Business Durability in the Age of AI

Written by Sean Linehan. Published on Feb 23, 2026.
There's no shortage of wild hypothesizing about what the world looks like on the other side of AI. Every week brings a new prediction about which industries will be obliterated, which jobs will disappear, and which trillion-dollar markets will materialize out of thin air. Most of this speculation focuses on what will change.
But Jeff Bezos once made an observation that I think is the better starting point. "I very frequently get the question: 'What's going to change in the next 10 years?' I almost never get the question: 'What's not going to change in the next 10 years?' And I submit to you that that second question is actually the more important of the two."
His answer was deceptively simple. Customers will always want lower prices, bigger selection, and faster delivery. You can't picture a future where someone says "I wish shipping were slower and more expensive." So Amazon built everything around those invariants.
That same logic applies to understanding which businesses are durable in a world where artificial intelligence is rapidly compressing the cost of cognitive work. Instead of asking "what does AI disrupt?" we should be asking what types of complexity are permanent, and which businesses are anchored to them.

What's In Peril

Selling Intelligence by the Hour

Professional services businesses that are fundamentally "body shops," selling organized human intelligence at a billing rate, face a structural challenge. There are two sub-categories here, and neither is particularly enviable.
The first is the AI-resistant body shop. Think Compass Group, the international food services company. They're selling cafeteria staffing. AI isn't replacing the person who serves you lunch anytime soon. But these businesses were never high-margin or high-growth to begin with. Being AI-resistant isn't particularly valuable if the underlying business isn't great.
The second is the intelligence-as-a-service body shop. Accenture, BCG, the big accounting firms. The work they do, synthesizing information, producing analysis, advising on decisions, is precisely the kind of cognitive labor that AI is getting very good at, very fast. These firms won't vanish overnight. They have brand equity, institutional relationships, and deep organizational knowledge. But the headwinds are real, and it's not obvious that being large is an advantage in the transition. A smaller, more agile competitor, or an AI tool itself, could be just as credible.

The Software Squeeze

Software businesses aren't necessarily doomed, but they face a subtle squeeze. The price at which you sell software now needs to remain meaningfully less than the combined cost of a customer (a) thinking through their own needs and (b) paying the token price to have AI replicate your functionality.
As token costs plummet, that second number keeps shrinking. The software that survives is software where there's deep muscle memory, regulatory entanglement, rapidly changing environments you wouldn't want to track yourself, or enough accumulated domain expertise baked into the product that reimplementing it from scratch is harder than it looks.
The software that's most vulnerable does a discrete, well-defined task you could describe to an AI agent in a paragraph. Summarize this document. Clean up this spreadsheet. Generate this report. If the job-to-be-done can be expressed as a prompt, the software is competing with a prompt. That's a losing position long-term.
There's also an important distinction within enterprise software. Products like Workday and ADP have historically been protected by massive integration moats. Ripping them out is an 18-month, multi-million-dollar project. But if AI compresses the cost of building and maintaining software integrations by an order of magnitude, those moats erode. Rippling demonstrated, even in the pre-AI era, that a sufficiently motivated competitor could simply go and do the integration lift that everyone assumed was prohibitive. That effort only gets cheaper from here. Integration moats are on a decay curve, and AI is accelerating the rate of decay.

What Endures

So what endures? I see five categories of businesses that are anchored to permanent sources of complexity.
Scarce physical assets and access rights. AI doesn't change the value of owning a railroad, a pipeline, a spectrum license, a landfill, or an airport slot. These are toll booths on physical reality. AI might make the operations of these businesses more efficient, which is great for the owner, but it doesn't compete with the asset itself. Think Brookfield, the Class I railroads, cell tower companies, utilities. The reason to own these has nothing to do with AI and everything to do with scarcity and replacement cost.
Regulated gatekeepers. Businesses where the government has mandated that a transaction must flow through an approved entity. Exchanges, clearinghouses, credit rating agencies, certain testing and certification labs. AI can't disintermediate you if the law says you're required.
Atoms, not bits (with bits as a tailwind). Industrial businesses that sell highly engineered physical products into fragmented niche markets with deep application expertise. Think Danaher or IDEX. AI probably makes their R&D faster and their sales process more efficient, but you can't download a replacement for a mass spectrometer or a specialized pump. The switching costs are embedded in the customer's process validation and workflow.
Brands that represent trust in high-stakes decisions. Many consumer brands are vulnerable if AI agents start making purchasing decisions. But in domains where the stakes of a bad choice are very high and quality is hard to verify in advance, a trusted brand still does real work. Medical devices, infant products, professional liability insurance. The key test is whether the trust was earned through repeated verified performance (durable) or is just familiar from advertising (vulnerable).
Orchestration businesses. This is the category I find most interesting, and probably the most counterintuitive. It deserves its own section below.

