Cheaper AI, sharper scrutiny: what sponsors should change this month.

A widely tracked gauge of large language model prices hit a fresh low this week. Silicon Data's LLM Token Expenditure Index fell to 97 cents on Monday, its lowest reading since launch and more than half below its high earlier this summer ([1]). The fall is driven in part by open-source Chinese models priced below the frontier labs, alongside OpenAI's price cuts on two GPT-5.6 models in late July ([2]).

For a portfolio company running AI in customer service, finance or software development, tokens are an input cost. Cheaper tokens improve the payback on use cases that looked marginal a quarter ago. Value creation plans sized on last year's unit costs are now conservative on cost, which gives operating partners room to widen scope without raising the budget. The discipline is to bank the saving in EBITDA rather than let usage grow to absorb it.

The other side of the trade hits software holdings. Syz Group's Charles-Henry Monchau argued that the moat "must shift away from raw model capability" towards "distribution, memory and context" ([3]). That logic reaches well beyond the model labs. A portfolio software business whose AI feature is a thin layer over a third-party model will see its pricing power erode as the underlying model gets cheaper and more interchangeable. Businesses with proprietary workflow data, embedded distribution and deep customer context are better placed to hold price.

In diligence, this changes the questions. Buyers should ask what share of a target's AI revenue depends on capability anyone can now buy cheaply, and what share rests on data and distribution the target owns. Sellers preparing for exit should evidence the second. Silicon Data's head of research suggested there may already be enough supply to "provide sufficient capabilities for most tasks" ([4]), which makes model capability alone a weak equity story.

UKG, the privately held HR software group formed from the 2020 merger of Kronos and Ultimate Software, gave a rare, numbered account of an AI programme this week ([1]). Its chief information officer, Prakash Kota, said his first major project was not AI at all. It was finishing long-delayed IT integration from the merger: centralising the technology function, consolidating ERP and CRM onto single platforms and merging data warehouses ([2]).

Scale came after that. UKG has launched 387 internal AI applications from more than 1,400 employee-submitted ideas and built more than 12,000 agents ([3]). AI voice and chat agents now resolve an estimated 27% of customer calls autonomously, and the company says AI adds 8,500 hours of productivity each month ([4]). It measures customer service AI on upselling and customer sentiment as well as productivity, and asks its leaders to allocate spend across "talent, tools, and tokens" ([5]).

The lesson for buy-and-build platforms is sequencing. Many sponsors pencil AI savings into year one of a merger plan, before data and systems are joined up. UKG's order (integrate, consolidate data, then deploy) is the one that produces numbers a buyer will accept at exit, because the savings can be traced to specific workflows rather than asserted.

Sponsors are applying the same logic to their own firms. Partners Group reported on 1 September that total operating costs fell 5% to CHF 414 million in the first half, "mainly driven by lower variable performance fee-related personnel expenses and AI-driven productivity gains" ([6]). A listed sponsor naming AI in its own cost line gives operating partners a reference point when they ask portfolio CEOs for the same standard of evidence.

Labour-intensive services are where AI disruption risk and AI margin upside meet, and this week's data cut against the simple displacement story. Apollo chief economist Torsten Slok noted that call centre employment in the Philippines rose every year from 2016 through 2025, nearly doubling to 2 million ([1]). He argued that if AI were displacing white-collar work at scale, it would show first in the Philippines and India, yet unemployment in both countries has kept trending lower ([2]).

That is despite heavy exposure. The Brookings Institution estimated 86% of customer service representative tasks had high automation potential ([3]). Slok's explanation is Jevons paradox: as AI makes call centre work cheaper and faster, companies buy more of it, not less ([4]).

For sponsors holding business process outsourcing, contact centre or tech-enabled services businesses, the implication is not that the risk has passed. It is that the risk shows up in price before it shows up in volume. If clients buy more work at a lower unit price, revenue holds only if the provider keeps part of the productivity gain rather than passing all of it through. Contracts priced per seat or per hour are the most exposed. Contracts priced per outcome or per transaction let the provider keep the margin that AI creates.

UKG's experience is the client-side version of the same shift: an estimated 27% of its customer calls now resolve without a human ([5]). Every in-house deployment like that is volume an outsourcer does not win. Value-creation leads should model two cases for services holdings, volume growth at lower prices and in-sourcing by large clients, and track the share of revenue already moved to outcome-based pricing. Exit buyers will ask for all three.

Where sponsors see AI risk in software and services, they are finding AI demand in the physical trades. PE activity in construction and engineering hit a record 529 deals in the second quarter, up 56.5% from a year earlier, according to PitchBook estimates ([1]). HVAC contractors logged 76 PE deals worth $3.8 billion in the first half of 2026, already more than the 63 completed in all of 2025, and electrical contractors recorded 38 deals ([2]). Contractors working on data centres carry an average backlog of 11 months, against 8.5 months for everyone else ([3]).

These are fragmented, labour-heavy businesses, which makes them natural buy-and-build targets. They are also businesses where AI value creation is unglamorous: quoting, scheduling, job costing, field workforce planning and working capital. A platform that adds bolt-ons without joining up its data will struggle to show those gains at exit, for the same reason UKG put integration before AI.

In the UK, sponsors opened September with fresh take-private agreements. Veritas Capital agreed to buy Bodycote in a deal valuing it at £1.9bn, and Gamma Communications recommended a £1.1bn all-cash offer from Epiris ([4]). City AM counts more than 50 London-listed companies that have accepted offers or drawn interest from private or overseas buyers this year ([5]). The Bodycote offer is a 41.4% premium to its twelve-month average share price to May, and the Epiris bid a 53% premium to Gamma's price before takeover speculation emerged ([6]).

Premiums of that size need operational gains to earn a return. For UK mid-market sponsors, the value creation plan for a take-private now has to show where AI lowers cost to serve or lifts throughput, how it is phased and how it will be evidenced, from the first hundred days.