Why Trade Capture is Still a People Problem in 2026

OTC trading desks have spent the last decade investing in execution technology, connectivity and risk infrastructure. Matching engines are faster when using on-chain technologies. Data distribution is more sophisticated. AI is being piloted across front-office workflows.

And yet, at the end of most OTC trade workflows, there is still a person reading a chat message and typing what it says into a system.

That step, the transcription step, has survived every wave of automation. It sits between the deal agreed in chat and the record that needs to exist in the OMS, the ETRM or the reporting system. It is manual, error-prone and, in markets with five or fifteen-minute reporting windows, it is a risk.

Four reasons automation hasn't reached trade capture

1. Accuracy wasn't good enough

Early attempts to automate trade capture from chat ran into a fundamental problem: OTC markets don't communicate in structured data. A physical gas trader confirming a deal might type "same as before +2 for Nov, done." A rates desk might use shorthand that only makes sense with knowledge of a prior conversation. A credit trader might confirm across three messages with corrections in between.

Rule-based systems couldn't parse this reliably. The fear of a misbooked trade, one wrong field in a high-value transaction, outweighed the cost of manual entry. That was a legitimate concern, and it kept automation out.

2. The governance question was unanswered

Regulated firms need to demonstrate that every trade is traceable. A system that produced a booking without a clear, auditable chain of evidence posed a compliance risk regardless of its accuracy. Nobody wanted to explain to the FCA or CFTC how a trade was booked by a model they couldn't interrogate.

Without a credible answer to the governance question, automation stalled at the point of adoption. IT could build it. Compliance wouldn't approve it.

3. Every desk does it differently

Different chat platforms, different confirmation formats, different OMS and ETRM systems, different asset classes with their own shorthand and conventions. Building a solution that worked for one desk meant rebuilding it for the next. The economics didn't work for most vendors, and bespoke builds carry a maintenance overhead that compounds over time. So the problem stayed manual because solving it at scale seemed too hard.

4. The cost was diffused and therefore invisible

The person doing the manual entry was rarely the person with budget authority. The cost of the problem didn't appear as a line item. It was spread across hundreds of small inefficiencies every day: time lost, errors corrected, booking windows missed, late reports filed. Each one is individually absorbable, but never added up into a number that forces a decision.

What has changed

Each of those four constraints has now been resolved, or is in the process of being resolved.

Accuracy is now provable

LLMs trained on domain-specific OTC chat data can now reach an accuracy higher than humans alone typically achieve. At a large North American physical gas trader, field-level accuracy has reached over 90%, a figure that reflects the model learning the abbreviated, context-dependent language of energy-market chat and improving through use. That number is verifiable and tracked against real trades in production.

Manual data entry carries its own error rate, amplified by time pressure and volume. At the field level, the automated system now outperforms the manual process it replaces.

The governance question has been answered

The human-in-the-loop model preserves accountability without sacrificing speed. Every inference is traceable to a source message. Every confirmation is logged as a discrete linked event. The audit trail is built as the workflow runs, because each step from chat message to booked trade is recorded as it happens, not reconstructed after the fact.

Regulators can follow the chain. Compliance teams can interrogate the output. A human reviews and confirms before any record is created. The automation handles interpretation. The decision remains with the desk.

The desk-by-desk problem is solved

A configurable orchestration layer means the same core engine adapts to different chat platforms, confirmation formats and downstream booking systems without rebuilding from scratch. Capture from chat, email, Excel, API or custom application. Deliver into exchange block entry systems, ETRM platforms, OMS, risk systems or reporting infrastructure. The same architecture that works for one desk scales to the next without a bespoke rebuild each time, which changes the economics for vendors and clients alike.

The cost is now visible

For years, the problem sat below the level where decisions get made. What has changed is that the cost can now be quantified and placed in front of the right people: trade-entry time halved, field-level accuracy that exceeds manual entry, and booking windows met rather than missed. These are numbers a business decision can be built around, in a way that "the desk spends a lot of time on this" never was.

What the workflow looks like now

Traders and brokers stay in their native chat environment. Nothing changes on their side. When a trade is done, the message is passed to ipushpull. AI fills a draft booking ticket, extracting the relevant fields. A human reviews the draft, confirms it, and the trade flows into the downstream system in seconds.

Where confirmation arrives across multiple messages or where corrections are made mid-conversation, the system handles that too. The system learns how your own desks communicate, not a sanitised version of it.

The same approach works across asset classes: energy, OTC derivatives, fixed income, FX and credit. Any workflow where someone is currently reading a chat message and manually filling in a form with what it says is in scope.

Read more: From chat to booked trade in near real-time

The remaining problem is organisational

The desks still doing this manually in 2026 are not doing so because no solution exists. They are doing so because the decision hasn't been made yet. After all, the cost remained invisible for long enough that it became normal.

That is the last barrier. And unlike the others, it cannot be solved with engineering.