What It Means to Be an AI Native Company And What It Means for IT
Airbnb didn't buy a few AI tools. It rebuilt how the company works, and the CEO now describes it as AI native. That distinction is the whole ballgame and it lands squarely on the desk of whoever runs IT.
"AI native" has already been flattened into a marketing phrase. Strip the marketing away and it means something specific and uncomfortable: the company's core work is now designed around models doing the first pass, with people directing, reviewing, and owning the result. Turn the models off and an AI-enabled company goes back to normal. Turn them off at an AI native company and the company stops.
That is not a hypothetical distinction. On its Q2 2026 earnings call in August, Airbnb CEO Brian Chesky described a business rebuilt from the ground up to be AI native — and put numbers behind it. Roughly 60% of Airbnb's new code is now written by AI. Nearly 45% of issues that start with its AI assistant are resolved without a human agent. Support cost per booking is down about 16% year over year. On key initiatives, concept-to-launch time has been cut by as much as 60%, and the company says it shipped nearly 80% more features than the year before.
Six months earlier, on CNBC, Chesky put the strategic case more bluntly: AI, he said, "is the best thing that ever happened to Airbnb," adding that "if you don't disrupt yourself, someone else will." You can argue with the framing. You can't argue that a $2.7-billion-a-quarter travel platform reorganized its engineering, support, and product functions around models and then reported faster shipping at lower cost per unit.
Most mid-market IT organizations are one to three years into multi-year carrier, UCaaS, and contact-center agreements that were sized for a human-first workflow. If AI resolves 30–45% of your front-line tickets in the next 18 months, you will be paying for seats, minutes, and licensed agents you no longer need and you will be short on the egress bandwidth and secure-access capacity you suddenly do need.
The window to reshape those agreements is at renewal, or at a mid-term amendment — not after the AI rollout has already changed your consumption profile.
What "AI native" actually means
Three postures get lumped together, and conflating them is why so many AI programs stall after the pilot.
AI-curious is a pilot: a few licenses, a champion, an internal chatbot, a slide deck about "exploring use cases." AI-enabled is a rollout: real tools deployed at scale, sitting on top of processes that were designed for humans and never redrawn. AI native is a rebuild: the process itself assumes a model does the first draft, and the human role is specification, judgment, and accountability.
| AI-Enabled | AI Native | |
|---|---|---|
| Unit of change | Tools added to existing workflow | Workflow redesigned around model output |
| Default first attempt | A person does it; AI helps if they remember | AI does it; a person directs and approves |
| Where capacity comes from | Headcount | Cycle time and throughput per person |
| What IT provisions | Licenses and endpoints | Identity, data access, network, and guardrails for agents |
| Primary metric | Tickets closed, projects delivered | Time-to-resolution and cost per unit of work |
| If you switched the models off | Slight slowdown | The operation stops |
AI-enabled is a procurement decision. AI native is an operating model decision and operating models are owned by leaders, not by tools.
The Airbnb signal: what it does and doesn't prove
Airbnb is a useful reference precisely because it is not an AI company. It is a marketplace with a support organization, an engineering org, a trust-and-safety function, and a finance team the same functional shapes most mid-market companies have, at a larger scale.
What Chesky described in August 2026 is a company that changed three things at once: how software gets built (AI writing the majority of new code, with humans specifying and reviewing), how customers get served (an AI assistant handling the front line in 50+ languages, escalating what it can't close), and how fast decisions turn into shipped product (concept-to-launch compressed by more than half on target initiatives). He also hired a CTO out of Meta's Llama team, which tells you the change was treated as an architecture problem, not a software-purchasing problem.
Two cautions, because vendor decks will strip them out. First, these are company-reported figures on an earnings call, not audited operational metrics. "60% of new code written by AI" says nothing about how much of that code survived review, and Airbnb has been explicit that it is not, in its words, "token maxing" for the sake of it. Second, Airbnb's economics are unusually forgiving: leadership noted that inference costs are minor relative to the value of a booking. If your gross margin per transaction is thin, the same architecture needs a much tighter cost model.
