Newsletter, 13 November 2026

Agents are becoming customers, and what BCG says about the gap to revenue

This week: a16z on the business-to-agent economy, BCG's data on why so few companies have controls for agents, UK adults now spending over a billion pounds a year on AI subscriptions, and how Anthropic's agents learn from their own past sessions.

Our four views on how AI is changing iGaming are on the blog: Acquisition is changing, A new front door, How products get built and businesses run and Regulation. This week's news sits underneath them.

Agents are becoming customers, and the growth numbers are big

Our view: A new front door

Olivia Moore and Justine Moore of Andreessen Horowitz (a16z) published "Welcome to the B2A Economy" on 11 November, the first in a series. Their argument is that AI agents are becoming the main users of the internet, and that companies which spent years keeping bots out will now compete for their business (a16z, 11 November).

They back it with figures from other companies. Shopify reported AI-driven traffic to its stores up eightfold year on year in the first quarter, and orders from AI-powered searches up nearly thirteenfold. Human Security saw agentic browser traffic grow eightfold between August and October. On the build side, they cite GitHub's Octoverse: upwards of 40% of pull requests as of September started with an agent draft, up from 17% six months earlier. They also cite Vercel, where agent-triggered deployments rose 28 times over twelve months (same article). They say enterprises still often block agent traffic over fraud and chargeback worries.

The firms closest to commerce and code, Shopify, Stripe, GitHub and Vercel, are all seeing agent activity grow fast, and a16z is putting a name on it: a business-to-agent economy. The point worth taking to your own team is the last one in the article. Many enterprises still block agents by default, and that is a choice an operator can make on purpose.
Why this matters for iGaming. We think a player's personal agent will become a new front door next to the website and the app, starting with research and comparison. Operators already run fraud and bot controls designed to keep automation out, so there is a good conversation to have about how a friendly agent acting for a player gets told apart from a hostile one.

BCG: only 5% of companies have the full set of controls for agents

Our view: How products get built and businesses run

Boston Consulting Group's Applied AI Index 2026, published on 30 September, is based on a survey of more than 1,300 senior leaders across over 20 sectors. BCG says almost half of companies now get value from AI. It labels 7.5% "future-built" and 41% "scaling", and says companies strong on both strategic clarity and applied AI generate five times the AI value of those that are not (BCG, 30 September).

On agents, BCG puts their share of total AI value at 22%, up from 17% in the 2025 sample. It expects 42% of companies to give agents autonomy by 2030, and says only 5% have the full set of critical controls in place today. BCG says about 80% of AI spend now sits outside enterprise IT, and that firms tracking AI against the P&L realise about three times the value of those that don't. Its rule of thumb is a 10-20-70 split of effort: 10% on algorithms, 20% on technology and data, and 70% on people, organisation and process (same report).

BCG's numbers say the gap is less about the technology than about everything around it. Almost half of companies get value from AI, only 5% have the full set of controls for agents, and BCG puts 70% of the effort on people, organisation and process. The firms it calls future-built are the ones tracking results against KPIs or the P&L, 95% of them, and running multiyear programmes.
Why this matters for iGaming. BCG's numbers match what we have been hearing this week. Many organisations have started with personal productivity by giving everyone the tools, which we think of as mass training: unstructured and in some cases very expensive. With about 80% of AI spend sitting outside IT, it is easy to see how that spreads into admin for other teams, with no one owning it. Everyone is building and everyone is learning, but the line from that spend to revenue is weak, and that is the gap BCG's P&L finding points at. The firms it calls future-built track AI against results, and they realise about three times the value. In iGaming the revenue levers are well known: acquisition, conversion and retention. We think the leaders who pick one of them, point an agent at it with approval and a revenue measure from day one, will turn what their teams have already learned into a result, and the unstructured training becomes a head start.

More than 5% of UK adults now pay for AI themselves

Our view: Acquisition is changing

Bloomberg reports (4 November) on a study by academics at Warwick, LSE and Imperial, based on 14 million UK bank and card accounts. UK consumers spend over £1 billion ($1.4 billion) a year on AI subscriptions. More than 5% of adults paid for their own subscription as of July, about double the share a year earlier. Most of the spending still goes to ChatGPT, but Claude's share of spending quadrupled between December 2025 and July 2026, partly because its users are more likely to take pricier plans (Bloomberg via Yahoo Finance UK). Separate research from Stripe puts UK consumer AI adoption about 10% behind the US (same article).

This is built on payments data, so it is real money, not survey answers. Paying for an assistant has gone from a niche habit to something one in twenty adults does, and with Claude's share of spending quadrupling in seven months, the leader's position is already being tested.
Why this matters for iGaming. Someone who pays for an assistant is likely to ask it things often, and we think that is where more first impressions of operators will start to form. Operators can't advertise in ChatGPT today, so being described accurately has to be earned.

Anthropic shows how agents can learn from their own sessions

Our view: How products get built and businesses run

Lance Martin of Anthropic describes how memory works for its agents and proposes "dreaming": an offline pass that reads past session transcripts, fixes mistakes, writes notes and promotes the useful ones into future sessions (Anthropic, 6 November). On Anthropic's CL-Bench database task, mean reward for Opus 5.5 rose from 0.655 to 0.838 with memory, and for Sonnet 5.5 from 0.644 to 0.833. In one example, a corrected note about a location let an agent playing Pokemon reach the next city, while the run with the wrong note never did (same post).

Memory lifts scores across Anthropic's models, and the Pokemon example shows why it matters: one wrong note can send an agent the wrong way until someone corrects it. For a team building with agents, what the agent remembers is worth reviewing, the same way you would review a new starter's notes.
Why this matters for iGaming. We think agent-assisted build cycles are where software factories start, and an agent that remembers your conventions and decisions is a big part of what makes them useful. Product and engineering leaders who treat what an agent learns as something the team reviews and owns will get more out of each week of use.

Other things we've been reading

  • -Rhythm at Meta. The @Scale team describes an internal system that turns linked project sources into plain-language project status with citations, and shows each insight only to people who are allowed to see all of its sources. Meta's authors report 200+ projects and a drop in the critical edit rate from 26% to 6% over three weeks (@Scale, 2 November).
  • -The work between the work. Stuart Winter-Tear argues that speeding up single tasks leaves handoffs, approvals and reconciliations as they were, so how work is divided is worth another look. The essay introduces his book, A Management Theory for the Age of AI (Unhyped AI, 1 November).
  • -What a services firm expects. Eddie Siegel, co-founder and CTO of Ode, sets out how the firm expects its work to move towards specs, agent factories and outcome-based pricing as code gets cheaper (Ode).

Work with us

If one of these is useful to you, send us a message or visit bettingproduct.ai/products. We start with a fixed-scope AI Discovery and Pilot: one problem, one pilot, one decision.

Ian and Vlad | BettingProduct.ai