1. Freed capacity does not become P&L impact on its own.

In our Q3 edition, we wrote that the AI conversation had shifted from capability to results. This quarter showed us where results break down. The chain from tool to profit and loss (P&L) impact has four links: (i) a tool is deployed; (ii) a task gets faster; (iii) the freed capacity is reallocated; and (iv) the reallocation appears in financial results. In our observation, most organizations stall at the third link, and by that point, the technology has done its job. What stalls is the management follow-through. AI can create capacity, but the authority to change a quota, a panel size or a staffing plan sits with the management team.

Two examples from our portfolio illustrate the point. At one healthcare organization, clinical staff completed documentation roughly twice as fast after an AI deployment, yet throughput targets never changed. At another, clinicians gained an hour a day, but the organization did not have a concrete plan to measure where that hour went. In both cases, the technology delivered, but the operating model did not adjust.⁽³⁾

What to do about it: Start by asking one question. Now that the work is faster, which quota, staffing plan or budget has changed as a result? If the answer is none, select a forcing function before year-end. We see four options that can help convert freed capacity into a financial number: (i) establish a no-backfill default on attrition; (ii) set a flat-headcount budget year; (iii) develop an explicit plan — co-owned with revenue leadership — to grow into the new capacity; or (iv) redeploy the freed capacity to an opportunity that previously went unstaffed. We believe the costly choice is not choosing, because unclaimed capacity quietly disappears back into the workday.

2. Always-on agents need enterprise-grade controls.

Managing AI to a return assumes you know what your AI is doing, and that assumption is getting harder to defend. Last quarter, we described agents moving from assistants to actors, and since then, persistent agents (agents that run continuously rather than per request) have become a standard feature of platforms that many companies already license. Microsoft, Google and Meta all shipped agent capabilities this summer, either generally available or as recurring-task agent. As adoption increases, incidents are rising. In one vendor-commissioned industry survey, 65% of organizations reported at least one agent-related security incident, and 82% discovered agents running that they did not know existed.⁽⁴⁾ According to a survey by 1Password, a credential-management vendor, roughly 40% of developers grant agents persistent access to systems or credentials.⁽⁵⁾ The frontier labs are documenting the risk themselves. In September, OpenAI published a report detailing six incidents in which its own models acted outside their controls, from inserting instructions into their own summaries to uploading files to public services. OpenAI notes that all six were observed during training or evaluation rather than live use.⁽⁶⁾

The market has responded with a control layer built on three elements: per-agent identity with expiring credentials, inline policy enforcement and rapid revocation. Approval prompts alone are not a control. When every action generates a prompt, users simply fall into the habit of clicking “yes.” We examined the security dimension of this shift in From Detection and Response to Zero Trust.

What to do about it: Treat agents as a class of identity and ensure your team can answer four questions at any time. (i) What did our agents do in the last 24 hours? (ii) How many do we run, and who owns each one? (iii) Do they hold their own credentials or shared keys? (iv) What shuts one off, and how fast? A leadership team that cannot answer these questions does not yet control its agents; it hosts them.

3. The next buyer of your product may be an agent.

A quarter ago, agent-mediated purchasing was largely speculative. Now, it runs on production-grade infrastructure. One infrastructure software company attributes roughly 20% of its revenue to deployments initiated by AI agents rather than by people, only two quarters after the pattern first appeared.⁽³⁾ AWS made agent payments generally available, allowing agents to discover and pay for APIs and tools autonomously within spending limits set at the infrastructure level. In September, Meta launched Muse, a consumer agent with its own browser that can operate existing websites and pay with single-use cards.⁽⁷⁾ A growing share of consumer shopping journeys now begins in an AI assistant rather than a search engine or storefront.

For anyone selling a product, the implication is direct. An agent selects vendors by parsing structured data on products, pricing and terms, and it is indifferent to the brand and design investments that win over human buyers. Companies that sell well through polished pages will need to sell equally well through machine-readable ones.

What to do about it: Make product and pricing information structured and machine-readable. Decide deliberately whether agents transact inside the chat or are routed to your own properties. And we recommend beginning the pricing analysis now, before the “user” behind a seat turns out to be another company’s agent.

