Automating routine work with AI: six bookkeeping tasks and the line where AI stops helping
In the summer of 2026 a bookkeeping outsourcing company came to us wanting to take routine work off their people. The client had already split the work into six blocks and marked which ones they considered simple and which hard.
The project reached the estimation stage and never became development, so this is not a case study. It is a breakdown of the brief, and that brief shows clearly where AI works and where it only looks like it works.
The six blocks
Moving sums between spreadsheets. Issuing invoices based on emails and chat messages, with a trip into the accounting system. Answering common questions. Preparing payment orders. Producing shipping documents. Checking tax calculations before filing.
The list is useful because it contains two different kinds of work that clients perceive as one and that a vendor is obliged to separate.
First kind: moving and preparing
Sums, invoices, payment orders, shipping documents. The work repeats, the data is machine readable, the result is verifiable: the document is either correct or it is not.
The difficulty here is not the model, it is access. Can the system read the mailbox and the chats. Does the accounting system allow writes or only reads. Who inside the company is authorised to issue those credentials. Until those questions are answered, discussing the interface is premature, which is why we start such projects with an access check rather than a design.
Second kind: judgement and accountability
Checking tax calculations before filing looks like the same category and behaves differently. The system can find a discrepancy, it cannot take responsibility for a signature on a filing.
That task also had a direct competitor: the checking mechanism built into the accounting system itself, which the client considered weak. That detail matters for the estimate, because you are not competing with manual labour, you are competing with an existing tool, and the difference has to be visible or the project will not pay for itself.
The sentence that removes the fear
The main takeaway from preparing that conversation was not technical. When a non technical client hears the architecture explained in jargon, they do not relax, they become cautious.
The formulation that works: the system prepares a draft, a human makes the decision. It is honest, and it removes the biggest fear, that a machine will start filing documents with the tax authority on its own.
It also sets an interface requirement: a person must be able to see what the suggestion was based on and correct it in one step. Automation where the draft cannot be fixed quickly turns into extra work.
Where automation hits something other than technology
We met a similar boundary in another project, a system for a college. Automating intake and processing was straightforward. Anything involving judgement was not: changing a status, recording a reason for refusal, assessing a disputed case all require a person, and the bottleneck is not data entry speed but the discipline of keeping data current.
The rule that came out of it: if a process today runs irregularly and depends on one employee’s memory, automation will not save it. First the process has to exist, then it can be made faster.
How we propose to do it
One process at a time, with a measurable frequency: how often it repeats per week and how long it takes now.
An access check before development: a small paid step that answers whether data can be read and a document produced inside the client’s real system.
A success criterion written down in advance, for example the share of documents a person accepts without edits. Without that number, two months later nobody can tell whether things improved or merely changed.
And an honest conversation about who is accountable for a mistake. That is not legal formality, it is a design requirement: the answer determines how many manual confirmations the system needs and what the audit log looks like.
FAQ
Which tasks automate most easily
Repeatable operations on machine readable data with a verifiable result: moving data between systems, producing standard documents, answering common questions from a known base. Where the output can be checked automatically, rollout is fastest.
Where does AI not help
Where judgement and accountability are required: a decision to refuse, assessing a disputed case, choosing between options on incomplete data. A system can prepare the option, the confirmation stays with a person.
Where should a company start
With one process that has a measurable frequency and a clear success criterion. Test data access first and discuss the interface later. Broad rollout across every process usually ends as a pilot nobody uses.
How long does such a project take
For one process it is usually weeks rather than months, provided access to the data exists. The delay is almost never development, it is obtaining credentials and agreeing who owns the result.