Enterprise adoption of AI inside capital markets firms has reached a point where the frontier models are widely available, the coding agents and orchestration frameworks are broadly the same from one firm to the next, and the cost of building something has fallen far enough that a business user with no engineering background can now build and run their own tools.
For trading desks, technology functions and the risk teams alongside them, the challenge has moved on from getting these systems to work, to identifying where they should be allowed to run, how their consumption should be paid for and measured, and who inside the firm gets to build on them.
That shift framed a panel session at A-Team Group’s recent TradingTech Summit New York entitled ‘AI and Automation – From Coding to Governance’, which set out to explore the evolution of agentic AI in trading and the transition from pilots to governed and scalable AI. The panel kept returning to a single idea: that what now separates the firms pulling ahead from those stalling in pilot phase is organisational discipline rather than access to technology.
When does a workflow actually need an agent?
The panel observed that much of what firms are trying to build does not actually need to be built at all, or at least does not require the technology being thrown at it. One panellist described a steady stream of frivolous requests, citing a proposal to build an agent that would track order flow to see how far a desk was running behind its VWAP schedule. Any desk not already doing that is in trouble regardless, the response ran, and an agent adds token cost and overhead to something a straightforward graph can handle with direct control over each step. The distinction was between genuinely needing an LLM for a task and reflexively using an agent simply because the tools made it easy.Panellists noted that early front-office wins were concentrated in highly manual, time-sensitive workflows. In sales coverage, for instance, a salesperson can spend hours preparing for a meeting by gathering flow, market colour, news, research, and client preferences. Automating this synthesis compresses hours of prep work into seconds. This obviously saves time and improves outcomes, provided the end user has the expertise to spot check the results. Firms are progressing from building isolated tools to developing ‘digital teammates.’ As one panellist observed however, although these AI assistants execute tasks traditionally given to junior staff much faster, they require the exact same level of oversight.
Why are firms measuring the wrong things?
An audience poll identified the lack of clear productivity metrics as the top risk of agentic AI adoption, but panellists were nearly unanimous that firms struggled simply because they measured the wrong things. They warned against easy-to-capture activity metrics like prompt counts, the percentage of code generated, or token volume.
One panellist cited a large tech company that incentivised token consumption, predictably leading developers to game the system without producing anything useful. Instead, the panel repeatedly urged firms to do the harder work of measuring actual outcomes. For developers, this means tracking whether teams ship faster, improve quality, or expose new bottlenecks. For sales or trading roles, it means measuring whether staff cover more clients or generate more revenue. Firms already using agile frameworks can leverage existing data (such as sprint velocity or time spent on refactoring and security) to establish a solid cost baseline. Ultimately, the panel cautioned that if a firm never defined what ‘good’ looked like before these tools arrived, it would struggle to prove what the AI actually changed.How do you stop tokens from quietly bankrupting a project?
Cost was a recurring theme, highlighting a strong need for organisational discipline. Panellists acknowledged surface concerns about rapidly escalating spend (such as reports of seven-figure daily token bills) and agreed that dedicated teams had to actively manage AI consumption. However, they warned against fixating strictly on token volume. One speaker argued that an agent’s economic value simply needs to exceed its operating cost; because consumption varies widely across users, artificially capping tokens inherently caps business value.
Furthermore, the panel cautioned that cheap tokens often conceal bad engineering. Inexpensive consumption allows inefficient agents to generate unnecessary activity unnoticed, quietly normalising poor coding practices. This lack of discipline around context engineering only becomes visible – and highly expensive – when firms upgrade to higher-tier models, which one panellist noted could multiply costs twentyfold. The recommended remedy was a deliberate, multi-model strategy. By routing basic summarisation tasks to lower-tier models, firms could reserve expensive, higher-tier models for complex reasoning and coding. Because few employees would ever deeply understand AI pricing, panellists suggested that a small central team publishing clear guidance on model selection was the most effective solution.
Can governance scale without becoming the bottleneck?
The panel framed governance not as a compliance chore, but as an enabler of scale. They acknowledged a familiar risk: lowered barriers to entry meant employees could easily build unmonitored tools, echoing the era when hidden spreadsheets ran critical business processes. Since this trend could not be reversed, the consensus was to constrain it rather than prevent it, with constraints operating at multiple levels, including runtime protection against prompt injection, static code scanning, and strict inventory management to track agent identities, data access, and patching. Anticipating stricter future regulations, panellists highlighted sandboxing as a key strategy. By isolating business-built tools, adding egress controls, and routing traffic through a central gateway, firms can effectively measure consumption and apply permissions without hindering non-technical users. When built correctly, the panel argued, these guardrails are exactly what allow innovation to scale.
Inserting a human reviewer from legal or compliance into every loop creates an immediate bottleneck. Instead, the panel advocated for letting AI govern AI. By baking compliance policies directly into the enterprise platform, the system could automatically hard-block, alert, register, or route tasks for human review. This approach relies on non-negotiable central configurations like logging, cost tracking, and safeguards against destructive commands – features that out-of-the-box coding tools lack for regulated environments. Ultimately, panellists favoured a middle path between unrestricted building and heavy centralised development: applying AI on top of pre-built, standards-based frameworks with baked-in permissions, ensuring teams delivered production-quality solutions rather than simply generating more code.
Where does this leave the people who used to write the code?
Panellists agreed that AI has significantly changed the developer’s role without replacing human judgement. The job had shifted from merely writing code to architecting workflows and defining intent. Successful developers now act as orchestrators, working closely with the front office to plan specifications before a single line of code is changed. Conversely, panellists warned that dumping an entire repository into an AI tool and asking for a sweeping fix was one of the worst ways to use these systems.
However, the panel expressed scepticism about who truly benefited from this shift. One speaker argued that AI could actively hinder junior developers by subsidising too much of their critical thinking. These systems are most powerful in the hands of experienced professionals with deep domain knowledge, who use them for targeted optimisation rather than generating applications from scratch.
The session ended on an unresolved question: could a new generation learn the rigorous discipline required – such as restraining agent builds, carefully managing context windows, and optimising token costs – if AI shortcuts the very experience needed to develop that judgement? Time will tell.
Subscribe to our newsletter
