AI agents that already know your company's documents
AIDA's AI agents work on business documents that AIDA has already classified, read and linked together. That is why they can search the entire archive, calculate exact totals instead of estimating them and follow the links between contracts, invoices and payments. You create them with Actio, AIDA's agentic architecture, by describing the work in plain language, without writing code.
Picture someone who has to assess a portfolio of non-performing loans (NPLs). Each position comes with loan agreements, collateral, appraisals, court filings and account statements: hundreds of documents even for a handful of files. They upload them to a general-purpose AI assistant and ask for the total value of the collateral. The answer arrives in a few seconds and is well written, but it is based on the documents and passages the assistant read for that question, and nothing in the answer tells you what was left out.
It's an objection we hear often: “I already do the same things with Copilot, or with ChatGPT.” Understanding where the two diverge is also the quickest way to understand how AIDA's agentic architecture is built.
What is an agentic architecture?
An agentic architecture is the way a system's AI agents are organized: what data they work on, how they split up tasks, when they start and who checks the result. An AI agent is software that receives a goal, plans the steps to reach it, uses the tools available to it and checks what it has achieved, rather than simply answering a question.
In document processes, what matters most is where the agent starts from. If it works on documents that are already classified, with extracted fields and links from one document to another, it reasons on the company's facts. If it receives raw files, it first has to search through them, and everything else depends on that search. The basics are explained in our introduction to agentic AI.
Can I do the same things with Copilot, ChatGPT or Claude?
With a few documents, yes; with a company archive, no. Microsoft 365 Copilot, ChatGPT and Claude are general-purpose assistants, designed to work on files uploaded to or connected to the conversation: for each question they read a limited portion of the documents, search the text instead of querying data, and know neither the history of the documents nor how they are linked.
They read only a few documents per question
Every general-purpose assistant has a limit on the amount of text it can consider at once, known as the context window. Beyond that limit, it looks for the passages most similar to the question and answers based on those alone. This approach is called RAG (retrieval-augmented generation), and the official documentation states its limits (October 2026):
- Microsoft 365 Copilot: when you reference a SharePoint site, it accepts no more than ten files or pages from that site;
- Copilot Studio, Microsoft's tool for building agents: for each answer it uses the top three results from each source;
- ChatGPT: a project accepts 5 to 40 files depending on the plan, and for long documents part of the content goes into the context while the rest is searched in an index;
- Claude: it accepts up to 20 files per chat, and when a project's files exceed the context window it retrieves only the information most relevant to the question.
Microsoft itself states that RAG works better for specific questions than for in-depth document analysis, and that it is not designed to compare entire documents or check their compliance. Even a document that fits entirely in the context is not read evenly: Copilot sometimes focuses only on the beginning, the Lost in the Middle study found that language models make poorer use of information located in the middle of a long text, and research by Chroma on 18 models saw reliability drop as input length grew.
They search text, not company data
A general-purpose assistant finds passages that resemble the question. It doesn't know that one file is an invoice and another a contract, which fields they contain or how they are linked. That is why Copilot Studio can't tell you how many files are in a folder or filter them by a property, and for Excel sheets pulled from SharePoint, Microsoft warns that answers to analytical questions may not be optimal.
In AIDA, every document is already classified by type, with its data extracted. And agent search doesn't stop at the text or the extracted properties: it follows the relationships between documents, the chain linking a contract to orders, invoices and payments. So it finds documents the way someone working in the company would look for them, starting from the business question: “Has ACME paid all the invoices linked to last quarter's contract?”
Totals and counts, moreover, are not estimates made by the language model. AIDA calculates them mathematically and precisely across all the documents found in the archive, with no limit on their number.
The gaps are invisible
The most serious risk is an answer that looks complete but isn't. Copilot Studio, for example, indexes at most 1,000 files for each connected SharePoint or OneDrive folder: any additional files are not processed, and the system doesn't say which ones were left out. Documents with a confidential label or password protection show as ready but produce no answers, and files uploaded from SharePoint and OneDrive refresh every 4 to 6 hours.
AIDA's AI agents work directly on the archive, with no copies to synchronize. Archive storage is unlimited on all plans, so documents from past years remain just as available to the agents as today's.
The model fills in the gaps
When a piece of data is missing or appears twice, a language model doesn't stop: it completes the answer with whatever seems plausible. This happens even with the most advanced models. Research by OpenAI explains that language models, when unsure, guess rather than admit uncertainty, and that these hallucinations persist even in the latest systems. Filling in the gaps is not a task you can delegate to a model: the context it receives must already be correct and certified.
In AIDA, agents start from exactly that kind of context. Duplicate detection recognizes duplicate documents from their text content or extracted data and flags them. If the company chooses, AIDA also blocks their validation: the duplicate never enters the archive, so it doesn't end up in the statistics or in the agents' context, and an invoice received twice doesn't count twice toward a total. Overview, AIDA's dashboard, turns extracted data into charts, statistics and business intelligence indicators that update in real time with every new document. Agents use those figures, the same ones the team sees, rather than reconstructing them from the text.
