
The most expensive mistake in an automation programme happens before anyone writes a bot: picking the wrong processes. A good shortlist compounds. A bad one gives you a year of brittle bots and a disappointed steering committee.
The trouble is that most scoring models in circulation were designed around classic RPA. They heavily penalise variable inputs and high exception rates, because in 2019 a bot could not cope with either. Large language models have changed that, and the scoring needs to catch up.
This post lays out an updated model and a tiny script to run it. It is adapted from the process-selection section of TechCirkle's buyer's guide to robotic process automation consulting services.
Six criteria, honestly measured
Score each candidate 1 to 5 on:
-
Volume: at least ~500 transactions or ~80 human hours a month to be worth it.
-
Stability: no planned system or policy change in the next 12 months.
-
Measurability: you can count inputs, outputs and current handling time.
-
Exception profile: you know roughly how many cases are exceptions and why.
-
Risk class: a wrong result is cheap and reversible, or there is a human checkpoint downstream.
-
Ownership: a named business owner who will sign off designs and own the bot after go-live.
Then weight by annual hours at stake. That is the classic part.
What LLMs change in the weights
Input variability stops being a veto. Reliable extraction from semi-structured documents means "the forms all look different" is no longer disqualifying.
Exception rate becomes an opportunity signal. If exceptions are mostly reading or matching problems, an LLM resolver can absorb many of them. A 30% exception rate with the right type of exceptions can now be a strong candidate.
Risk class gains weight. AI steps fail quietly and confidently. A process where errors are expensive or irreversible should score lower for AI-assisted automation even if it scores well otherwise.

A small scoring script
const WEIGHTS = { volume: 1.0, stability: 1.0, measurability: 0.8,
exceptions: 0.8, risk: 1.4, ownership: 1.2 };
function score(c) {
// Hard gates: reject before scoring
if (c.monthlyHours < 80 && c.monthlyVolume < 500) return { reject: "low volume" };
if (c.plannedChangeWithin12m) return { reject: "system or policy changing" };
if (!c.owner) return { reject: "no business owner" };
// LLM-era adjustment: readable exceptions count as upside, not downside
const exc = c.exceptionType === "reading_or_matching"
? Math.min(5, c.exceptionScore + 2)
: c.exceptionScore;
const s = { ...c.scores, exceptions: exc };
const weighted = Object.entries(WEIGHTS)
.reduce((sum, [k, w]) => sum + w * s[k], 0);
return { value: weighted * c.monthlyHours * 12, tier: c.hasApi ? "integrate" : "bot" };
}
Notice the tier field. If the target system has an API, the recommendation is an integration, not a bot. Screen automation over a usable API is technical debt from day one.
Measure the baseline before you trust the score
Scores are only as good as the inputs. Before ranking anything, spend a week collecting real numbers: transactions per month from system logs, a timed sample of twenty or thirty cases per process, and a rough tally of exception reasons. Teams that skip this step routinely overestimate volume and underestimate exception handling time, which inflates exactly the processes that disappoint later. A baseline also gives you the "before" figure you will need to prove savings after go-live.
Expect to reject half the list
In practice, applying hard gates and honest scores removes roughly half of the initial candidate list. That is the model working. Common rejections: processes due to change with an ERP migration, low-volume tasks better fixed with a form, and processes with no one willing to own them.
A good consultancy will reject candidates too. When evaluating partners, ask what share of their discovery recommendations ended up as something other than RPA. Somewhere between 20% and 40% suggests they are recommending on merit rather than selling bots.
Turning scores into a first wave
Take the top two or three. Ideally pick one classic deterministic process and one with an LLM-assisted exception step, so your team learns both patterns in the first 90 days. Measure the baseline before you build, because without it you cannot prove the saving.
If you want help running this assessment on your own processes, TechCirkle's AI development services team runs discovery engagements that end in a ranked, costed pipeline.
Frequently Asked Questions
What makes a process a good candidate for RPA?
High volume, stable systems, measurable baselines, a known exception profile, low risk from errors and a committed business owner. If an API exists, integrate instead of using a bot.
Should high exception rates disqualify a process?
Not anymore. If exceptions are mostly about reading or matching data, AI-assisted exception handling can resolve many of them, making the process worth reconsidering.
How many processes should be in the first automation wave?
Two or three is typical. Choosing one deterministic bot and one AI-assisted process lets the team learn both patterns quickly.
What volume threshold justifies automation?
A common rule of thumb is around 500 transactions or 80 human hours per month. Below that, build, licence and maintenance costs rarely pay back.
How do I know if a consultant's scoring model is current?
Ask whether it accounts for AI-assisted extraction and exception handling. Models unchanged since around 2021 tend to undervalue document-heavy processes.