What the example does prove is the direction and the sequence. Airbnb did not get here by issuing licenses. It rebuilt the substrate, then the workflows, then the metrics. That order matters more than any single number.
What AI native means for the IT leader
For 25 years the IT leader's job has been, roughly: provide people with tools, keep them secure, keep them running. AI native breaks that framing in a specific way the "people" you are provisioning for are no longer only people.
In an AI native company, IT stops being the department that hands out capability and becomes the department that sets the ceiling on how fast everyone else can move.
Concretely, five parts of the job change:
- Identity gets a second population. Service accounts and agents need scoped, revocable, auditable credentials, not a shared API key in a config file. Non-human identity becomes a first-class part of your IAM program.
- Data access becomes the bottleneck, not storage. A model is only as useful as what it is permitted to see. Years of tolerated over-permissioning in file shares and SaaS become an immediate risk the moment a model can read everything a user can read.
- The network profile changes. Continuous outbound API traffic, larger egress, latency sensitivity on interactive tools, and new inspection requirements at the secure web gateway. Circuits and SASE policies sized for email and video conferencing are not sized for this.
- Audit becomes about actions, not just access. "Who logged in" is table stakes. "Which agent took which action, on whose behalf, with what data, and who approved it" is the new evidence standard your auditors, insurers, and enterprise customers will ask for.
- Shadow AI is already in your environment. Unlike shadow IT, it needs no budget and no install — a browser tab is enough. The practical response is a fast, generous approved path, because a restrictive one guarantees the unapproved one.
The uncomfortable part: the IT leader who treats this as a security-and-licensing exercise will get a compliant environment that no one uses, while the business quietly routes around them. The one who treats it as an operating-model exercise gets to set the pace of the company.
Five steps to make your IT team AI native
This is the sequence we walk clients through. It is deliberately ordered. Steps three through five fail if you skip step two.
Set an AI-first standard of work and define it precisely
"Use AI where it helps" produces nothing. An AI-first standard says: for a named list of repeatable tasks, the first attempt is made with AI, and if a person does it manually, that's a decision they can explain. Start with the tasks your team already does dozens of times a month — first-pass ticket triage, runbook and documentation drafts, log and config analysis, change summaries, vendor quote comparison, script writing.
Make it real by removing the friction: licenses for everyone on day one, not a request queue; a shared internal library of prompts and patterns that worked; fifteen minutes in your existing weekly staff meeting where someone demos what they automated. What gets demonstrated gets copied.
Measure: % of named tasks with an AI-first path · target 60%+ within two quartersFix the substrate before you scale: identity, data, network
This is the step everyone wants to skip, and the one that determines whether steps three through five produce value or incidents. Three workstreams, run in parallel:
- Identity: extend SSO and conditional access to every AI tool; create scoped, revocable identities for agents and service accounts; get non-human credentials out of scripts and into a vault.
- Data: classify what models may and may not touch, and run a permissions cleanup on the file shares and SaaS repositories the models will index. Over-permissioning that was survivable when only humans were browsing is not survivable when a model can retrieve.
- Network: re-baseline egress and API traffic, check inspection and DLP policy at the gateway for AI endpoints, and confirm your circuits and SASE licensing were not sized for a 2023 traffic mix.
Publish guardrails people can actually work inside
A one-page policy beats a thirty-page one nobody reads. It needs four things: an approved tool and model list with a documented path to add to it; a data classification rule stated in plain language (what never goes into a model, what goes only into approved enterprise tenants, what is fine anywhere); a human-in-the-loop rule naming which actions require review before execution; anything touching customers, money, production, or personal data; and logging so AI-assisted actions are reconstructable after the fact.
