After Token Usage Surged 86×, Man Group Found the Real AI Bottleneck Beyond the Model

How a large asset manager moved AI from the chat window into real investment-research workflows

Part of Frontwise’s Weekend Deep Read series: a closer look at how global financial institutions are putting AI to work.

In July 2026, a Bloomberg Odd Lots host asked Tushara Fernando, Man Group’s head of data and AI, how steeply company-wide token use had risen.

The answer: an 86× increase since January.

The more revealing point is who was using those tokens. Adoption was no longer confined to technology teams. Finance, operations, and HR were building AI-assisted software and automated workflows. On the investment side, AI tools had reached discretionary portfolio managers, quantitative researchers, and traders.

Man Group is not a newcomer experimenting with chatbots. It reported approximately US$253.6 billion in assets under management at the end of June 2026 and employed more than 1,700 people, including over 600 in quantitative or technical roles. In 2025, it invested more than US$135 million in AI, data, and technology capabilities.

That makes the 86× figure worth examining—but not because it proves that AI created 86 times more value. It exposes the harder questions that appear when a financial institution moves AI into production. How does a machine understand the meaning of internal data? How far may a research agent act? Who approves the output? Which business unit owns the token cost? How much institutional knowledge should be shared?

Key takeaways

• AI expands research capacity before it changes investment authority. Agents can monitor filings, broker research, alternative data, and podcasts continuously. To become part of an investment process, however, new information must be mapped to a position, a thesis, or a decision rule.
• Man Group is systematizing the research process itself. Agents can move from papers and datasets to economic hypotheses, code, market data, backtests, peer review, and finally a human investment committee. At the time of the interview, roughly 15–20 models had completed that chain.
• The newest model is not the moat. Proprietary data, institutional context, internal tools, backtesting standards, transaction-cost assumptions, and execution capabilities create a system that another firm cannot copy by buying the same model.
• An 86× increase in token use creates management questions. When tokens are consumed by persistent workflows rather than individual employees, budgets, ownership, and value attribution must be redesigned.
• The real constraints are organizational change and safe deployment. In a regulated industry, speed is not the only objective. Explainability, testability, permissions, and approval determine whether an AI workflow can enter production.

Estimated reading time: 12 minutes.

---

1. Not Summarizing for the PM, but Monitoring Every Relevant Signal

A traditional equity analyst may cover 50 companies at once. The analyst reads filings, broker research, company announcements, and alternative data, while also listening to industry calls and podcasts. The problem is not a shortage of information. It is a shortage of time.

Many changes that matter to an investment thesis do not arrive as major news. They may appear in a technical interview, an engineer’s aside, or a new relationship between two datasets.

Man Group offered a concrete example. A portfolio manager following the AI supply chain needs to identify the next infrastructure bottleneck. In a podcast, the head of engineering at a large cloud provider said that GPUs still mattered, but that sufficiently large data centers were becoming scarce and connectivity between data centers would become increasingly important.

That speaker might never appear on an investor call, and a conventional research process might miss the conversation. An agent can transcribe the podcast, identify the change, compare it with the existing investment thesis, and alert the portfolio manager.

The distinction is important: a summary answers, “What did the speaker say?” An investment-research system must answer, “Which hypothesis does this new information change?”

If AI merely compresses ten documents into ten paragraphs, the researcher still has to make the final connection. A useful research agent needs to know the portfolio, the original thesis, the variables being monitored, and the threshold for further investigation.

The objective is not to let the machine decide for the investor. It is to expand the investor’s field of attention.

Frontwise note: Many institutions treat “being able to retrieve all the data” as the finish line for an AI project. Man Group’s next step is to map retrieved information back to a specific investment hypothesis. Without that layer, faster information does not produce faster research.

---

2. Beyond Reading: Systematizing How Research Gets Done

In discretionary investing, agents help people detect change. In quantitative research, Man Group has gone further by involving AI across the research chain.

The process begins with papers and curated datasets. One agent proposes an economic hypothesis. Another turns the hypothesis into code, retrieves market data, and runs a backtest. Other agents review the work. The output then goes to a human investment committee, which evaluates the logic, evidence, robustness, and permitted use.

According to CTO Gary Collier, roughly 15–20 models had completed this path from idea generation through signal construction, validation, and committee review, and had been judged suitable for use in managing client assets.

That statement can easily be misrepresented as “AI already trades autonomously.” It does not mean that.

What matters is that AI is beginning to handle the slow and failure-prone middle of the research process: forming a testable hypothesis from a paper, translating it into code, connecting that code to internal data and backtesting systems, and packaging the result as evidence a human committee can review.

