The main question: where exactly intelligence is needed
Break the process down into steps and ask of each one: does it require understanding meaning, or only following rules? Moving data from one system to another, checking against a list and calculating by a formula follow rules. Working out what an email is about, extracting the substance of a document and matching a customer’s wording to the catalogue can only be done through meaning.
AI is needed exactly at the steps that involve meaning. Everything else is cheaper, faster and more reliable in ordinary code. That is how working systems are built: a deterministic frame, with AI at the nodes that cannot do without it.
Which processes go first
A good first process has five traits:
- It is frequent: hundreds of operations a month, not a dozen.
- It has one goal but varied forms of input. This is exactly where rules stop coping.
- Its result can be checked.
- A mistake is reversible or caught by a check. Processes with irreversible consequences are not taken on at the start.
- The data for it already exists in digital form.
Typical working candidates
- triaging incoming enquiries: classification, extracting the substance, routing;
- checking documents against requirements, with the discrepancies pointed out;
- preparing drafts of replies and commercial documents from the company’s templates;
- matching data from different sources where the wording does not coincide word for word;
- a first search of the internal knowledge base in place of asking a colleague.
What not to automate with AI
- Processes with no criterion of correctness. If two experts argue about what is right, the system will not learn it.
- Rare operations. The preparation would take longer to pay back than the process itself will live.
- Strictly regulated calculations. They need reproducibility, and the flexibility of a model only gets in the way.
- Processes the company plans to change. It makes sense to automate what is stable.
- Processes whose data is not in digital form. Data collection first, AI after.
Why the process is redesigned first
Many processes are badly built for historical reasons: extra approvals, duplicate data entry, checks for a problem that disappeared long ago. Automation makes such a process fast and still bad, and it also fixes the process in code, so changing it becomes more expensive.
So the first stage is an analysis of which steps are needed at all. Often, once two unnecessary steps are removed, what remains is covered by an integration with no AI at all. For the client this is a normal result, and it saves money.
What implementation looks like, step by step
- Process analysis: steps, volumes, time, errors, who does what by hand.
- Data check: does it exist, in what form, who owns it, can it be used.
- Choosing the nodes for AI and the rules for everything else.
- A pilot on a limited part of the process, with a metric fixed before the start.
- Integration into working systems. This is usually the longest part, and the model has nothing to do with it.
- Gradually widening autonomy as quality statistics build up.
What to measure
- the share of operations completed without a person;
- the time an operation takes from start to finish;
- the share of corrections made after the system;
- the cost of one operation, including running and review;
- the number of cases where the system correctly declined and handed the work to a person.
The last metric is often underrated. A system that recognises its own limits is safer than one that always produces an answer.