The CFO's Blind Spot: Why Tax Provision Still Runs on Manual Overrides
The quarterly tax provision is the most consequential financial calculation that most CFOs do not fully understand. It is also the one most likely to be adjusted manually at the last minute. Both facts are connected.
In every quarterly close I have been part of — across Big 4 clients, Fortune 100 corporates, and FTSE 250 multinationals — the tax provision follows the same pattern. The tax engine runs. The ETR comes out. Someone frowns. Then the manual overrides begin.
The overrides are not errors. They are corrections for things the tax engine cannot model: the pending consolidation journal that is "not a topside, but a late adjustment" the R&D credit study that will not be finalised until next month, the transfer pricing adjustment that the TP team communicated verbally but has not yet journalised. These are legitimate adjustments. The problem is that they live in the tax manager's head, not in the system.
The Provision Is a Model, Not a Calculation
Most finance leaders treat the tax provision as a calculation: inputs go in, the ETR comes out. In reality, the provision is a probabilistic model. The current tax expense is reasonably deterministic, it follows the tax return logic. But the deferred tax components involve predictions: will this temporary difference reverse in a period when the enacted rate is 25% or 28%? Will the net operating loss be utilised before it expires? Will the valuation allowance need to increase? And all this gets compounded with the introduction of Pillar Two
These predictions are where manual overrides concentrate. And because they involve judgment, they are rarely documented with the rigour that the auditors would prefer. The tax director makes a call, enters an adjustment, and the ETR moves 30 basis points. The audit trail says "management estimate", which tells the auditor nothing about the reasoning.
Why Automation Has Stalled
Tax technology vendors, Oracle, Thomson Reuters, Longview, have invested heavily in automating the mechanical parts of the provision: data collection, entity-level calculations, consolidation, and reporting. These work well.
But none of them have cracked the judgment layer. The manual override problem persists because the overrides encode business context that the system does not have. A pending audit settlement requires knowing the litigation strategy. A transfer pricing adjustment requires knowing the TP team's preliminary findings. A valuation allowance change requires knowing the company's five-year revenue forecast.
This is not a technology limitation. It is a data integration limitation. The information exists — in emails, in meeting notes, in the TP team's working papers, in the FP&A team's forecast models. It is just not connected to the tax provision system.
The AI Opportunity
This is where large language models become genuinely useful in tax — not as calculators, but as context aggregators. An LLM that has access to the TP team's draft report, the legal team's audit settlement memo, and the FP&A team's latest forecast can draft the narrative justification for each manual override before the tax director even opens the provision workpaper.
The override still requires human judgment. But the LLM can pre-populate the adjustment with the supporting context, flag when an override contradicts a data source, and ensure that the audit trail contains more than "management estimate."
This is not speculative. I am building this capability into real workflows today. The initial results are promising: the time spent on override documentation has dropped by 60%, and the audit queries on provision adjustments have dropped by half — because the documentation is substantive from the start.
What CFOs Should Ask
If you are a CFO, ask your tax director one question: "How many manual overrides were in last quarter's provision, and where is the documentation for each one?" If the answer involves a long pause, you have found your blind spot.