Man Group has used names such as AlphaGPT and Alpha Assistant for these systems. Its public descriptions are careful. A general-purpose model may write plausible, executable strategy code, but it does not know the firm’s portfolio-construction methods, in-sample and out-of-sample testing standards, volatility estimates, turnover controls, transaction costs, or slippage assumptions. Without that context, it produces code that looks like research but is not ready for use.

The value of an agent, therefore, is not merely that it writes code quickly. It must operate inside an established research discipline and leave behind a record that allows people to judge why the work is valid, where it may fail, and whether it may proceed.

Frontwise note: When financial AI moves from demo to production, the dividing line is not whether it can generate an answer, but whether it can produce an auditable research record. Answers can be fast; accountability cannot disappear.

---

3. Three Layers of Data—and the Most Underrated One

Man Group describes three broad categories of data.

The first is market data: highly structured and extremely large. Tick data from major exchanges alone can approach 1 TB per day.

The second is alternative and unstructured data. It is less consistent in origin, format, and meaning. A language model can read a sentence, but that does not mean it understands what a field represents in business terms.

The interview used credit-card data as an example. A model can see rows and columns without knowing whether a row represents a purchase, an authorization, or a settlement. It may not know how merchants, companies, industries, tickers, and time periods align across datasets. Teams must add descriptions, labels, metadata, and shared definitions before AI can move reliably between sources.

The third category is institutional knowledge and context.

This layer tells the system how Man Group runs a backtest, which outputs are trustworthy, how internal code should be called, how an investment report should be read, and which steps require review. It is not simply another document library. It is a machine-readable account of how the organization works.

At this layer, the newest frontier model becomes less decisive. A stronger model may improve a standalone coding task, but it cannot replace proprietary data, market access, internal tools, execution assumptions, and research standards accumulated over many years. Models are easy to switch; institutional context is difficult to reproduce.

That is why two financial institutions can buy access to the same models and achieve very different results.

The model sets the technical ceiling; organizational context determines whether the capability can be used.

Frontwise note: Many data projects optimize for access to more information. AI production systems must also answer whether the machine understands what the data means. Moving from volume to semantics is one of the most underestimated steps in data engineering for agents.

---

4. After an 86× Increase, Token Cost No Longer Belongs to One Person

Man Group anticipated rapid growth in AI use by late 2025 and built budget models for different users and scenarios. In 2026, it pushed usage data and budgets down to business units so that people closer to the work could decide where tokens should be spent.

The firm did not immediately build an automated router that assigned every request to the model it considered cheapest or best. It began with education.

The reason was practical. Some employees would write investment logic in one context window, then ask the same conversation where to eat lunch and what the weather would be. Experienced users recognize the problem immediately; a firm with more than 1,700 employees cannot assume consistent habits.

Other waste is less visible. A coding agent may execute a deterministic version-control command and then send the entire output back to the model, consuming tokens unnecessarily. Some steps should remain outside the agent loop.

Transparent budgets, user education, and tool-level optimization proceeded together—and token use still rose 86× in six months.

That is not necessarily bad. It suggests that AI spread from a few pilots into many real tasks. But it is not a performance measure either. As the Odd Lots hosts noted, usage must eventually be linked to another set of outcomes: lower cost, higher revenue, broader research coverage, or better investment decisions.

The deeper issue is that future token use will increasingly be driven by persistent workflows rather than individual employees. One agent may serve research, operations, and compliance simultaneously. Which department owns its cost? Who decides whether it keeps running? Who is accountable for the result? Man Group acknowledged that these questions were not yet fully resolved.

Frontwise note: Token spending cannot remain an undifferentiated IT budget. A better unit is a workflow with a named owner, defined output, and acceptance criteria. Once cost is mapped to specific work, the organization can distinguish waste from genuinely new capacity.

---

5. The Real Bottleneck Is Whether the Organization Can Change Safely

When asked whether Man Group would prefer more computing power or more data, Collier did not choose between them. He argued that the larger bottleneck was organizational change.

A regulated financial institution must balance rapid experimentation with controlled deployment. Methods for evaluating agents are still developing. The industry has no universal answer for how to test output, limit permissions, or prevent an automated action from causing an irreversible loss.

At the investment level, Man Group’s boundary is clear: every trade must ultimately be explainable. Before an AI system writes code, it states the economic hypothesis in natural language. Other agents test the work, and a human investment committee reviews the result.

