AWS pushes software partners to charge per result, and a 45% price cut became a 20% bigger bill

Zendesk charges per support ticket its AI closes, not per seat. Intercom bills $0.99 each time its Fin assistant resolves a conversation. Salesforce sells Agentforce at $2 a conversation, or at about $0.10 for each action an agent takes.

AWS is pushing the rest of its software partners the same way. Allison Johnson, who directs the AWS Americas technology partners team and works with roughly 400 of them, told Channel Dive there is "no way to go back to that world of a five-year-deal." An AWS market study this year found 80% of customers moving to outcome-based commercial models.

Intercom's help documentation counts a resolution when Fin answers and the customer either confirms it helped or leaves without asking for more help. A customer who gives up in silence bills the same $0.99 as one who was helped.

Anthropic released Claude Fable 5.1 this week with its list price unchanged, $10 per million input tokens and $50 per million output tokens, and cut the price of cached reads by 75%, to $0.25 per million. It puts the saving at about 25% on ordinary work and as much as 45% on agent-heavy work. Artificial Analysis, which tested the model before release, found that at the highest effort setting Fable 5.1 uses about 1.7 times as many output tokens as Fable 5 and costs 20% more per finished task: $3.76 for one task on its Intelligence Index, against $2.34 on Anthropic's Opus 5, which scores three points lower.

Infosys, Deloitte, MIT and Gartner each published research this year that arrives at the same place: AI budgets grow every year, and most projects stay small tests that never demonstrate money saved or earned. A test that stays a test still costs the licences, the staff hours and the consultants who set it up.

Companies that measured the return on each AI project separately, and closed the ones that did not deliver, reported payoffs on 81% of their initiatives.

The projects that pay worst are the ones companies most like to announce: automated threat detection and the IT help desk.

About 40% of Kier's 11,000 employees use Microsoft Copilot every day. The UK construction group got there through short sessions where staff show each other prompts that worked, rather than through required training, and it pointed the tools at the riskiest parts of the work first: agents that track safety data and read drone footage from sites. A Copilot licence costs the same whether the person opens it daily or twice a quarter.

M&T Bank, a 170-year-old regional lender with $212.9 billion in assets, spent years moving its technology work back in-house before it handed copilot tools to more than 15,000 employees. It knew where its own records were held before it bought anything.

Delta has opened its AI travel assistant to every SkyMiles member. Its technology chief says the payoff depends less on the software than on whether staff use it, which is why Delta trained staff before turning anything on for customers, and why anything a customer sees keeps a named employee behind it.

e.l.f. Beauty used to reply to 30% to 40% of the comments customers left under its TikTok and Instagram posts. It now replies to more than 80%. An AI tool the company built drafts as many as 90% of those replies, and a person approves each one before it appears.

Answering comments used to scale only by hiring: someone reads each one, decides whether it needs a reply, and writes it in the brand's voice.

The staff finding these uses are not engineers. e.l.f. gives them pre-approved building blocks they can assemble into small tools for their own work, so a marketer with an idea does not wait in a queue for the technology team.

That is how a cosmetics company ends up running 85 further AI pilots at once, and pushing most of them into daily use within six months.

Meta released Muse Voice Transcribe this week at $0.18 an hour of audio. It transcribes as people speak, tells apart more than 20 speakers in one recording, handles dozens of languages, and gets 3.1% of words wrong, a lower error rate than the comparable services from Google and OpenAI, which also charge more.

That is the work underneath the meeting-notes and call-recording subscriptions sold per person per month: a claims call, a sales meeting, a service line, separated by speaker and searchable while it is still going on.

Text written by Claude now carries a hidden watermark, and Anthropic has opened the tool that reads it to regulators, journalists and businesses. The EU rule behind the change took effect in August: AI companies must mark the text their models generate, and penalties reach 15 million euros or 3% of worldwide sales. A report, a tender response or a contract clause drafted by Claude can be identified as such later, by whoever received it.

Perplexity released a feature this week for the same worry from the other end. Hybrid Compute, on Mac, keeps sensitive files on the person's own computer and sends only the non-sensitive parts of a task to cloud models. The sensitive file never leaves the laptop.