning. A twelve-point dump of related ideas tends not to land; people argue with each bullet and the combination gets lost. What follows is one route through the selection argument.

Two contrasting ideas, labelled hypothetical. Pass candidate: an accountable relationship-stewardship service for mid-market firms that already outsource event and introduction coordination — fitted per client objective, learning transfers across engagements, priced on completed bounded engagements with verified follow-through. Fail candidate: “CRM with a chatbot the incumbent lacks” — same package, AI vocabulary on the interface, memory of tickets, no change to what the customer can buy.

1. Remove the AI from the complete offer

IBM’s remove-the-AI test is strong for products. For a company, apply it to the whole customer promise: price, turnaround, specificity, coverage, availability and quality — not merely whether a human could theoretically perform the task.

Ask: Without AI, could we still deliver this offer to these customers on these terms?

A human might perform the work. That does not establish that the same commercial offer remains viable. This distinguishes “we use AI to draft our emails” from “our service only makes economic sense because we can bring this much intelligence to each engagement.”

2. Is learning both possible and valuable?

Amble separates remembering from learning, and profession knowledge from institution knowledge. Memory recalls preferences and last week’s decision. It does not tell the system whether the work was good, why an expert changed it, or what to change next time. Profession knowledge transfers across customers; institution knowledge often stays local.³

Ask: What will each engagement teach us, and how will that improve a different engagement?

That is stronger than any generic “data flywheel” claim. Remembering is not learning. The preferred territory is enough variation for judgement to matter, and enough shared structure for lessons to transfer — for example, “for groups with many beginners where relationship-building matters more than competition, a less competitive format may produce a better experience” travels; “repeat Client A’s day for Client B” does not.

3. Price the cost of a verified outcome — not the cost of generating it

Model tokens are not the delivery cost. Account for preparation, review, corrections, exceptions and the real-world work required to complete delivery. Bessemer’s durable-monetisation framing is the right commercial instinct: tie revenue to something the customer can independently verify, not to an unverifiable assistance claim.⁸

An imagined excellent testimonial is a design target. The actual response is the evidence.

4. Creating value versus keeping enough of it

State the economic mechanism explicitly:

Does the opportunity make an existing service cheaper, make a much better service affordable, or create demand for something that was not previously viable?

Then ask why the provider can earn an attractive margin. A personalised experience might justify a premium. A dramatically cheaper service might unlock many more customers. Those are different businesses and require different evidence. If AI collapses price and you cannot keep surplus — because buyers capture it, or because undifferentiated competitors race to zero — you have created value for the world and a weak company for yourself.

5. An easy first commercial transaction

Sequoia’s wedge is practical: already-outsourced work. The buyer has accepted external delivery, a budget exists, and they are already purchasing an outcome. Replacing an outsourcing contract can be a vendor swap; replacing headcount is a reorg.⁷

The large eventual opportunity can be ambitious. The first thing a customer buys should be understandable. This is not a reason to restrict ideas to existing categories. It is a way to ask whether the first sale can establish evidence without requiring the buyer to embrace the entire future vision.

6. The two counterfactuals

Tan warns that incumbents have distribution and can eventually incorporate the technology; the startup window is not guaranteed to remain open.⁵ Amble adds the general-agent front door. So test both:

1. The incumbent gets equally capable AI. Why do customers still choose us?
2. The customer has a capable general-purpose agent. Why do they still need our company?

Plausible answers involve trusted relationships, access to scarce resources, superior accumulated judgement, cross-party coordination, or responsibility for getting the job completed. These are hypotheses to prove — not benefits conferred by calling the company AI-native.

7. Separate the unit of value from commitment, pricing and proof

As I have argued elsewhere, every opportunity needs four clear answers that must not be collapsed into one another: what the customer values; what the company commits to delivering; what the customer pays for; and what proves the company is improving. Experience can be the unit of value without becoming vague. Completed work, margins and repeat revenue can be measured without replacing customer value as the score.

I will not rebuild that table here — it is owned by the experience-design argument. The selection test is simply whether the idea has four distinct answers, or whether “experience,” “engagement,” “subscription” and “tasks completed” are being used as synonyms to hide a hole.

Check your own instruments

Even a good brief fails if the scoring curve quietly selects another shape.

Curves built around ARR and time-to-revenue encourage non-subscription businesses to present themselves as subscriptions — or undervalue attractive engagement-based businesses. For experience-led companies, evaluate revenue quality, contribution margin, repeat purchasing, cost-to-serve and dependence on scarce human judgement. Keep genuine recurring revenue distinct from projected annual sales, and gross booking value distinct from revenue the company retains.

A company should not need to pretend it is SaaS to look investable.

One small generation heuristic from our own trials: changing a single assumption about incumbents — treating them as running on ordinary tools rather than as already agent-native — produced a visibly different set of opportunities overnight. Useful as a provocation for idea generation. Not a substitute for the competent-incumbent counterfactual above. The opportunities still have to survive Tests 1–7 when the incumbent is competent and gets the same models.

Proof order

First, show that the new orientation generates different, commercially coherent opportunities — assessed independently, not rewarded for vocabulary such as “adaptive,” “personalised” and “compounding.”

Then deliver a bounded engagement in the world. Capture actual customer response, delivery costs, exceptions and useful knowledge.

Then run a second engagement with a meaningful variation — different objective, stakeholder mix or constraint — and ask:

Did the company adapt using its accumulated capability — or did the founder rebuild the service by hand?

That is a stronger test of the thesis than “the agents completed their tasks.” Track customer value, the cost of worthwhile adaptation, and demonstrated reuse together. Maximising only one can hide failure in the others.

What changes in the search

I would express the expanded doctrine as:

We look for valuable experiences that existing business models cannot deliver economically or adapt to readily. We build companies that can understand the customer, assemble the delivery and retain what they learn. Software is infrastructure or working material; the commercial asset is an increasingly capable business.

That does not abandon products. A bounded, buyable promise can still be a product even when every instance is fitted differently. It does not abandon category tests either. It refuses to confuse category membership with an opportunity worth funding.

Capital should still follow evidence: customer behaviour, willingness to pay, repeat use and economics. The change is what you ask the opportunity engine to discover, and what you count as proof.

Not a clever application that can be built quickly.

Not an agent team that performs lots of tasks.

Not a chatbot the incumbent has not shipped yet.

But a company that can make it materially better to be its customer — and can keep improving that experience without rebuilding itself every time — even when the incumbent is competent, and even when everyone has the same models.

If you can do that without me in the room, the piece has done its job. If you still need the chatbot slide to feel like an opening, it hasn’t.