AI-native tools removed the settings screen. The setup work reappeared below the waterline: in data access, permissions, evaluation, escalation, and attribution. The 2026 market data shows what that trade actually cost. Adoption is faster than SaaS ever was, retention at the cheap end is catastrophic, over 40% of agentic projects are forecast to die by 2027, and vendors bill in credits and conversations because effort is what they can measure without doing attribution work. Market figures in this piece are current to August 2026; our own numbers carry their exact dates. The durable work is connecting AI exposure to revenue; the tools themselves keep changing hands. We show our own chain below, zeros included.
The settings screen is gone, and the setup work is not. AI-native tools sold 2026's buyers on configuration-free adoption: describe the outcome you want and the software arranges itself. The market ran that experiment at full scale this year, and the results are legible in the churn tables, the project cancellation surveys, and the pricing pages. The configuration migrated below the waterline, into data access, permissions, evaluation, escalation, and attribution, where no demo call ever goes. We have written about single-metric misreads before: SEO Didn't Die. It Split in Two. and Citation Share Is Not Market Share each unpacked one number that gets sold as more than it is. This piece reads the whole tooling market the same way, then shows the one piece of setup work we believe outlasts every tool that promised to make setup obsolete.
The ease is real
Enterprise AI deals reach production at nearly twice the rate of traditional SaaS. Menlo Ventures' 2025 enterprise survey puts 47% of explored AI deals in production against 25% for conventional software, inside a $19 billion application-layer market. Product-led growth carries 27% of that AI spend against 7% for SaaS, and 76% of deployments are bought rather than built. Those figures are modeled from a survey, not audited from contracts, and Menlo says so.
Adoption friction collapsed because natural language replaced the configuration project. The unit of setup used to be the field mapping, the workflow builder, the admin console walkthrough. Now it is a described outcome: a team that once budgeted six weeks for onboarding types what it wants in a sentence and watches the tool comply in the demo. Every vendor built its funnel around that moment, and the funnel works. Money follows the absence of setup screens.
A settings screen was never decoration, though. It was the visible record of decisions the software needs made: who can see what, what happens on failure, what counts as done. Deleting the screen deleted the record, not the decisions.
Retention is the tell
The same tools that are easy to adopt are proving easy to abandon. ChartMogul's study of 200 AI-native companies found that above $250K ARR, the median AI-native product keeps 40% of its revenue year over year, with net revenue retention at 48%. Products priced under $50 a month retain 23%. Products above $250 a month hold 70% gross retention and 85% net. ChartMogul calls its higher-ARR buckets, roughly 50 companies each, "directional rather than statistically bulletproof," and that hedge belongs in every citation of the number, including this one.
Read the price tiers as a proxy for operational embedding and the curve explains itself. A $29 tool that wrote its own onboarding never earned a place in anyone's process, so cancelling it costs exactly what adopting it cost: nothing. The expensive tiers survive because somebody did the integration work the cheap tiers promised away: connected the data, defined the permissions, wrote the escalation path. The retention curve is tracing where the below-the-waterline work actually got done. The failure mode is visible at the vendor level too: Relay.app, a funded agent-automation product, announced in July 2026 that it is shutting down outright, with paid access ending September 14.
Where the configuration went
The work the settings screen used to represent moved into nine places: data access and normalization, knowledge and context curation, permissions and identity, evaluation datasets, failure thresholds, escalation rules, exception handling, auditability, and cost controls. None of them appears on a pricing page. All of them decide whether the tool survives contact with production.
The cancellation data maps onto that list with uncomfortable precision. Gartner projects that more than 40% of agentic AI projects will be cancelled by the end of 2027, citing costs, unclear business value, and inadequate risk controls, and coined "agent washing" for vendors rebranding chatbots and RPA as agents. S&P Global's 451 Research survey of 1,006 organizations found the share abandoning the majority of their AI initiatives jumped from 17% to 42% in one year, with 46% of projects scrapped between proof of concept and broad adoption. Across these studies the recurring causes are cost, data, integration, and unclear value. User experience appears on no root-cause list we could find. Nobody cancels these projects because the buttons are confusing; they cancel because the underwater work was never scoped, priced, or assigned to anyone.
The vendors' own operational numbers tell the same story from the inside. Among AI companies surveyed by ICONIQ in Q2 2026, 21% run proactive adversarial testing on their own agents, and almost half report their agents still need a human to step in on at least 30% of tasks. Escalation design, the least glamorous line item in the old configuration budget, turns out to be a live production dependency for half the industry. The settings screens are gone; the decisions they encoded are being made late, under incident pressure, by whoever is on call.