Write it so that the compliant path is the fastest path. Every hour of friction you add to the approved route is an hour of adoption you hand to the unapproved one.
Measure: time from tool request to approval · target < 5 business daysRebuild your five highest-volume workflows. Don't buy five more tools
This is the actual line between AI-enabled and AI native. Pull your ticket and request data, rank workflows by volume times cycle time, and take the top five. For most mid-market IT organizations that list is some combination of service-desk triage and password/access requests, employee onboarding and offboarding, change and incident documentation, asset and license reconciliation, and vendor or carrier quote comparison.
Redesign each one end to end on the assumption that the model does the first pass: draft the response, propose the classification, assemble the change record, generate the comparison. The human's job is to direct, correct, and approve. Then delete the steps the redesign made unnecessary. That deletion is the point; teams that add AI without removing steps get more work, not less.
Airbnb's version of this was cutting concept-to-launch by as much as 60% on key initiatives. Yours will be a first-response time that drops from hours to minutes, and an onboarding that takes twenty minutes of human attention instead of three hours.
Measure: cycle time per workflow · touches per ticket · cost per resolutionChange what you measure and how you staff
You will not get AI native behavior while you reward human throughput. Retire tickets-closed-per-tech as a performance measure and replace it with cycle time, first-contact resolution, cost per resolution, and time-to-deliver on projects. Report those to your executive team monthly so that the operating-model change is visible in business terms, not IT terms.
Staffing shifts too. The scarce roles become the ones that specify and verify: an automation and integration owner, a data-quality owner, and a security engineer who understands non-human identity. Rewrite job descriptions and interview loops accordingly: ask candidates to show you how they'd solve a problem with AI, not whether they've heard of it. And retrain the people you have first; the person who has run your service desk for six years knows which exceptions matter, and that knowledge is exactly what the model can't supply.
Measure: cost per resolution · time-to-deliver · capacity absorbed without added headcountIf five steps is too much to start, do this instead: put licenses in every hand this month, extend SSO over them, pick your single highest-volume workflow, rebuild it end to end, and publish one number to your executive team — the cycle time before and after. One proven workflow buys you the mandate for the other four.
The part nobody budgets for: the infrastructure bill
Every AI-native program eventually collides with the contracts underneath it. We see the same four collisions repeatedly:
- Bandwidth and egress. Continuous API traffic and larger payloads change the shape of your circuit utilization. Sites provisioned for a video-conferencing peak start showing latency on interactive AI tools.
- SASE and secure web gateway. New inspection, DLP, and CASB requirements for AI endpoints — often a licensing tier change, sometimes a platform change.
- Contact center and UCaaS seats. When AI resolves a third or more of front-line contacts, seat-based agreements signed for a human-first volume become the most expensive line item you own.
- Model and platform spend. Inference costs that look trivial in a pilot behave very differently at production volume, and they rarely sit in the IT budget where anyone is watching them.
None of these are reasons to slow down. They are reasons to sequence the commercial work alongside the technical work because the cheapest time to change a carrier, SASE, or contact-center agreement is before your consumption profile makes the incumbent's renewal quote look reasonable.
Frequently asked questions
The window is narrower than it looks
The gap between AI-enabled and AI native is not a technology gap. Every company on this list has access to the same models. The gap is willingness to redraw a process that currently works, in order to get one that works several times faster and that is a leadership decision, made by a small number of people, one of whom runs IT.
You don't need to rebuild your company this quarter. You do need to pick one workflow, rebuild it properly, prove the number, and use that number to buy the mandate for the rest. Meanwhile, look hard at what you've already signed. The infrastructure and contracts underneath your environment were sized for the way your company worked when the ink dried — and if the operating model above them is about to change, the cost structure below them should change with it.
Is your infrastructure ready for an AI native operating model?
We'll review your circuits, SASE posture, contact center and carrier agreements against where your AI roadmap is actually heading and show you what to fix before renewal, not after.