Write the process down. Then let a machine do the boring half.

• The process only exists in one person’s head

  Nobody wrote it down, so quality depends entirely on who is at the desk that day. When they resign, take leave or get promoted, the close slips and nobody can say exactly why.

• Every month is rebuilt from scratch

  The same reconciliations, the same reformatting, the same copy and paste between the accounting system and the reporting pack. Highly paid people spending days on work that has no judgement in it.

• Controls exist on paper but not in the workflow

  An approval matrix in a policy document that the actual payment run does not enforce is not a control. It is a document.

• Software was bought before the process was fixed

  A new ERP on top of an undesigned process produces the same numbers, later and more expensively. This is the most costly mistake we are called in to unwind.

• AI has been tried on the wrong thing

  A pilot that summarises documents nobody reads, while three people still key invoices by hand. Enthusiasm aimed at the visible problem rather than the expensive one.

• Map what actually happens, not what the manual says

  We sit with the people doing the work and document the real process, including the workarounds. The gap between the documented process and the real one is usually where the risk lives.

• Design the SOPs

  Written procedures for the month-end close, accounts payable, accounts receivable, payroll, expense claims, fixed assets, banking and approvals. Owner, trigger, steps, evidence, and what to do when it goes wrong. Written so a new hire can follow them in week one.

• Rebuild the process before automating it

  Automating a bad process makes the bad outcome arrive faster. We fix the sequence, the ownership and the controls first, then decide what is worth automating.

• Automate the clerical layer

  Invoice capture and coding, bank reconciliation matching, recurring journal preparation, report assembly and distribution, document filing, data movement between systems. High-volume, rule-bound work where a machine is genuinely better than a person.

• Build controls into the workflow

  Approval limits enforced by the system rather than remembered by a manager. Exceptions raised automatically. An audit trail that exists because the process creates it, not because someone maintained a spreadsheet.

• Hand it over with the documentation

  You get the SOPs, the configuration and the reasoning. Not a black box that only we can maintain, which is the thing most automation vendors are quietly selling.

• Reading documents that arrive in any format

  Supplier invoices, receipts, bank statements, contracts. This is the single highest-value application in an SME finance function, because it is the highest-volume manual task and the error rate of a tired clerk at month end is worse than most people admit.

• Matching and reconciling

  Bank lines to ledger entries, purchase orders to invoices to receipts, intercompany balances between entities. Rule-based matching with a human reviewing only the exceptions.

• Drafting the first version

  Variance commentary, board pack narrative, policy documents, procedure notes. A competent first draft that a human edits is faster than a blank page, and the human stays responsible for what is said.

• Answering questions of your own data

  Letting a manager ask what happened to a cost line without waiting for the finance team to build a report, within properly scoped access.

• Judgement calls with money attached

  Provisioning, revenue recognition on a complex contract, impairment, a transfer pricing position. These need a person who will be accountable for the answer.

• Anything where a wrong answer is silent

  If a mistake would not be caught by a downstream check, do not put a probabilistic system in that seat without review. Finance has plenty of places where an error surfaces only at audit, twelve months later.

• Approvals

  A machine can route an approval, prepare it and flag the exceptions. A named human approves the payment.

• Confidential data you have not thought about

  Payroll, personal data, commercially sensitive contracts. Before anything goes near an external model there is a conversation about what may leave the building and under whose terms.

Automating a finance process is not primarily a technical problem. The hard part is knowing what the output has to survive: an auditor, a tax inspection, a lender's covenant test, a buyer's due diligence. A developer who has never closed a set of books cannot specify that, and will build exactly what you asked for rather than what you needed.

We come at it from the other side. We have run the close, defended the audit and signed the filing, and we work with the tooling directly rather than briefing someone else to. That combination is rarer than it should be, and it is the entire reason this works.

It also means we are willing to tell you when the answer is no. Plenty of processes in a small company are not worth automating, and plenty of AI pilots exist because somebody wanted to run an AI pilot. You are paying us for that judgement more than for the build.

The rest of the scope · If you are already the CFO