Decoded: AI News for Alternative Assets (September-2026)

September 2026 Newsletter

Decoded: AI News for Alternative Assets

So…how much does this all cost?

The industry spent the summer chasing AGI while simultaneously arguing that it should slow down, sometimes in the same interview. Neither position changes anyone’s Monday. Two more practical questions came up this quarter instead: what these systems cost once they run at volume, and what happens when the model behind a process stops answering.

Here's what we at Aithon found interesting this quarter:

Cheap models, expensive habits

The Translation

Forecasting software expenses used to be a fairly simple and accurate exercise, but gone are the days of simple per-seat pricing. We are now in the agentic world and here the billing unit is the token, or roughly a word or part of a word. The meter starts with one request and continues running through increasingly long AI-driven workflows.

One single instruction can produce a very large bill. Ask an agent to complete a single task and it may run several searches, read whatever those searches return, call outside services, draft its answer more than once, check its own work and hand part of the job to another agent. The person who typed the instruction sees one request and one answer, but the Finance department sees an amalgamation of the hundred steps in between.

Counterintuitively, falling token prices does not solve this on their own. Accenture expects per-token pricing to drop 19% over the next two years while consumption rises 78%.

The So What?

Let’s start with the measurement issue, which is where most firms go wrong. Cost per token prices what goes in and says nothing about what comes out. “Out” in this context means a fully complete, timely and accurate result. The model and its tokens won’t get you there alone (generally)

A model that is 25% cheaper per token but has 2x the errors will end up costing you more in the long run. The trade-off won’t show up in token usage but will in the analyst hours needed to correct the issues and any downstream ramifications. Do not evaluate token costs in a vacuum.

With that said, cheaper open-weight models are on quite a hot streak right now, hence the shift in this summer’s buying behavior. Stanford found that a small model on local hardware answered 71.3% of real-world queries correctly, against 23.2% two years earlier, while the price gap stayed wide.

Make the cheaper model the default then make anyone who wants the frontier model give a reason for it.
Set a spending limit for each person, with a fast route to ask for more. A hard cap which blocks real work is a quick way to stall progress.
Show people their own consumption. Alerts at 50%, 75% and 90% of an allowance can change behavior before anyone has to intervene.
Train the team on how a model reads a conversation. Shorter prompts, a fresh session for each task, no resending history every exchange, etc.
Switch on caching and batching. In some scenarios, work that can wait runs at roughly half price overnight. The technology may be new but the approach to governing it doesn’t have to be. Measure twice, cut once…

The technology may be new, but the approach to governing it does not have to be. Measure twice, cut once.

The outage nobody planned for

This was by far the largest and most impactful model shutdown and it made no difference which cloud a firm ran it on. The direct API went down, as did AWS Bedrock, Google Cloud, Microsoft Foundry and Snowflake.

The AI race has been political ever since it became a race, and politics will continue to have an impact which will be difficult to forecast. Just as your trading teams may weigh political developments in their trading models, you must weigh them in your operating model choices. Thankfully you won’t have to pay too much attention if you set up your AI architecture with redundancy, or hire a vendor who will do that for you.

The So What?

Every firm reading this already buys redundancy in some shape or form. You run production in more than one availability zone. You have a documented recovery time objective. You test a failover on a schedule, and your administrator's SOC report describes theirs. Nobody signs a hosting contract without asking what happens when a region goes down.

That discipline does not always carry over to the model layer but it needs to asap. Running in three clouds protected nobody in June, because all three were serving the same model. The equivalent of a second availability zone is a second model and many have not selected that second model, nor tested its outputs.

Which model answers each request, and where does the processing happen? The answer should be specific and concrete.
What is the substitute when that model becomes unavailable? Be wary of any substitute that requires a code change rather than a configuration setting.
Has anyone put your our own work through the substitute? Send fifty of your own documents through both and compare the answers.
Will you be told when a substitution happens? This may not be necessary assuming the above exercise has been completed, but extra caution should be paid to critical outputs (i.e. anything going to investors)

We are all on a learning curve when it comes to AI – vendors, users and CTOs alike. Unfortunately, AI will not wait for us to get up that curve so it is paramount that you do your due diligence now.

Quick Hits

What You Need to Know

Our Take:This is an interesting way to think about translating processes for an AI operating model. Written procedures, training guides, implementation plans all fall over when you get to the last mile of customization. Why not let AI shadow you for a day to watch, and in the case of Anthropic, even listen? As practitioners, we have our reservations. But there is certainly a better way to capture that last mile than endless Visio diagrams.

Our Take: This feels like groundhogs day, likely because similar product suites have been launched before but also because they both target… you guessed it, the front-office. Neither one reaches the middle and back office, where our readers spend the day. Neither product touches fund accounting, waterfall calculations, expense allocations or investor reporting. That gap is the recurring shortfall of a generic tool sold into alternatives.

Our Take: Integration has always been the expensive part of fund technology. Alternative asset operations is multi-party and multi-vendor by nature, so a multi-agent future is not hard to imagine. A common protocol means the manager's agent and their third parties’ agents connect once through a single standard, rather than through custom code written for each pairing. A2A will be an important language for your vendors and your IT team to learn.

The Numbers That Matter

7%

Share of companies running fully autonomous agents in production, against 38% that require human approval on every action 

Nine out of every 10 companies keep a human in every AI-enabled workflow. Don’t feel like you are falling behind if you still have heavy human-in-the-loop

Jargon Decoder

Demystifying AI Terms

The Gut Check

This Month’s Question

About Aithon

Aithon Solutions delivers intelligent automation and data solutions purpose-built for investment management operations. Our proprietary technology seamlessly integrates with existing systems to enhance operational efficiency, improve reporting accuracy, and unlock deeper business insights. By combining domain expertise with applied AI, we help asset managers do more with less—adding new products and clients faster while driving better outcomes through reimagined processes.

Know more

Aithon

September 2026 Newsletter