
Analytics software specialist, Alteryx, recently launched a new set of capabilities to its platform that enables clients to take a firm grip of their artificial intelligence usage costs. The California-based company has connected AI agents to reusable “build once” business context workflows that dispense of the need to recreate business logic from scratch. The company estimates this can reduce firms’ token consumption by as much as 93% and accelerates completion of tasks involving raw, ungrounded data by as much as 85%.
Ben Canning, Alteryx Chief Product Officer, told Data Management Insight about the origination and development of the agents.
Data Management Insight: Hello Ben. What was the genesis of this development – did Alteryx see a need for this or was it prompted by client feedback?
Ben Canning: The need was already visible, and customers made it impossible to ignore. Enterprises have invested heavily in LLMs and data infrastructure over recent years but continue to encounter a hurdle to get the payoff from those investments. Business logic that determines how a company actually reconciles, forecasts, calculates and approves often exists outside central cloud data platforms (trapped in spreadsheets, workflows, applications and people’s heads) and beyond the reach of AI systems – which is a significant handicap on what those systems can be used for. This problem has become more pressing with the introduction of AI agents, which require context and an understanding of business logic to integrate with how a business actually works.
DMI: How can this be leveraged by capital markets participants?
BC: Capital markets firms run on reconciliation. Front-office and back-office data has to match. Positions, exposures and regulatory reports are also calculated according to defined rules. Those are deterministic processes. Processes that aren’t suited to an LLM that gives a slightly different answer to the same query each time it’s run. In many capital markets workflows, outputs need to be exact and reproducible: the same inputs have to produce the same number every time, and someone has to be able to show why.
The opportunity for firms is to give staff the ability to interface with trusted processes without having AI recreate the underlying logic. A user could interact with an agent to investigate a reconciliation break, interrogate an exposure or understand the data behind a report, while the underlying calculations continue to be performed by the governed Alteryx workflow. It’s a much more intuitive way of interacting with complex financial data without ceding control or oversight that’s crucial to good governance.
This is already running. NextWave reconciles trades this way: a set of 120 trades that take around four hours now resolves in about thirty seconds, and the governed workflow identified every break, where the model working on the raw data on its own missed two of them. HSBC has gone from four days to minutes on a comparable process.
DMI: How does the system handle cases where incoming data is ambiguous or ungrounded before passing it off to an Alteryx workflow?
BC: Alteryx can bring messy data that is siloed across different systems together, cleaning, reconciling and preparing it so that AI has trusted data and business logic to work from.
Where the data is genuinely ambiguous, what matters is what the workflow does rather than what the model guesses. The workflow is explicit about what it expects, so an input that does not match is not quietly absorbed into an answer: it comes back as an exception with a reason attached. The agent receives “this cannot be reconciled, and here is why” instead of a fluent answer with an unflagged assumption buried inside it. That distinction matters more than it sounds, because an ungrounded answer from an LLM looks exactly like a grounded one. Resolving ambiguity at the point of data processing, in logic the business can inspect, repeat, audit and is the only place it can be caught before it reaches a person.
DMI: The announcement cites up to a 93% reduction in token consumption. Which parts of the task execution are shifted away from the LLM context window into Alteryx to achieve those savings?
BC: The data-heavy processing and deterministic calculation are shifted into Alteryx. Instead of loading raw data into the LLM and asking it to work through the analysis itself, Alteryx performs that work and gives the model the result to interpret and present.
Alteryx processed sales data and analysed margin across categories, channels, regions and suppliers in a mock margin leakage scenario and performed calculations in an eCommerce test. LLMs are reserved for focus on interpreting and communicating the answer rather than consuming tokens reasoning through all of the underlying data which is where token consumption spikes.
Those were controlled tests outside financial services. On the finance side, our partner NextWave achieved a 20x (~95%) reduction in token consumption on trade reconciliation by moving the calculation into the workflow.
DMI: Does Alteryx offer pre-packaged agent connectors out of the box for major enterprise platforms (e.g., Salesforce, SAP, ServiceNow), or does each integration require custom workflow endpoints?
BC: There are two directions here, and neither requires custom endpoints. Inbound, the Alteryx MCP (Model Context Protocol) Server is a single governed access layer: Claude, OpenAI, Gemini and Microsoft Copilot Studio connect to it once and can reach any workflow the organisation has chosen to publish, so there is no bespoke integration per agent or per model. Outbound connectivity to Salesforce, SAP, ServiceNow and the rest is what Alteryx has shipped for years: over a hundred sources, including the warehouses, the legacy databases and the spreadsheets that never reach a warehouse at all.
The practical consequence is that the workflow is the integration. What a firm builds is the logic it wants an agent to be able to call. It does not build plumbing to each system, and it does not rebuild anything when it adds a second agent or swaps the model underneath.
DMI: How does the platform handle workflow exceptions or unexpected edge cases triggered by an agent? Is there a built-in mechanism to halt execution and route to a human for validation?
BC: Yes, and the mechanism is the workflow itself rather than a policy layer sitting on top of it. A workflow returns typed outputs, so an exception is a different kind of output from an answer: the agent receives an explicit exception with its reason, not a best guess it can present as a result. Where that output goes next, including stopping and routing to a named person for validation, is designed into the workflow by the person who owns the logic rather than left to the agent’s discretion at runtime.
Teams confirm that behaviour before anything is exposed to an agent. Our test tooling lets a designer run the workflow against the edge cases they are worried about and see exactly what comes back, which is the part most firms in regulated functions ask about first. And because every run is versioned and logged, a decision can be replayed for an auditor months later: which inputs, which version of the logic, which exception, and who approved it.
Subscribe to our newsletter