What Orchestration Really Means

There's a type of business that might be classified under professional services, but it's fundamentally different from a body shop. The product isn't labor hours or software. The product is that they've already done the painful, slow work of assembling a network of relationships, integrations, and processes so that you don't have to.
Benefits administration is a perfect example. A company like this bundles together dozens of value streams.
  • Pre-negotiated vendor relationships with insurance carriers
  • Pre-built enrollment processes
  • Customer support infrastructure for employee questions
  • Regulatory compliance across jurisdictions
  • Access to a menu of benefits that would take years to assemble independently
Some of that is software. Some of it is people. But the real product is the assembled position, the fact that they've already done the painful, slow work of stitching together all the relationships and processes so you don't have to.
Freight forwarding is another pristine example. From the customer's perspective, whether the forwarder runs on software or manual processes doesn't matter that much. What the customer is buying is smooth-running shipments at good prices. To deliver that, forwarders need:
  • Pre-negotiated rates with carriers based on volume commitments
  • Relationships with customs brokers at specific ports
  • Knowledge of which carriers are reliable on which lanes
  • The ability to get on the phone at 2am when a container is stuck and know exactly who to call
Flexport learned this early, which I saw firsthand. In principle, software can automate a lot of the work and create delightful, real-time experiences for customers. But in order to deliver on that promise, you have to brick-by-brick build the network of relationships, which requires sufficient scale, internal expertise, physical presence, and years of earned trust.
Other businesses in this cluster include ADP's payroll processing (they have pre-negotiated tax filing relationships with thousands of state and local jurisdictions, many of which don't have APIs and instead have a person with a specific process), pharmacy benefit managers, insurance brokerages, group purchasing organizations in healthcare, and franchise systems.
What makes these businesses durable, and what actually makes them better in an AI world, is that AI compresses the cost of their delivery without eroding their position. They can use AI to automate customer support, speed up enrollment, handle bureaucratic processing. Their margins expand because the cost of doing the work drops, but the switching costs and network position remain.

Integration Moats vs. Relationship Moats

The critical distinction, particularly within orchestration businesses, is the difference between integration moats and relationship moats.
An integration moat says "switching is expensive because reconnecting all the technical plumbing takes forever." This moat is eroding as AI compresses the cost of building and maintaining software. If the bottleneck to switching was an 18-month, $2M reintegration project, and AI turns that into a 3-month, $200K project, the effective switching cost has dropped by 90%.
A relationship moat says "switching is expensive because you'd need to reform dozens of bilateral relationships that exist in the physical, contractual, and regulatory world." AI doesn't help with this. The State of Ohio's tax office doesn't have an API. They have a person who processes things in a specific way, and ADP has a 20-year relationship with that office. A benefits administrator hasn't just integrated with 40 carriers. They've negotiated rates with 40 carriers based on their book of business, and those rates are a function of their scale. That's a demand aggregation moat.
The practical test for any specific business is this. If a well-funded competitor could build all of your software from scratch in 6 months using AI, what would they still be missing? If the answer is "not much," you're a software company with integration lock-in and you're on borrowed time. If the answer is "they'd still need 5 years to build the carrier relationships, the regulatory approvals, the negotiated rate cards, the institutional trust," you're looking at something much more durable.

The Bezos Test for Durability

This brings us full circle. The Bezos question, what won't change, gives us the filter.
Some complexity is permanent.
  • Regulatory fragmentation across jurisdictions
  • Multi-party contractual negotiation
  • The physical movement of atoms across borders
These will always be complex. Other complexity is temporary.
  • Connecting two software systems
  • Building an integration
  • Generating an analysis
These are only complex because the tools haven't caught up yet. The clock is ticking on that kind of complexity, and AI is speeding it up.
The businesses worth owning are anchored to the permanent kind. Their moats live in the relational and contractual layer, not the technical layer. AI will compress the value of organized human cognition while doing essentially nothing to the value of assembled network positions, negotiated relationships, regulatory standing, and physical scarcity. The businesses on the right side of that line will see their margins expand and their positions harden as delivery costs fall and competitive moats stay exactly where they are.