Five things moved. Each one changes a decision you are already making: what you pay, what you delegate, what you sign, what you protect.
Read the whole thing in three minutes. Where it touches your business, there is a tool that does the work for you on the spot. We talk about the market here, not about ourselves.
Between June 9 and July 9, all three frontier labs shipped new model generations. OpenAI's GPT-5.6 arrived in three tiers, the cheapest at one dollar per million input tokens. Anthropic's Sonnet 5 launched near the bottom of the market while matching much of what flagships did in spring. July's budget tier beats January's flagship on most business work.
Every AI workflow quoted, piloted, or contracted before June is now mispriced. Support, document processing, research, drafting: the unit economics moved by multiples, not percentages. A vendor quote from Q1 is built on costs that no longer exist.
Do not sign an annual AI contract priced against last quarter's models. Do not compare per-token prices without token efficiency: a cheaper model that burns twice the tokens per task is not cheaper.
The price of good-enough intelligence fell to commodity levels while the ceiling rose. That splits every AI budget into two problems: route the routine work to the cheapest tier that passes, and reserve premium models for the work that deserves them. Companies still paying flagship prices for budget-tier tasks are funding the gap out of margin.
This also explains the failure statistics everyone quotes. S&P Global found 42% of companies abandoned most of their AI initiatives in 2025, roughly double the year before. The failure is rarely the model. It is the absence of a baseline: nobody measured cost per completed task before or after, so nobody noticed when the economics moved.
Anthropic published usage data from 1.2 million sessions across 600,000 organizations. A third of the work is business operations: reports, checklists, reconciliations, spreadsheets. Software development is 8.7%. In July, both major labs put deliverable-producing assistants on ordinary subscriptions.
The unit of AI work moved from the answer to the deliverable. Recurring reports, reconciliations, and first drafts can be handed over with files attached and a definition of done: a subscription feature now, not a custom build. Non-technical staff are shipping internal tools too. Lovable reports a million new projects a week.
Do not run these like chatbots, typing questions at a tool built to finish work. Do not point one at the whole shared drive: scope what it reads and writes first. And do not measure adoption in logins. Measure completed units of work and error rate.
The quiet shift of 2026 is that the leading AI products stopped being places you type and became staff you brief. The working pattern is the same one you use with a good hire: here is the task, here are the files, here is what done looks like, here is what you may touch. Companies getting value write that down. Companies getting demos do not.
The usage data makes the point. Out of 1.2 million anonymized sessions, business process operations took 33.4%, content and copy 16.4%, and software development, the use case everyone assumed was the whole story, 8.7%. Your competitors' back offices are already doing this quietly.
The right first candidates share three traits: recurring, file-based, and low blast radius. The weekly ops report. Month-end reconciliation against invoices. Proposal first drafts. Meeting notes to action lists. Wrong first candidates: anything customer-facing, anything touching payments, anything you cannot check in minutes.
The internal-tools lane deserves the same discipline. Lovable hit $500 million in annualized revenue in June, mostly from non-technical users building CRMs, inventory systems, and HR tools in plain English. The rule that keeps it useful: internal and low-stakes first, and a professional security review before anything holds customer data or money.
OpenAI shipped GPT-Live on July 8: a voice model that listens while it talks, takes interruptions, and hands hard questions to a bigger model mid-call. Sierra raised $950 million at a $15 billion valuation in May, with over 40% of the Fortune 50 as customers and published deployments live in five to ten weeks. Vendor numbers, but named customers.
Deployment moved from quarters to weeks, and the awkward turn-taking pause, the thing that made callers hang up, is gone. The same capability sells at mid-market prices. After-hours calls, bookings, order status, tier-one questions: every call you miss is now a decision, not a limitation.
Do not pilot on your hardest calls. Angry escalations will fail first and poison the project. Route the status checks and bookings, with a clean handoff to a human. And say it is AI. Callers and regulators both punish the alternative.