The pricing confession
Billing units are a confession about what a vendor can measure. Decagon, which sells AI customer-service agents, writes that the "vast majority" of its customers choose per-conversation pricing, because outcome-based pricing brings "constant renegotiations" over what counts as a successful outcome. Salesforce meters Agentforce in "agentic work units," 3.8 billion delivered cumulatively to date. AthenaHQ sells 3,600 credits a month for $295, one credit per AI response. Otterly tiers by monitored prompts. Conversations, work units, credits, prompts: all of these meter effort, and effort is what you can bill without doing attribution work inside the customer's business.
The units are also visibly in motion. Among AI companies in ICONIQ's two 2026 survey waves, consumption pricing rose from 35% to 42% in six months and outcome-based pricing from 18% to 23%, while 37% plan to change pricing again within 12 months. Growth Unhinged's 2026 monetization survey of 230+ companies finds 29% of B2B software companies already sell AI credits and another 33% say they plan to introduce them within a year. An industry that priced its predecessor generation in seats is repricing itself midflight, which is not what settled value measurement looks like.
The exception proves the rule's price. Intercom's Fin charges $29 per seat plus $0.99 per outcome, with "outcome" operationally defined: a resolution counts when no further help is requested after Fin's last answer. Getting to that one sentence took definitional and measurement work most vendors have not done. Outcome pricing is not a billing toggle; it is the migrated configuration, productized. Kyle Poyar's CAMP framework reaches the same conclusion from the analyst side: outcome pricing requires consistency, attribution, measurability, and predictability. Attribution is on the list because attribution is the work.
Conversions from AI: the answer-to-revenue chain
An answer-to-revenue chain connects four links: AI exposure, attributable visit, qualified lead, closed revenue. We coined the term because we could not find an existing one. Through August 2026 we could not locate a public controlled study showing that raising an AI mention share causes incremental revenue, nor a published closed-loop framework connecting AI exposure to closed deals. The chain is the honest response to that gap: instrument the whole path and read what it says, zeros included.
Here is ours, measured on our own site across three separate hosting layers, June 1 to August 3, 2026. Nineteen named AI crawlers made 85,194 requests. About 15,000 of those were live fetches by assistant user agents answering a real user in real time, not training crawls. AI crawlers out-requested traditional search-engine crawlers 2.6 to 1 over the window, and their volume grew 36% from June to July. All of that machine reading sent back 63 AI-referred visitors, 0.036% of our human traffic. Of 75 tracked lead events in the period, zero arrived carrying an AI referrer.
Our 0.036% share sits at the low end of published ranges, which is what a B2B services site should expect. The zero is a reading from one B2B services funnel, ours, over one nine-week window; it is not a claim about what the same chain would show on an ecommerce store or a clinic, which is why the chain gets built per business instead of quoted across them. Ahrefs measured 0.17% average AI referral share across 3,000 sites, and Conductor's benchmark, the largest published sample, puts AI referrals at 1.08% of all traffic.
The intent evidence is the reason to keep the instrumentation running. Ahrefs reports AI referrals are 0.5% of its own traffic but 12.1% of its signups, and Visibility Labs measured ChatGPT referrals converting at 1.81% against 1.39% for non-branded organic across 94 ecommerce stores in 2025, a 31% edge riding on volume roughly 70x smaller. The consistent pattern is tiny volume carrying materially higher intent.
Referral counts are also a floor, not a total. A buyer who reads an AI answer and later searches your brand or types your URL arrives with no referrer, so the chain's first links undercount the influence its later links eventually collect. That is exactly why the chain must run all the way to revenue: lead staging and disposition are where invisible influence becomes visible, or provably absent.
None of this came in a box. Referral classification, lead-event staging with dispositions, offline conversion upload with retraction of disqualified leads: every link is configuration that used to be a settings screen and is now engineering and process. Anyone selling AI visibility should be able to show you this chain on their own business, current to a stated date. Ours currently ends in a zero, and we would rather publish the zero than sell the crawl volume as success.
What survives the absorption
The tools themselves are features on a countdown. In one 32-day stretch (May 22 to June 23, 2026), Cisco closed its acquisition of Galileo, the agent-evaluation platform; Sitecore bought Scrunch, an AI-visibility vendor, for a reported $225 million; Salesforce agreed to acquire Fin for roughly $3.6 billion; and Superhuman bought GPTZero. Four control points of the AI stack moved inside platforms in a month. Salesforce also disclosed that more than half of Agentforce and Data 360 bookings come from existing customers; distribution is deciding who owns the category, and distribution favors incumbents.
For an agency, the planning assumption follows directly. Any standalone tool in your stack is a candidate for absorption, repricing, or shutdown on someone else's schedule. What sits on nobody else's schedule is the migrated configuration: your clients' data access, their permission boundaries, their evaluation criteria, their escalation rules, and their attribution chain. That work transfers intact across whichever tools survive.
We compete on that layer, and we publish our own chain to keep the claim checkable. If you want the same instrumentation pointed at your business, talk to us. Building and reading answer-to-revenue chains is what our SEO practice and AI marketing work does now.