Daniele presented a three-agent architecture built on IBM Planning Analytics, with a human checkpoint at every stage. He was explicit that this is meant to work as a copilot, not an autopilot.

Agent 1: month-end close acceleration. Confirms data availability and loads it. IBM Planning Analytics then runs intercompany elimination, FX translation, and the other model calculations. The agent checks trial balances by entity, flags unmapped accounts, and classifies any open issues by materiality. If something material is open, it stops and hands off to a human. Once cleared, it writes a narrative summarizing what it did.

Agent 2: variance analysis. Runs continuously against GL actuals, budget, and forecast across entities and a defined set of KPIs (Daniele noted the business chooses which KPIs to watch, to keep the output focused). It applies absolute and percentage thresholds to sort variances into priority tiers, for example: priority 1 if a variance exceeds $75,000 against both actual and budget, priority 2 if it exceeds $75,000 against either forecast or the prior-period variance. It drills into root cause across dimensions such as region, product, and channel, can identify the responsible budget owner, and writes a scored narrative that both a human reviewer and the next agent can use. It also draws on a "skills" library where finance teams encode business-specific context that a generic threshold rule wouldn't catch, such as a production line stoppage that lowers sales and COGS at the same time.

Agent 3: autonomous forecast. Triggers on a schedule, a material variance event, or a manual request. It reads current actuals, the prior forecast, agent 2's variance commentary, and existing driver assumptions, then assesses which driver logic still fits each signal, flags one-off items to exclude, and logs every change it makes. Calculation itself is handed to IBM Planning Analytics rather than done by the model. Daniele's phrase for this: "you don't want a probabilistic engine trying to do calculation if a deterministic engine does it every time the right way." The agent runs scenario comparisons against the prior forecast as a baseline, then posts its output as a separate version for the CFO to review, question in plain language, adjust, and approve before publication.

During Q&A, Daniele added that the platform is model-agnostic: customers can bring their own LLM, local or cloud, connect their own OpenAI or Anthropic API keys, or run Apliqo's SaaS version, which uses AWS Bedrock with frontier models including Anthropic's Claude. He also said the skill system, more than the underlying model, is where most of the business value sits, since that's where finance teams encode judgment a generic model wouldn't have.