Build vs buy: AI trading agents for exchanges and brokers

Building an AI trading agent means owning every layer: models, data and order connectors, guardrails, advice-bait testing, a confirmation UX, evaluation and ongoing maintenance. Buying means integrating a vendor’s agent and holding the vendor to those same standards. The decision usually comes down to time to launch, the team you can dedicate, how much control you need, and who carries the risk.

This guide lays out what building actually involves, compares the two paths, and ends with the questions that tend to settle it. It avoids cost and headcount figures on purpose. Those vary too much by venue to be useful as general numbers.

What building an AI trading agent requires

A demo that answers “what is BTC doing?” takes days. A production agent that traders can rely on, and that compliance can sign off, is a different project. These are the parts.

1. Models. A reasoning model to interpret questions and write answers, plus faster models for routing and simple lookups. Someone has to choose them, tune prompts, and re-test when a provider ships a new version.
2. Data connectors. Read access to live prices, order books, funding, positions and balances, with a timestamp on every answer. Stale or wrong data is the fastest way to lose trader trust.
3. Order connectors. A write path to the order API that fires only after confirmation, handles rejections, and maps natural language to exact quantities.
4. Guardrails. Output rules that stop the agent from recommending trades, predicting prices or ranking assets. See how AI trading agents avoid investment advice.
5. Advice-bait testing. A large suite of questions written to provoke advice, run against every model and prompt change, with the pass rate as a release gate.
6. Confirmation UX. An order ticket that shows every field that matters and waits for the trader. The reasoning is in why confirm-by-default matters.
7. Evaluation. Accuracy checks on research answers, quantity resolution, and order drafting, plus human review of failures.
8. Ongoing maintenance. Models change, APIs change, new products launch, and traders find new phrasings. The work does not end at launch.

Each item is buildable. The question is whether your team builds all of them, keeps them current, and still ships the rest of the roadmap.

Build vs buy, side by side

Dimension	Build in-house	Buy and embed
Time to launch	Long. Every layer above has to exist and pass review before traders see it.	Shorter. Integration work only, if the vendor’s agent is production-ready.
Team needed	Engineering, ML, product design, compliance and QA, dedicated for the long term	Engineers to provide APIs and auth, plus product and compliance review
Ongoing cost drivers	Staff, model usage, test-suite upkeep, monitoring, re-testing after model updates	Vendor fees, integration upkeep, internal review time
Control	Full control of behaviour, roadmap and data	Control through contract, configuration and the vendor’s roadmap
Advice-boundary risk	Carried entirely by your team, including building the test suite	Shared with the vendor; depends on the vendor’s testing and evidence
Differentiation	Possible, if the agent is a core product bet	Comes from how you use and position it
Exit	No vendor dependency	Depends on contract terms and data portability

Neither column wins on every row. A venue with a strong ML team and a long horizon may value control enough to build. A venue that wants traders using an agent soon, without pulling engineers off the core exchange, often buys.

Where building tends to stall

A common pattern: the first version works in a demo. Then three problems arrive.

• The advice line under pressure. Direct questions are easy to handle. Multi-turn pressure, role-play and leading questions are not. Catching them takes a large, maintained test suite, not a prompt.
• Quantity resolution. “Close half”, “buy 5”, “sell the rest”. Each phrase needs an exact, echoed quantity before confirmation.
• Maintenance load. Every model update can shift behaviour. Without automated regression testing, each upgrade becomes a manual review.

None of these is a reason not to build. They are the reason to plan for them from the start.

What buying still requires from you

Buying is not zero work. The exchange’s engineers still provide market data access, account and position APIs, an order API, and authentication. The integration pattern matters too. A thin-client panel asks for much less than deep integration into native screens; see thin-client SDK vs deep integration.

Compliance and legal review also stay with you. A vendor can show how its agent is designed to stay on the information side of the line. Your team still reviews each market you launch in.

Questions that decide it

Work through these with product, engineering and compliance in the room.

1. Is the agent a core product bet or a capability you need? Core bets justify building. Capabilities often justify buying.
2. When do traders need to see it? Compare that date with an honest estimate of every layer above.
3. Who would own the advice-bait suite? If no team can build and maintain one at scale, that layer has to be bought.
4. What happens at the next model update? Name the person who re-tests and the process that gates the release.
5. How much of your native app can change? If the answer is very little, the integration pattern narrows the options.
6. Who owns the trader relationship? Confirm that any vendor’s agent stays loyal to your venue and submits orders only to it.
7. How will you measure impact? A controlled pilot with a holdout group answers this for either path; see measuring AI agent impact at an exchange.

If buying is on the table, the AI trading agent vendor checklist turns these into 20 specific questions.

Related

For the basics, start with what an embedded conversational trading agent is. For how agents differ from the tools they get confused with, read trading agent vs support chatbot. More in embedded agents.

Hippo is an embedded trading agent built for exchanges and brokers; askthehippo.com explains how a pilot works.