Most companies we talk to have some version of the same line in this year's plan: we have to do something with AI. The line rarely says what the AI is supposed to do. A tool gets bought, a pilot runs for a few months, and at the end the work is still done the way it was before. In our experience the tool is usually not the problem. The problem is that nobody asked what kind of work the process consists of before choosing how to automate it.
We sort work into three categories before we build anything. The categories apply equally to finance, operations, sales support and customer service, and most of the value of an automation project comes from sorting correctly.
1. Work with a known right answer
A large share of office work has exactly one correct result. Matching supplier invoices to purchase orders is an example, and so are assembling the monthly management pack from this month's numbers, checking that a new customer file contains every document compliance requires, copying an order from an email into the ERP system, and producing the same client report every period from new data. If two competent people did any of these tasks independently, they should arrive at the same result, and if they did not, one of them made a mistake.
Work of this kind belongs in a pipeline made of rules, code and templates. A pipeline produces the same output from the same input every time, and anyone can open it and see why a particular value ended up where it did.
This work should not be given to a language model at run time. Consider a team that processes 2,000 invoices a month with a model that is right 99% of the time. Twenty invoices will be wrong every month, and they will be scattered at random among the correct ones, so someone has to check all 2,000 to find them. The automation has turned into a review job. A rule that either finds a match or flags the invoice for a person does not create that problem.
2. Work that needs a person
Other work has no single right answer. Someone has to decide whether a mismatch is a data error or a genuine dispute with the supplier. Someone has to explain to the board why the quarter came in under plan, reply to an unhappy customer, or judge whether an unusual order should be accepted at all. These tasks require judgement, and a named person is accountable for the result.
A language model is useful here as an assistant. It can prepare a first draft, summarise a long thread, or read a 40-page contract and point out the clauses that matter. We apply one rule to this category: when a model drafts something, a person reviews it and signs it off. Responsibility for the result stays with that person.
3. Work to leave alone
Some processes should not be automated at all, because they are rare, cheap to do by hand, or about to change. The yearly insurance renewal is one example. A one-off data migration is another, and so is any process that will disappear when a new system arrives next spring. Automating them would cost more than it saves.
For that reason every assessment we deliver includes a list of processes we recommend leaving as they are. Knowing where not to spend money is as useful to a company as knowing where to spend it.
Where AI fits in the first category
We do use AI for work that has a known right answer, but we use it to build the pipeline rather than to run it. A model is good at reading messy inputs, proposing how the fields in one system map to the fields in another, drafting transformation rules and writing a first version of the code. A person reviews that work once. After that the pipeline runs as ordinary deterministic software with no model involved, so the same input always produces the same output.
We built our own product this way. Octigen Reporting uses AI to set up a recurring report: it reads the PowerPoint template, proposes how the data maps to each chart and table, and drafts the pipeline. Every later run of that report is deterministic. We started with reporting because both of us spent years doing it, but the approach applies to any process in the first category.
How to try this yourself
You can apply the same sorting to one of your own processes in an afternoon. Pick a process that recurs, write down each step, and ask three questions about every step:
- If two competent people did this step independently, should they get the same result? If so, the step belongs in a pipeline.
- Would you want a named person to answer for the result? If so, the step stays with a person, who may use a model to prepare a draft.
- Does the step happen only a few times a year, or differently every time? If so, leave it alone.
In most processes we look at, the majority of steps fall into the first category and a small number fall into the second. The people who run the process usually spend most of their week on the first kind of step, which is the kind a pipeline could do for them.
This sorting is the core of our Assessment. Over 20 hours we map how the work moves through your organisation, place each step in one of the three categories, and deliver a written report that says what to build, what to keep with your people and what to leave alone. If you have a process that costs your team more time than it should, tell us about it.
