Lots of unstructured data
Client emails, contracts, invoices, chat comments: the information is there, but someone has to read it.
AI automation
AI automation for business processes that need analysis and decisions
Simple rules are easy to automate. It gets harder with things someone has to read and judge: a client’s email, a contract, an unusual request. We hand those processes to AI, and the decisions that need a human stay with a human.
What it is
What is AI business process automation
AI automation means using artificial intelligence for the steps of a process that used to need human attention: read and understand a document, judge a situation, pick a direction, prepare a decision. Classic rules and integrations stay at the base, and AI adds understanding.
| Classic automation | AI automation | |
|---|---|---|
| Input data | Structured: form fields, spreadsheet columns | Unstructured too: emails, PDFs, chats, free text |
| Logic | Fixed “if… then…” rules | Judging context by your criteria and examples |
| Non-standard case | Stops or gets it wrong | Weighs it up, suggests options or hands it to a person with an explanation |
| Changes in the process | The rules have to be rewritten | Changing the instructions and criteria is enough |
| Result | A completed action | A completed action plus the reasoning behind the decision in the log |
AI automation pays off most where a process is hard not because of the number of steps, but because something has to be understood and decided.
Who it suits
Signs of a process worth giving to AI
Not every process needs artificial intelligence. These are the signs we use to pick tasks for a pilot.
What we automate
Which business processes we automate with AI
Control over decisions
How AI makes decisions: three levels of independence
You set the level of independence for each step of the process separately. We usually start with the first and raise the level once the quality is stable.
-
1
Advises
Analyzes the situation, suggests a decision and explains why. A person carries it out.
-
2
Acts with approval
Prepares the whole action: a document, an order, an email. A person only confirms.
-
3
Acts alone within the rules
Acts on its own as long as the situation fits your rules. Anything outside them goes to a person.
Every AI decision is written to a log: what it saw, what it decided and why. The log shows where the AI is wrong, so the rules can be fixed without guessing.
Ground rules
What we build in so that the automation can be trusted
Examples
What it looks like on real processes
The first four cards are processes of our own team, with approximate figures. The last two are typical scenarios for client projects, without numbers.
Commercial offers
Reads the request and the client’s history, picks items and prices by the rules and builds the offer. Discounts above the limit go for approval.
2 min instead of 20 per offer
Contracts, invoices and acts
Prepares a set of documents from the deal data and checks it before sending.
7 min instead of 1 hour per set
Checking tasks against requirements
Checks finished work against the requirements and sends it back for rework before any human review.
~5 min instead of 1 hour per task
Partner correspondence
Handles the correspondence, remembers agreements and reminds about them on time.
~25 partners, saving ~6 hours a week
Sorting incoming requests
Requests arrive as emails in different formats. The AI pulls out items and quantities, checks them against the price list and stock, and prepares an offer for the manager.
Routing requests
Defines the type of request, urgency and owner, collects missing details from the client and creates a task in the CRM.
Approximate values from NextDoorCoders’ internal use.
How we implement it
From a process breakdown to everyday work
-
1
Process breakdown
In 30 minutes we go through how the process works now, where time is lost and which decisions people make.
-
2
Rules and criteria
We fix what counts as a correct result, which exceptions exist and where your approval is needed.
-
3
A pilot on one process
We run the AI on real data in “advises” mode and measure accuracy and time saved.
-
4
Integration and launch
We connect it to your CRM, email, spreadsheets or accounting system and raise the level of independence where the quality is stable.
-
5
Support and expansion
We watch the decision log, refine the rules and move on to the next process.
Pricing
What the cost of AI automation depends on
The price depends on the process, so we don’t quote a universal figure up front. After a 30-minute conversation you get a proposal for a specific pilot with a fixed scope.
What affects the price
- How complex the process is and how many exceptions have to be handled
- How many systems it has to integrate with (CRM, email, accounting system, spreadsheets)
- The level of independence: from advice to fully automatic actions
- The amount of training material: documents, instructions, a knowledge base
If the solution is built on Claude-based AI agents: a subscription from $20/month on your side, hosting on your server or on ours for $15/month.
Questions about AI automation
It is the use of artificial intelligence for steps of a process that used to need human attention: reading a document or an email, judging a situation, choosing the next action, preparing a decision. The rules and integrations stay, and AI adds an understanding of unstructured data and context.
Ordinary automation works by fixed rules and breaks on unusual cases. AI automation understands free text, takes context into account, explains its decision and hands the cases it is unsure about to a person.
Those with a lot of unstructured data, where the decision depends on context and there are clear criteria for a correct result: handling requests and emails, documents, purchasing, order processing, quality control, reporting.
You decide that for each step. The AI can only advise, prepare an action for your confirmation, or act alone within your rules. Non-standard situations always go to a person.
During the pilot the AI works in advice mode and we compare its decisions with those of people. Every decision is logged, so the quality is visible in numbers, and independence is raised only where it is stable.
CRMs, email, messengers, spreadsheets, accounting and custom systems, as long as there is an API or code access. We don’t rewrite your systems.
The process breakdown takes 30 minutes, and a pilot on one process launches in a few weeks. The exact timing depends on the complexity of the process and the number of integrations, and we give it in the proposal.
On your server or on our separate one. Only the context needed for the task goes to the language model, and access is issued through separate keys that you can revoke.
Shall we go through your process?
Tell us what takes the most time. In 30 minutes we will work out whether the process suits AI and prepare a pilot proposal.
Discuss automation