Voice AI has been promised for a decade, and for most of it the tell was the pause: the half-second of silence while the machine waited for you to finish. Full-duplex models remove it. They listen and speak at the same time, acknowledge while you talk, and recover when you interrupt. That is the difference between a caller finishing the call and a caller asking for a person.
The enterprise evidence says the operational bar has been cleared at companies with far more brutal call volumes than yours. Resolution rates are now quoted like service-level metrics, and a five-to-ten week deployment at Fortune 50 scale means a focused mid-market rollout is a quarter's project, not a transformation program.
Before any vendor call, price your missed calls. That is what the tool on the right does. Then sort a month of answered calls into buckets: status checks, bookings, simple questions, account changes, everything else. The first two buckets are the pilot. The last bucket never is.
Salesforce agreed on June 15 to buy Fin, formerly Intercom, for about $3.6 billion, and now sells support by the resolved ticket. Its AI line hit $1.2 billion in annual recurring revenue, up 205%, billed by completed units of work. The Wall Street Journal reports roughly a quarter of McKinsey's fees are now tied to outcomes.
Both sides of your ledger. As a buyer, your renewals will arrive with usage or outcome components: the paperwork decides whether that cuts idle-seat cost or blows the budget. As an owner, AI-native competitors are entering services with structurally lower cost bases, and the same funds are the likeliest acquirers.
Do not sign an outcome-priced contract without a written definition of the billable event, a monthly cap, and your own baseline volumes. Do not renew a multi-year, hours-billed retainer at legacy rates this year: that locks in the pre-AI cost structure right as it deflates.
Salesforce currently runs four pricing models at once: per seat, credits, per conversation, per resolution. A market that has not settled its own billing unit is a market where a prepared buyer sets terms. The practitioners closest to these deals give the same warning: "resolution" is a contested metric. What counts as resolved gets defined in the contract or it gets defined in the dispute.
The rollup thesis is blunt: buy firms running at 5 to 10% operating margins, automate 30 to 70% of the workflows, and run them at 30 to 40%. General Catalyst has put about $1.5 billion into buying property managers, contact centers, and accountants to do exactly this. If you buy services, incumbent cost bases just became negotiable. If you own a services firm, the choice is to capture the margin yourself or compete against someone who did.
The insurance industry's standards body drafted three endorsements that strip generative-AI claims out of ordinary liability policies, and at least one major carrier sells an absolute AI exclusion on directors-and-officers coverage. Meanwhile Deloitte projects $40 billion in US fraud losses from synthetic media by 2027. The benchmark case: Arup wired $25.6 million after a video call where every other participant was generated.
Risk is being transferred to you quietly, at renewal, in endorsements nobody reads. And your finance team is the soft target: your executives' voices are on podcasts, your controller can move six figures alone, and crime policies often exclude wires you authorized, which is exactly what a deepfake makes you do.
Do not rubber-stamp this year's renewal. Do not let any payment or banking-detail change above a set threshold move on the strength of a call, video, or message alone. Ferrari's near miss is the lesson: one personal verification question, zero dollars lost.
The forms have numbers: CG 40 47, CG 40 48, and CG 35 08 exclude generative-AI claims from general liability, advertising injury, and completed operations. W. R. Berkley's absolute exclusion goes further, reaching claims arising out of AI use and AI-related disclosures, which can include what your investor deck says about your AI strategy. The same market is happy to sell the risk back as a specialty line. That is not a conspiracy. It is a repricing, and it only hurts the companies that do not read it.
In one large survey, 85% of organizations saw at least one deepfake incident inside twelve months; of those that lost money, most lost over $100,000. The targets are not naive people. The Arup employee was on a video call with what looked and sounded like the CFO and colleagues. The defense that works costs nothing: it is procedure, not technology.
The Briefing arrives weekly: what moved, what it changes, what not to do, and a tool for each. No product tours, no victory laps. If it stops earning the read, unsubscribe.
This is back issue No. 001. Every tool in it still runs. Read the current issue · Open the archive