Choosing the first workflow to automate when every team has a different idea

Why every team proposes something different

Each department proposes the work that hurts it most, which is reasonable and rarely a good basis for the first project. The most visible pain often sits in a process with poor data, unclear outcomes or a legal review in front of it. A first project has a second job besides its own result: it sets up the test set, the system access and the review routine that the next projects reuse. That only happens if it finishes.

Six questions to score each candidate

1. 1

   Volume

   How many cases arrive per month, and how much skilled work does each take? Multiply the volume by the measured benefit per case and compare it with the build cost, the running cost and the review effort. A small volume of expensive cases can justify a project, and a large volume of trivial ones may not.

2. 2

   Variation

   Do cases differ in ways a rule cannot capture, such as free text, mixed documents or missing fields? If a rule works, write the rule. If every case needs deep judgement, an agent will pass most of them back to a person.

3. 3

   Access

   Can a service account read the inputs and write the results through an existing API today? A candidate that first needs a new integration, screen scraping or a data protection review has a project in front of the project.

4. 4

   Review effort

   How long does it take a person to check an output compared with doing the work? Checking two amounts against a purchase order takes seconds. Checking a contract summary means reading the contract.

5. 5

   Cost of a mistake

   Can a wrong result be caught and reversed before it does harm? An internal routing error costs a delay, while a wrong payment or a wrong customer commitment costs more.

6. 6

   Known outcomes

   Are there past cases where the correct result was recorded? Without them you cannot build a test set, and without a test set nobody can say whether the system works.

These questions mirror the scoping phase in the AWS Generative AI Lens, which asks whether generative AI is relevant to the problem at all, how success will be measured, whether the project is technically and organisationally feasible and whether the data it needs exists in usable quality.

A worked comparison of three proposals

The following example is hypothetical. An operations director receives three proposals: customer service wants drafted replies to every incoming email, finance wants supplier invoices matched to purchase orders, and legal wants incoming contracts screened for non-standard clauses.

• Invoice matching scores well. Volume is high, cases vary through scans, partial deliveries and credit notes, the ERP has an API, a clerk checks a proposed match in seconds, and years of booked invoices supply known outcomes. A wrong match can be stopped before payment.
• Email replies score unevenly. Volume is high and the mail system is reachable, but a good reply is a matter of judgement, past replies were never graded, and checking a complex reply takes almost as long as writing it. A narrower slice, such as classifying and routing incoming mail, scores much better.
• Contract screening scores poorly as a first project. Volume is low, a missed clause is expensive, and the lawyer reads the whole contract anyway, so the review effort barely changes.

Invoice matching goes first. Email routing is a reasonable second candidate once the first project has set up the test and review routine. Contract screening waits until there is a reason to expect the review effort to drop.

Check whether a rule would do the job

Before scoring a candidate as an AI project, take fifty past cases and try to write rules for them. Google's Rules of Machine Learning open with the advice not to be afraid to launch without machine learning, because a simple heuristic often gets you much of the way. Anthropic's 2024 article on agents makes a similar distinction: fixed workflows suit well-defined tasks, and agents suit open-ended problems where the steps cannot be predicted. The NIST AI Risk Management Framework also asks that viable non-AI alternatives be considered.

If rules cover most of the fifty cases, build the rules and send the remainder to people or, later, to a model. That is a successful outcome, and it usually costs less to run and explain.

Cut the winner down to a first slice

Automating invoices is a programme. Matching single-currency invoices that reference a purchase order, for the forty suppliers with the most volume, is a first project. NIST's framework asks for the targeted application scope to be specified and documented, and a short scope statement does that: which cases are in, which are out, what the system produces, who reviews it and where it hands over.

Publishing the scores to the teams

Share the scores with every team that proposed something, along with what would change them. A candidate that lost on access can return once the API exists, and one that lost on known outcomes can start recording them now. If you would like a second opinion on a shortlist, our AI agents work begins by mapping one process and measuring what a case costs today.

AI agents and process automationThe data and access a project needs before it starts