This operating model did not appear overnight. For roughly 15 years, Man Group has used embedded engineers who work alongside quants and discretionary investors, learning the business continuously rather than delivering software from a distance. What the industry now calls forward-deployed engineering has existed inside the firm for years.

The lesson is easy to miss: for an AI agent to enter a workflow, someone must first understand that workflow in detail.

If engineers do not understand how investors form judgments, and investors do not understand what the system can and cannot do, both groups can settle for an attractive chat interface. Embedded collaboration builds something more valuable: shared judgment about when an output is usable.

Hiring standards are changing as well. Man Group expects new employees across technical, operational, and investment functions to strengthen the firm’s AI capability. For engineers, value is shifting from executing every detail manually toward designing end-to-end processes, coordinating multiple agents, and connecting teams.

Execution is becoming cheaper. Deciding what should be done is becoming more valuable.

Frontwise note: Man Group’s foundations were not created by one software purchase. Data platforms, research standards, embedded engineering, and cross-team collaboration accumulated over years. Agents amplify the value of that infrastructure.

---

6. Share Institutional Knowledge Without Pouring Every Secret Into the System

For AI to understand an institution, it must have access to human experience. In an investment firm, that immediately raises an incentive problem: if an individual’s distinctive judgment determines their value, why would they place every method into a shared system?

Man Group’s answer is not “share everything.”

Systematic teams have long used shared codebases and collaborative research, so knowledge is easier to accumulate. Discretionary teams are different. Common processes—how to run a backtest, read a report, or validate a dataset—can be shared. The most sensitive investment logic may remain restricted.

These two types of knowledge should not be treated the same way.

Common processes can become reusable AI playbooks or skills. Strategy-specific judgment requires tighter permissions, explicit provenance, and narrower use. Forcing every investor to contribute their “secret sauce” to a central knowledge base creates incentive problems and makes it harder to distinguish a general method from an individual view.

A financial institution’s knowledge layer should answer at least three questions: who supplied the rule, where it applies, and how it can be updated or withdrawn.

Without those answers, “institutional memory” becomes an unattributed pool of advice.

Frontwise note: Knowledge sharing does not mean making everything public. A scalable approach codifies common processes, keeps sensitive judgments inside the right permission boundaries, and preserves provenance and accountability for every rule.

---

Frontwise View: Before Choosing a Model, Answer Four Questions

The most useful part of the Man Group interview is not the quantity of AI in use. It is the specificity of the questions that follow adoption.

First, which workflow will AI enter?

“The research department uses AI” is too broad. Is the system reading documents, forming hypotheses, writing code, running backtests, drafting reports, or taking action? Without a defined step, authority and responsibility cannot be defined either.

Second, what institutional context does it receive?

A general model knows market concepts, but not your data definitions, research standards, internal tools, or execution assumptions. Without that context, a fluent model can produce work that looks professional but is not actionable.

Third, who decides whether it passes?

What evidence is required at each stage? Who may approve the next step? How is the workflow stopped or rolled back after a failure? “Human in the loop” is not enough; the institution must define where the human intervenes and which criteria govern the decision.

Fourth, which numbers measure value?

Tokens, calls, and active users show that a system is being used. Outcomes should be tied to research coverage, delivery time, error rates, approval efficiency, revenue, or risk-adjusted investment performance—with credible baselines for comparison.

Seen this way, the 86× increase means something different. It is neither proof of runaway cost nor a badge of success. It is a stress test of whether the institution’s data, workflows, budgets, and governance can support AI doing real work.

Man Group’s answer is not simply “buy a stronger model.” It is building the slower, harder-to-copy system around the model: proprietary data, institutional knowledge, internal tools, execution capability, and human approval.

In financial AI, the competitive unit is shifting from the model to the entire workflow.

If you are building AI workflows in research, investment advice, risk, or operations, which barrier is hardest today: data, evaluation, permissions, cost, or organizational coordination?

---

Sources and methodology: Bloomberg Odd Lots, “One of the World’s Largest Hedge Funds on Its 86x Growth in Token Spending” (July 9, 2026; interview with Gary Collier and Tushara Fernando); Man Group’s Technology and About pages; Man Group’s 2025 annual report and 2026 interim disclosures; and Man Group research including “What AI Can (and Can’t Yet) Do for Alpha” and “A Trend Following Deep Dive: AI, Agents and Trend.” The 86× token growth, nearly 1 TB of daily tick data, and approximately 15–20 research models are company statements from the interview; they are not independently audited measures of cost, benefit, or investment performance. This article is for industry research and information exchange only and does not constitute investment advice.