This month we are writing down the sorting method itself, because choosing which processes deserve a rebuild decides more about how a modernisation turns out than any of the engineering that follows it. An estate assessment produces a ranked list, and the ranking matters more than the inventory. Most estates contain a handful of processes worth rebuilding properly, a larger group that should keep running exactly as they are, and usually two or three that nobody has used in a year. Sorting them takes six questions per process, and the order below is the order we put them in.
1. Does every system it touches expose an API
This is the first question because it constrains every other answer. For each application the process reads from or writes to, we check whether there is a documented programmatic interface, whether your licence includes it, and what the authentication model is. A finance system with a supported REST API and a spreadsheet on a shared drive are both straightforward. A vendor portal with no API and no export is the case that forces a different approach. We check this against the vendor documentation and your licence terms rather than against what the team believes is available, because the belief and the entitlement disagree often enough to matter.
2. How often does it run, and what does a run cost
Volume changes the economics of every option. A process that runs four hundred times a day and a process that runs at month end are different problems even when the steps look identical. For screen operation in particular, cost per run matters, because driving an interface with a model is more expensive per execution than calling an endpoint. Low volume can make an approach viable that would be indefensible at scale, and high volume can justify engineering work that a quarterly process never would. This is the number we ask you to measure rather than estimate, since it decides which of the options below you can afford.
3. What happens when it gets one wrong
We ask what the process writes and who sees the result. A bot that drafts an internal summary and a bot that posts to the general ledger both fail occasionally. One creates confusion and the other creates a correction, an audit note and a conversation with your accountant. This answer sets how much validation the rebuild needs and where a person has to approve before anything leaves. Our bias is to spend the budget on the processes that write into a system of record ahead of the ones that produce reading material, and we will argue for that ordering if your own priorities run the other way.
4. How often does the interface underneath it change
Vendor release cadence is a decent proxy. A cloud application that ships changes every few weeks will break selector-based automation regularly. An on-premise system that has not been upgraded since 2019 will not. Maintenance history tells the same story: if the process has been repaired six times in two years, the interface is moving and the coupling is the problem. Where nobody can tell us the release cadence, the maintenance tickets answer the question well enough.
5. How much of it is judgement
We separate the deterministic steps from the interpretive ones. Matching an invoice to a purchase order on reference and amount is deterministic. Deciding whether a delivery note in a photograph corresponds to that purchase order is interpretive. Processes that are entirely deterministic are the cheapest to rebuild and the most reliable afterwards. Processes carrying real judgement are the ones where a language model earns its place, and they need explicit approval points. This is the question we think the market is loosest with at the moment, because a good deal of what is being sold as agent work is a set of database updates with a chat window in front of it.
6. Does anyone still use the output
Nothing in an estate switches itself off, so this one has to be asked deliberately, and it often ends in a recommendation to do exactly that. Reports that went to a manager who left, reconciliations superseded by a system upgrade, and feeds into a spreadsheet nobody opens all keep running because nothing tells them to stop. Checking who consumes the output is faster than modernising something that should be switched off.
What the four verdicts mean
Six answers give us one of four verdicts per process. We would be sceptical of any assessment that comes back recommending a rebuild of everything, because that result means the sorting was not done.
Rebuild. The process moves onto system APIs with Python carrying the deterministic steps, a language model used only where a step needs reading, classifying or interpreting, and a test suite that runs without a screen. It runs on the Hub with each execution logged and outbound work held for approval by default. This is where the processes that are high volume, high consequence, or sitting on a moving interface end up.
Replace with screen operation. Where a system exposes nothing programmatic, AI screen operation can take the place of the existing bot. We size this on volume, reliability requirements and cost per run before recommending it, and we say plainly when the numbers do not support it.
How we decide whether an AI agent can take over an RPA bot
Leave alone. A process that runs reliably against a stable interface, at low volume, with low consequence when it fails, is working. Rebuilding it costs money and leaves the process behaving as it already does. Processes regularly get this verdict, and recording it is part of the assessment.
Retire. Nobody uses the output. The recommendation is to switch it off, document that it existed, and reclaim the licence and the maintenance attention.
Why the order matters more than the list
An inventory tells you what you have. A ranked assessment tells you where to spend the first engagement, which is the only decision actually in front of you at the start. We rank by consequence and fragility rather than by effort, which puts the processes that can write bad data into a system of record at the top, and those are the ones worth engineering properly first. The ordering is the part of this we would keep even if you never engaged anybody to do the work.
Back to the blog