The practice behind VirtuBrain.

Most arguments about AI in finance start one level too high. They ask whether a process can be automated. The useful question sits at the step, and it is about checking. When this step is done, how would anyone know it was done right?

Some steps answer with a record. A bank reconciliation ties to the statement, item by item, or it doesn't. An intercompany match agrees to the cent on both sides, or it doesn't. When the check is that clean, an agent can own the step, because every result can be compared with something outside the agent. The check is logged, and the log is the evidence.

Other steps produce work that needs judgment to confirm. An accrual drafted from open purchase orders can look reasonable and still be wrong, and only someone who knows the business can tell. Flux commentary can be drafted from the numbers, but whether it explains the variance is a judgment. Here the agent drafts and a person approves.

Then there are steps where accountability can't be delegated at all. A judgmental provision. The close sign-off. The conversation with the auditor. Agents can prepare the file, and a named person decides.

Most processes contain all three kinds of step, often inside the same task.

Put the three side by side and they form a gradient. That is why blanket rules about AI in finance fail in practice. A policy that keeps agents away from journal entries blocks clean, checkable work. A policy that lets them run the close end to end ignores the steps a controller must own.

In practice we map the process to the step, ask how each step's output can be checked, and set autonomy to match. It is the first sort we run in every engagement, and the one the VirtuBrain X-ray makes visible for any of its 142 processes.

X-ray the close in the app

The arithmetic of shared services has always been seats. More invoices meant more analysts. A new country meant a new team. Quarter-end meant overtime. Pricing followed the same logic, as a rate per full-time equivalent or per transaction, worked out from what a seat can process in a day.

Agents don't fit that arithmetic. Their capacity is bought in compute, and it stretches. The agent that clears a quiet Tuesday's reconciliations can clear quarter-end's as well, without a hiring plan. Three things a service center has always planned in seats start to move.

Pricing. A charge per seat stops describing the cost of the work. Prices need to follow the work done and the checks passed, with people priced for the judgment they add.

Peak planning. Peaks used to be met with overtime and temporary staff. When agents run the checkable steps, a peak is mostly a compute question. The human constraint moves to review, meaning how many approvals and sign-offs your people can do well in the time available.

Service levels. When checkable steps run around the clock, their turnaround depends on systems and data. Steps that wait for a person still depend on people. Service levels should separate the two and promise each one honestly.

Leaders who keep planning in seats will under-use what they have already paid for.

They will also over-promise on the steps that still need a person. The fix starts from the same map as everything else, with every step sorted by who should run it, so capacity can be planned for each kind of work on its own terms.

The month-end close became a batch because checking used to be expensive. Reconciling an account meant pulling statements, ticking items and chasing differences by hand, so it made sense to do it once, at the end, when everything had landed.

Agents change the cost of checking. A bank reconciliation, a subledger tie-out or an intercompany match can run every day, each result compared with the record and logged. Differences surface on the day they happen, while the person who can explain them still remembers why.

Run that way, most of the close is finished before the period ends. The last days are left for judgment on estimates and provisions, explaining what moved, and sign-off.

Cut-off, review and certification still happen on dates the business and its auditors rely on. By then most of the work is done, with the evidence attached. Every daily check leaves a record, so the review at period end starts from a file that is mostly tested.

That is why close engagements start with the zone map. The steps an agent can run daily are the ones whose output can be checked against a record, and those move first. The judgment stays where it is, with more time to do it well.

Agentic Close Assurance