Agentic AI
TPDOC AI runs transfer-pricing documentation through specialist agents, each with a defined job, a hand-off and a human checkpoint. The arithmetic is done by code, not by a language model, so every number can be traced.
The agent family
Same machine, seven builds. Every robot has a brushed-metal chassis and a dark visor with the three ascending bars of the SBC mark glowing inside it — that visor is the family signature. What changes is the chassis: TAIGA is round with twin aerials, SCOUT is tall with a single probe, BENCH is wide with a dish, SCRIBE is a narrow page, GATE is armoured and has no aerial at all because it does not broadcast. Each also has its own icon, and its own voice.
These seven are the faces of the family. The full TPDOC AI platform has 170+ agents behind them.
Presence state
Breathing, nothing else. The mascot is present but silent — no rings, no waveform, no pointer. This is what a user sees 90% of the time, and it must never pull focus from the document.
Host
The morning briefing, the voice channel, and routing every question to whichever specialist owns it.
Speaks first, then hands the floor over. Never talks across a sub-agent.
voice · system default · pitch 0.98 / rate 1.02
Sub-agents
The IRD questionnaire — group structure, segmentals, agreements, DEMPE and benefit-test evidence.
Asks only for what it genuinely cannot infer from the file.
voice · system default · pitch 1.02 / rate 1.06
Tested-party selection and the most-appropriate-method test, with a written rejection for every method not used.
Shows its rejections, not just its conclusion.
voice · system default · pitch 0.92 / rate 0.99
The FAR map for each entity and the comparability factors that flow out of it.
Every FAR it produces ends at a human signature.
voice · system default · pitch 1 / rate 1.04
Screening, the interquartile range, and the adjustment computation when a tested margin falls outside it.
Keeps an accept/reject note for all 340 candidates, not just the 11 kept.
voice · system default · pitch 0.88 / rate 0.98
The Local File under Rule 10D and Form 3CEB — drafted from approved sections, never from a template with gaps.
Writes from single sources so two sections cannot contradict each other.
voice · system default · pitch 0.96 / rate 1.05
Release conditions, the change log, and the audit trail of who approved what and when.
Holds everything until a partner releases it. No exceptions, no silent overrides.
voice · system default · pitch 0.85 / rate 0.95
The Hear buttons use your browser's built-in speech and only play when you press them.
Four states of the same three bars. No new shapes invented, no illustration bolted on — the logo itself becomes the loader, the microphone, the thinking indicator and the confirmation.
Handoff protocol
01
You ask TAIGA
Voice or text. TAIGA classifies the question and names the agent that owns it out loud, so you always know who is answering.
02
The specialist takes the floor
The orb's badge switches to that agent's tint and code, travels to the relevant element, and points at it before speaking.
03
It shows the work
Answer, the reasoning behind it, and a checklist of what you do next. Every figure links back to the source it came from.
04
GATE records the outcome
Your decision, who made it, and when — appended to the audit trail. Nothing reaches a filing without passing through here.
Every AI skill, agentic module and MCP tool in the platform, arranged around TAIGA at the centre. Pick a pipeline, or let the walkthrough run, and watch the work hand off stage by stage. Drag to orbit.
Snapshot of the tracker’s network view. Open the live tracker for current data.
Book a demo
We use Google Analytics to see which pages are useful. It only loads if you accept, and you can change your choice later from Cookie settings. Privacy policy
