There is a specific claim circulating in finance operations this year that deserves scrutiny: that the transactional end of the function can now run itself, with people involved only at the exceptions. Invoices post themselves, payments match themselves, reconciliations resolve themselves, and the finance team moves up the value chain. The technology to do a meaningful portion of this exists now. What most organisations have not noticed is that autonomy changes what the finance team is for, rather than simply reducing how many of them you need.
Automating a task removes the work. Automating a judgement moves the work to whoever has to decide whether the judgement was sound — and that person needs to be more senior, not less
The organisations getting this right are hiring differently, not smaller.
What is genuinely automatable now
Be concrete, because the marketing is not. Invoice capture and coding. Extraction from documents is reliable, and coding against historical patterns is good enough that the exception rate on routine categories is low. This is the strongest case. Three-way matching. Where a purchase order and a receipt exist, matching is close to solved. Where they do not — services, ad hoc spend — it is not, and that is where your volume of pain actually sits. Bank reconciliation. Statement-to-ledger matching including partial payments and grouped settlements is now genuinely strong. Recurring journals and accruals. Preparation is automatable. Review is not. Collections correspondence. Drafting and sending routine chasers, escalating on a defined ladder, is effective and unglamorous. What is not automatable, despite the claims: anything requiring knowledge that exists only in a conversation, anything where the correct answer depends on a commercial relationship, and anything where being wrong is expensive and discovery is delayed.
The exception desk becomes the whole job
When ninety per cent of transactions flow through untouched, your team spends its day on the residual ten per cent — and that residual is, by definition, the hardest ten per cent. A clerk who processed a hundred invoices of which eight were awkward now handles only awkward ones. This has three consequences organisations consistently fail to plan for. The skill requirement rises. The person resolving exceptions needs to understand why the system was confused, which is a different and harder competence than data entry. The work gets less pleasant. A day of nothing but unresolved problems is harder than a day with a rhythm to it, and attrition in exception roles runs higher than in the roles they replaced. Training pathways disappear. The routine volume was how juniors learned what a normal transaction looks like. Remove it and you cannot develop the people who will staff the exception desk in three years.
Measure the right things
Straight-through rate is the headline number and it is easy to game by widening tolerances. Watch four others: the exception resolution time, the rate at which automated postings are later corrected, the proportion of exceptions that are the same recurring issue, and how long a new joiner takes to become useful. The last one is the leading indicator of whether your operating model survives staff turnover.
| Measure | Question it answers |
|---|---|
| Exception resolution time | Can people clear the harder residual work? |
| Later posting corrections | Are automated results needing rework? |
| Recurring exception share | Is a source issue being repeatedly handled? |
| New-joiner learning time | Does the training pathway still work? |
| Tolerance changes | Has apparent throughput improved by relaxing controls? |
Qualitative summary of this article's source text, not a measured outcome or performance estimate.
Practical Guidance for Autonomous Operations Assessment
- Separate automatable tasks from automatable judgements explicitly.
- Design the exception desk first, not as an afterthought.
- Raise the skill profile of the roles that remain.
- Track post-hoc correction rate, not just straight-through rate.
- Fix recurring exceptions at the source rather than resolving them repeatedly.
- Preserve a training pathway even though the routine volume is gone.
- Keep tolerance thresholds under change control.
- Plan for higher attrition in exception-only roles.
The Regional Angle
The first structural obstacle here is that a large share of regional transaction volume never had a purchase order. Purchasing happens by telephone, by messaging application, or through a relationship where the arrangement is understood and the paperwork follows — sometimes weeks later, sometimes not at all. Automated matching has nothing to match against, so every one of those invoices becomes an exception, and organisations that measure a straight-through rate of thirty per cent usually have a procurement discipline problem rather than a technology problem. The sequence that works is to establish purchase order coverage for the categories where it is achievable, accept that a negotiated tail will remain manual, and set the automation target against the addressable portion rather than total volume. The second concerns payment instruments, which are not automatable in the way the vendor material assumes. Post-dated cheques remain in active use across the Gulf for supplier settlement and rent, letters of credit govern a great deal of import trade, and the reconciliation of a cheque presented four months after issue against an invoice paid in instalments is a genuinely difficult matching problem that most products handle poorly. Similarly, bank statement formats and the availability of clean electronic feeds vary considerably between regional banks, and a reconciliation engine fed by a manually downloaded file is only partly autonomous. Check the bank connectivity before believing the reconciliation claim. The third is about the people question, which has a particular shape in this region. Transactional finance in Gulf shared service operations is staffed substantially by expatriate employees on employment-linked residency, which means the consequences of automating a role are immediate and personal in a way they are not in markets with portable employment. Two things follow practically. Redeployment is more valuable here than elsewhere, because the institutional knowledge of which supplier does what is concentrated in exactly the people whose roles are changing, and losing it makes the exception desk worse. And the training pathway problem is sharper, because you cannot easily recruit a mid-level exception handler locally if you have stopped producing them internally. Plan the transition of specific named people, not headcount.
The objection worth taking seriously
The strongest objection is that this is a familiar cycle dressed in new language. Optical character recognition, robotic process automation and rules-based matching were each sold as the end of transactional finance, and each delivered a real but partial improvement while the headcount stayed broadly flat because the exceptions and the reconciliation of the automation itself absorbed the saving. Expecting a different outcome because the pattern matching is now statistical rather than deterministic is the same optimism with better demonstrations, and finance directors who lived through the last two waves are entitled to be sceptical. That scepticism is well earned, and the prediction that headcount will not fall as much as promised is probably correct again. Where this wave differs is in what it does with unstructured input. Rules-based automation required the input to conform — a fixed invoice layout, a defined file format, a populated purchase order field — and the reason it plateaued is that the residual volume was residual precisely because it did not conform. Model-based extraction handles the non-conforming case tolerably, which moves the boundary rather than removing it. That is a genuine expansion of scope, and it matters most in exactly the environments the earlier waves failed in: messy documents, inconsistent naming, missing references. The honest expectation is not an empty finance floor. It is a materially higher share of automated volume, a harder exception population, and a team that looks more like analysts than processors.
Common Questions
What straight-through rate should we expect?
It depends almost entirely on purchase order coverage and document consistency, not on the software. Measure your addressable volume first; a rate quoted without that context is meaningless.
Should we automate before or after fixing our processes?
Fix procurement discipline first for the categories where it is feasible. Automating an undisciplined process produces an exception queue and a disappointed sponsor.
Who reviews automated postings?
Someone senior enough to know when a pattern is wrong, on a sample basis, with the sample weighted toward unusual categories rather than drawn evenly.
What should we expect over the next twelve months?
Expect the vendors to push from assistance into execution and to price it as a separate module. Expect the first serious reporting on exception-desk attrition. Expect the constraint to be revealed as process discipline rather than model capability. And expect the organisations that redeployed their experienced staff to be measurably ahead of those that reduced them.
Autonomous Operations Assessment — we size what is genuinely automatable in your volume, then design the exception desk that has to absorb the rest.