4. AI economics reward disciplined buyers.

The price companies pay for AI continued to fall in Q3. One leading lab canceled a scheduled 50% price increase in August, and the effective price paid per million tokens has fallen roughly 40% from its March peak. Pricing at the frontier is more mixed, with the newest flagship model launching in September at roughly 2.5 times its predecessor’s price.⁽⁸⁾ Lower unit prices, however, have not produced lower bills because usage is growing faster than prices are falling. Token usage grew roughly 4.5 times in 2025 while prices halved,⁽⁹⁾ and a single agent task can consume 10 to 50 times the tokens of a chatbot query.⁽¹⁰⁾

Leaders are feeling this strain. In KPMG’s Q2 pulse survey, 49% of leaders reported scaling back, narrowing, delaying or pausing agent deployments when costs outweighed the expected value, and only 7% reported established ROI to date.⁽¹¹⁾ Some of that spend is simply unused. Correlation One, an AI training provider, reports that its enterprise clients typically carry two to three times more AI licenses than trained, active users.⁽¹²⁾

Measurable results often follow management discipline. In the same KPMG survey, leaders who maintained visibility into their AI operating costs were five times more likely to report established ROI than those without that visibility.⁽¹¹⁾ In practice, that visibility means a live cost dashboard and a cost review in the AI approval process, disciplines that roughly half of surveyed organizations now report having.⁽¹¹⁾ Three levers are inexpensive and well understood: default users to lower-cost models and route up as needed; use prompt caching; and negotiate utilization-based licensing where it is offered. A fourth lever, open-weight models, is maturing quickly. Eight of the ten highest-volume models on OpenRouter, a leading model-routing platform, are now open-weight models. More telling, leading open models now deliver near-equivalent performance, based on standard benchmarks, and do so at roughly 60% of the price.⁽¹³⁾ For stable, high-volume tasks, running an open model on owned infrastructure is becoming a credible alternative to per-token pricing.

Two shifts warrant particular attention in 2027 budgets. First, we expect agent costs to enter headcount planning, as teams begin to weigh an agent’s cost per completed task against the fully loaded cost of a hire. Second, seat-based pricing is giving way to volume- and credit-based models, as GitHub did with Copilot in June when multi-hour agent sessions made flat pricing unsustainable.⁽¹⁰⁾

What to do about it: Before renewing any AI contract this budget season, run an audit of seat utilization against actual usage, as well as an inventory of personal AI accounts in use with company data. Cost control is not the end goal. It serves as evidence that the organization is managing AI to a return.

5. AI judgment is becoming the scarce skill.

In our Q3 edition, we asked which capabilities stay human. We are now beginning to see the answer in the labor market, as employers are paying a premium for AI skills while also placing greater weight on human judgment. According to PwC’s 2026 Global AI Jobs Barometer, job postings requiring AI skills grew roughly eight times as fast as the overall jobs market (69% versus 9% for postings overall), and the average wage premium for workers with AI skills has climbed to 62%, up from 57% a year earlier. The demand is not confined to technology teams. Indeed Hiring Lab finds that 63% of U.S. job titles that include AI now sit outside technology occupations.⁽¹⁴⁾ At the same time, AI has weakened hiring’s traditional signals. As more candidates use AI to polish their applications, the application itself may reveal less about a candidate’s true qualifications.⁽¹⁵⁾ Leading companies are responding by rebuilding interviews to test AI judgment directly, including exercises built around deliberately flawed AI output in which the candidate’s job is to catch the mistake.

What is happening to headcount may be more nuanced than headlines suggest. Layoffs attributed to AI coexist with rehiring, and firms more exposed to AI have grown headcount faster than those less exposed. Team composition is changing faster than team size, with junior hiring shifting sharply toward roles that demand judgment, leadership and other skills once reserved for senior staff.⁽¹⁴⁾

What to do about it: Make AI fluency a stated expectation for every role and give the organization a training runway before holding people to it. When a role opens, backfill the capability the team needs now rather than the role as it existed. And update interview processes before the January hiring cycle, because a hiring bar that cannot see AI judgment will fill 2027 with 2024 skills.

‍