Introducing Virtual Teammates

By Maël Garnier·1 June 2026·5 min read

A practical playbook for tax and finance leaders

Virtual teammates — AI agents that hold a standing role on the team rather than answering one-off prompts — are arriving in tax and finance functions faster than the operating models built to manage them. The leaders who succeed will not be the ones who buy the most capable model. They will be the ones who onboard it like they would any new hire: with a clear mandate, the right access, supervision, and a path to earn trust. This paper sets out how to do that.

From tools to teammates

A tool waits to be used. A teammate is accountable for an outcome. That distinction is the whole game. A spreadsheet macro or a query bot executes a task you direct; a virtual teammate owns a recurring responsibility — reconciling a tax provision, drafting a transfer-pricing memo, monitoring filing deadlines across jurisdictions — and is measured on how well that responsibility is met. Treating an agent as a teammate changes what you must put in place: not just a license, but a role definition, access rights, escalation paths, and a manager. Skip that, and you have an ungoverned tool making consequential decisions in a regulated function. Get it right, and you add reliable capacity that compounds.

The playbook

1. Define a named role before you choose a technology. Write the job description first: the outcome the teammate owns, the decisions it may make alone, the decisions it must escalate, and how success is measured. "Draft first-pass VAT return commentary for review" is a role. "Use AI for tax" is not. A sharp mandate is what makes everything downstream — access, controls, evaluation — tractable.

2. Scope the first mandate narrowly and reversibly. Choose a process that is high-volume, rules-based, and low-blast-radius if it goes wrong — deadline tracking, document classification, data-quality checks, or first-draft reconciliations. Avoid making the maiden mandate a final-signature decision. Early scope should let you build confidence without betting a filing on it.

3. Build the onboarding pack. A teammate is only as good as the context and access it is given. Assemble three things: the knowledge it needs (policies, prior workpapers, tax positions, templates), the systems it may touch (with least-privilege, role-based access and a clear audit trail), and the guardrails that bound its behaviour. Treat this exactly like provisioning a new joiner — because that is what it is.

4. Set the human-in-the-loop and escalation rules. Decide in advance what the teammate does autonomously, what it does with review, and what it must hand off. Make escalation triggers explicit: low confidence, anomalous figures, novel fact patterns, or anything touching a filed position. Every output should carry its reasoning and sources so a reviewer can check it in seconds, not reconstruct it from scratch.

5. Name an accountable owner. A teammate without a manager is a risk without an owner. A named person — not a committee — is responsible for its performance, reviews its work during probation, approves expansion of its mandate, and can pause it. Accountability for the function stays human; the teammate is leverage, not a transfer of responsibility.

6. Run a supervised probation. For an initial period, route the teammate's output through human review and log every correction. Those corrections are your richest signal: they tell you where context is missing, where guardrails are too loose or too tight, and when the teammate is ready for a longer leash. Expand autonomy on evidence, not optimism.

7. Measure like a teammate, not a tool. Track outcome quality (accuracy, exceptions caught, rework rate), reliability (consistency across runs), and trust (how often reviewers accept output unchanged) — alongside throughput and cycle-time. A teammate that is fast but needs everything redone has not earned its seat. The goal is dependable capacity, not raw speed.

8. Scale deliberately. Grow in two directions: deepen a proven teammate's mandate, and add new teammates for adjacent processes. Resist the urge to deploy a dozen half-governed agents at once. Each new role gets its own mandate, owner, and probation. A small number of trusted teammates beats a swarm of unaccountable ones.

What matters most

Across every deployment, four principles separate the programmes that stick from the pilots that quietly die.

• Trust is earned, not configured. Adoption follows confidence. Build it through transparency and a visible track record, not by mandating use. Reviewers who can see the reasoning will lean on the teammate; those handed a black box will route around it.
• Transparency is non-negotiable in a regulated function. Every consequential output needs traceable sources and an auditable trail. In tax, "the model said so" is never a defensible position before an authority or an auditor.
• Accountability stays human. Agents extend your team's capacity; they do not absorb its responsibility. Keep a person answerable for every mandate.
• Change management is the real work. The technology is the easy part. Redesigning workflows, retraining people to supervise rather than execute, and rebuilding controls around a human-plus-agent team is where most of the effort — and most of the value — actually lives.

Where to start

Pick one high-volume, low-risk process. Write the role description. Give it an owner, the right access, and a probation period with full review. Measure the outcome, not the novelty. When that teammate has earned trust, expand its mandate and add the next one. Introduced this way, virtual teammates stop being an experiment and become what they should be: dependable members of the team, managed with the same rigour you apply to everyone else who is accountable for the numbers.