Every business has a handful of processes that keep it running: closing the month, onboarding a client, paying suppliers, reporting to investors, preparing the board pack. They are rarely written down in full. They live in spreadsheets, shared inboxes and the heads of the people who do them.
They also take more time than anyone would choose to spend. The systems involved are usually good ones. The time goes on the gaps between them: someone downloads a report from one system, re-keys part of it into another, checks it against a PDF that arrived by email, chases a colleague for a missing figure and then formats the result for somebody else.
Two things have changed in the last few years that make those gaps worth closing. AI can now read and write the unstructured material that sits between systems. And the cost of building software around a specific workflow has fallen sharply. Together they mean a business can automate the process it already has, without replacing the systems it already owns.
Why standard software leaves a gap
Accounting packages, CRMs, fund administration platforms and ERP systems are built for the common case. They have to be, because they are sold to thousands of customers. Each one handles its own part of a process well.
No vendor knows how your business joins those parts together. Your approval rules, your client’s preferred reporting format, the way your lender wants covenant figures presented, the exceptions your team has learnt to handle: none of that is in the box. So people become the connection between systems. The spreadsheet that reconciles two of them, the checklist in someone’s notebook and the monthly email asking for the same five numbers are all software that was never built.
For years, the choices were to live with it, to buy another platform and reshape the business around it, or to commission a bespoke system at a cost only large firms could justify.
What has changed
AI can handle the unstructured part
Traditional automation needed clean, structured data. It could move a number from one field to another, but it could not open a supplier invoice in an unfamiliar layout, read a loan agreement or make sense of a paragraph in an email.
Language models can. They can extract figures from a PDF, classify a document, compare a contract clause with a policy, draft a commentary from a set of numbers and summarise a long thread into the three points that matter. The work that used to require a person simply because it involved reading can now be done by software, with a person checking the result.
Building custom software costs far less
The same technology has changed how software is written. A small team working with AI coding tools can now deliver in weeks what used to take months. Cloud services provide the hosting, security and storage that once had to be built from scratch. Most business systems now offer a way to read and write their data directly.
The practical effect is that a tool built for one process in one firm is no longer a luxury. It can cost less than a year’s licence for a platform that does eighty per cent of the job and leaves the awkward twenty per cent to your team.
What a customised automation looks like
The pattern is similar whatever the process. A good automated workflow does six things:
- Starts by itself. A document arrives, a date is reached or a balance changes.
- Collects what it needs. It pulls data from the accounting system, the CRM, the bank feed or the shared drive.
- Reads and extracts. It takes the figures and terms out of invoices, statements, contracts and emails.
- Checks. It applies your rules: does this match the purchase order, does this reconcile, is this within the limit?
- Drafts the output. The journal, the report, the client letter, the summary for approval.
- Stops for a person. Someone reviews what was done, sees the exceptions and signs off. Every step is recorded.
Some examples from the work we see most often:
- Month-end reconciliations. Bank, intercompany and supplier statements are matched automatically. The team sees only the items that do not agree, with the likely reason attached.
- Client and investor onboarding. Identity documents are read, checked against the requirements and filed. Missing items are chased without anyone writing an email.
- Supplier invoices. Invoices are read on arrival, coded, matched to orders and queued for approval.
- Management and investor reporting. Numbers flow from the ledger into the pack. A first draft of the commentary is written from the variances and reviewed by the person who owns it.
- Compliance deadlines. Filing dates, covenant tests and renewals are tracked in one place, and the supporting pack is prepared ahead of the date.
In each case the process is the one the firm already runs. The steps are the same. The difference is who performs them.
Where the productivity gain comes from
It is tempting to measure automation by hours saved on typing. That is the smallest part of the benefit.
- Waiting disappears. Much of the elapsed time in a process is work sitting in a queue until someone is free. Software does not wait. A month-end close that runs for ten working days rarely contains ten days of work.
- Rework falls. Re-keying produces errors, and errors are found late. When data moves directly and every item is checked the same way, fewer things come back.
- Checks become consistent. A person at the end of a long day checks a sample. Software checks every line, every time, by the same rule.
- Knowledge stops living in one head. Once a process runs in software it is documented by definition. Holidays and resignations stop being operational risks.
- People move to the work that needs them. Reviewing exceptions, speaking to clients and deciding what to do about the numbers are better uses of a skilled person than assembling them.
The last point matters most. The aim is to grow without adding the same proportion of manual work, and to give experienced people back the part of the job they are good at.
Buy or build?
Custom software is not the answer to everything. A simple test helps.
Buy when the process is the same in every business. General ledger, payroll and email are solved problems. Building your own would be a waste.
Build around what you bought when the process is specific to you, crosses several systems or depends on documents and judgement. This is where standard products stop and manual work begins.
Custom automation should sit on top of your existing systems and connect them. Replacing them is rarely necessary. The ledger stays the ledger. The new tool handles the work that happens before, between and after.
How to do it well
A few principles separate the projects that deliver from the ones that stall.
Start with one process. Choose something that happens often, follows rules most of the time, is disliked by the people who do it and can be measured. Monthly reconciliations and document-heavy onboarding are common first choices.
Map it as it really runs. The written procedure and the real one are rarely the same. Sit with the person who does the work and record every step, including the exceptions. The exceptions are where most of the design effort goes.
Fix before you automate. If a step exists only because of a problem elsewhere, remove the problem. Automating a poor process makes it run faster and leaves it poor.
Keep a person in control. Software prepares and people approve. Every output should show where each figure came from, so that the reviewer can check it in seconds.
Deliver in weeks. A first version that handles the main path and flags everything else is worth more than a perfect system that arrives next year. Extend it once it has proved itself.
Measure before and after. Record how long the process takes today, how many errors come back and how many people it involves. Without a baseline, nobody can say whether the investment worked.
Own what is built. Insist on documentation, access to the code and a clear plan for support. Client and financial data must stay inside an environment you control, with access limited to those who need it.
Where to begin
Ask your team one question: which task do you repeat every week or month that you would happily never do again? The answers usually come quickly, and they are usually the same three or four processes.
Pick one. Measure it. Automate it with a person signing off. Then use what you learnt on the next one.
If you would like help choosing the right process or building the workflow around it, get in touch.
Start with one conversation.
Tell us what you are trying to fund, fix, value or automate. In a 30-minute call we'll tell you honestly what we can do, roughly what it involves, and what a first step would look like.