Strategic Environment Analyser Agent
Monthly manual sweep → continuous agent over 250+ sources
- 250+ sources
- 1,000+ stories / cycle
- 100+ baseline cells
- 4 stages
A four-stage agent that scans for regulatory change, tests each lead against a financial organisation's compliance baseline, and routes confirmed changes through human approval.
Problem
A financial organisation maintains a regulatory and compliance baseline: five frameworks broken into 100+ individual cells, each one a statement of what the organisation is and does for a specific regulatory requirement. That baseline is only as good as how current it is, and currency depends on continuously watching everything outside the organisation — upcoming legislation, compliance changes, conferences, industry news, developing stories across the financial domain — for anything that might mean a cell no longer holds.
Before automation, the strategy team did this by hand, monthly, through a general-purpose AI chat assistant: exporting source emails, uploading them, enabling web access, and prompting their way to a baseline-change artefact each cycle. It worked until it stopped scaling. As the volume of source material grew, the assistant’s context window became the bottleneck, instructions stopped being followed reliably, inaccuracies crept in from prompt and context drift, and the team had no way to raise the cadence beyond monthly without making all of that worse.
The hard part isn’t summarising news. It’s doing it at a volume where one cycle means 1,000+ candidate stories, each of which has to be verified against its own source, deduplicated against every other subscription that carried it, and then tested against 100+ baseline cells — without a false positive quietly rewriting a compliance statement.
Approach
The brief was to turn a reactive, manual, monthly process into a proactive agent that monitors continuously. I split it into four stages, each with its own narrow failure mode, connected by human approval gates rather than one pass asked to do everything.
- Discovery — taps 250+ premium subscription email sources (arriving roughly twice daily, plus monthly reports), SME-flagged items and SharePoint documents, extracting each story’s headline, short description and source URL as a candidate. One cycle’s material runs to 1,000+ stories.
- Verification and deduplication — follows each story’s URL to confirm it’s accurate and current, writes a short justification of why it’s relevant to the organisation, and collapses the duplicates that recur across subscriptions. The output is a much shorter, cleaner set of leads.
- Baseline testing — reads the existing baseline and asks one question per lead: does this change our baseline? A lead that doesn’t becomes a finding, recorded in the artefact but not a change. A lead that might branches to a human gate, where an email of supporting facts goes to senior management for confirmation before any cell actually moves.
- Artefact construction — builds an interactive HTML dashboard (the current cycle’s findings and top stories, the prior five months for trend context, the baseline across its five frameworks, and a references tab with every source URL) plus an abstracted PowerPoint version of the same material.
Two design decisions did most of the work. Artefact generation treats the previous cycle’s dashboard and deck as templates, stripping them to a data JSON object, updating that, then plugging it back into the untouched template — the artefact itself never passes through the model, so branding and layout stay identical cycle over cycle while the data changes. And the human gates sit between every stage, not just ahead of a baseline change, with every intermediate artefact written to SharePoint rather than held in memory, so a cycle can pause on management sign-off and resume later instead of restarting.
Outcomes
Live in production as a four-stage pipeline on Airia.ai, replacing the manual monthly chat-assistant workflow.
- The strategy team runs cycles without the export-upload-prompt grind, and can pause a cycle at any gate rather than holding the whole thing in one sitting.
- Senior management approves baseline changes from an email with the supporting facts already assembled, instead of being asked to trust a summary.
- Compliance reviewers get a fully auditable path: every finding, lead and approval traces back to its source URL.
- Cadence moved from monthly to continuous, so the baseline reflects the environment as it changes rather than when someone finds time to check.