They don't know the history of documents
To a general-purpose assistant, a file is just text. In AIDA, every document has an event history: payments, approvals, status changes, manual notes. Agents read it, so they know that an invoice is disputed or that a contract was approved yesterday, and they can start from that very event.
| What changes | General-purpose AI assistant | AIDA's AI agents |
|---|---|---|
| Documents considered | The uploaded ones or the passages most similar to the question | All archive documents found by the search |
| How it finds documents | Similarity to the text of the question | Document types, extracted data and relationships between documents |
| Totals and counts | Depend on the passages read or on a manually prepared spreadsheet | Calculated mathematically across all results |
| Document history | Only the file content | Event history: payments, approvals, statuses, notes |
| Archive coverage | File limits per chat, project or source | The entire archive, with unlimited storage on every plan |
| Missing or duplicate data | The model completes the answer with whatever seems plausible | Duplicates flagged, and optionally blocked before archiving; real-time Overview statistics |
How is AIDA's agentic architecture built?
It has two layers: underneath, Intelligent Document Processing (IDP), which classifies documents and extracts data; on top, Actio's specialist agents, which reason on that data. In between, relationships, event history, duplicate control and Overview give the agents correct, certified company context.
| Layer | What it does |
|---|---|
| IDP | Recognizes the type of each document and extracts the data, accurately from day one and with no training |
| Relationships and event history | Link contracts, orders, invoices and payments in a knowledge graph and record approvals, payments and status changes |
| Duplicates and Overview | Flag duplicate documents, with the option to block their archiving, and turn extracted data into charts and statistics updated in real time, which agents use as context |
| Actio | Specialist agents that reason on the data and produce the results; a coordinator Actio distributes the work to the others |
| AIDA GPT | The team's chat, which knows every Actio and routes each request to the right one |
Every agent inherits what AIDA already knows about your documents. You don't need to explain document types, field names or where exceptions hide, and there's no parallel data pipeline to build. Anyone using AIDA GPT makes the request in chat and gets the finished work.
How do you create an AI agent with Actio?
Through a guided chat, in plain language and with no code. The wizard asks the questions and you answer in business terms:
- describe what the agent should do, for example “check every incoming supply contract and flag those without a termination clause”;
- choose the document types it works on, for example Supplier contracts;
- confirm the name, description and instructions the wizard proposes;
- if needed, set a tone of voice, for example formal for customer emails, and restrict the tools and actions the agent can use.
AIDA can suggest how to improve the instructions and shows a preview of the changes before saving them. Each Actio can have multiple versions, so you can test a change without losing the one that works. The configuration stays in plain text: anyone on the team can update it when the process changes.
For larger jobs, a coordinator Actio identifies the documents, hands each one to the right specialist and combines the results. AIDA GPT automatically knows every Actio the team creates, so all you have to do is ask in chat.
When does an AI agent go into action?
When someone calls on it in chat, at a set time or when something happens to a document:
- in chat, at the request of anyone on the team;
- on a schedule, for example every morning at 9 for the collections summary;
- on a document event, for example when a contract is approved, an invoice is archived or a case file is closed.
Every background run leaves a notification linked to the chat, with the full conversation, the documents consulted and the results produced. You can also see how long the task took and how the documents were searched.
Who checks the work of AI agents?
People. In AIDA, human in the loop is the rule: by default, no email, payment request or message leaves the platform without confirmation from someone on the team.
Drafts open in an editable canvas, where they can be reviewed, corrected and confirmed. Once a workflow has been proven and the company asks for it, automatic sending can be enabled: it's an explicit choice, not the starting point.
For long jobs there is Plan Mode. As soon as a request involves three or more items, AIDA GPT displays a plan on screen, updates each item in real time with a link to the source document and lets the team change direction, skip an item or close the plan early. Persistent memory retains the team's corrections, for example “ACME pays at 60 days”, so over time the same question on the same data leads to the same correct answer.
Results arrive ready to use and easy to verify: tables downloadable to Excel, charts in the chat, draft emails and draft payment requests.
What do AIDA's AI agents do in real-world cases?
They work where documents are numerous and rules are complex. A client in the legal sector, for example, processes hundreds of documents from NPL positions in a handful of minutes.
In NPL portfolio valuation, AIDA classifies documents of more than a hundred types into uniform case files. Specialist Actios extract dates, asset values and legal clauses, a coordinator assesses each position and a final Actio consolidates the results into a single valuation, with every total calculated by AIDA's verified arithmetic. The output is charts and tables in AIDA, an Excel file for stakeholders and a report with context and recommendations.
In cash flow management during due diligence, an Actio sorts incoming payments from more than a hundred payment methods worldwide, writes to customers in their own language and handles their replies, then matches each payment to its tax document and reconciles it with the invoice. A coordinator produces monthly, quarterly and annual statistics, and a forecasting Actio estimates future flows.
The same pattern applies to procurement, claims, compliance audits and supplier onboarding: to any recurring job that spans multiple document types.
Frequently asked questions
Do AIDA's AI agents send emails on their own?
Not by default. Emails, messages and payment requests open as drafts that a team member reviews and confirms. On request, automatic sending can be enabled for workflows that have already been proven.
How many documents can an AIDA AI agent analyze?
All the documents the search finds in the archive, with no limit on their number. Totals and counts are calculated mathematically across all results, and archive storage is unlimited on every plan.
Do you need a developer to create an Actio?
No. An Actio is created through a guided chat, by describing the work in plain language, and its configuration stays in plain text that anyone on the team can edit.
What's the difference between AIDA GPT and Actio?
AIDA GPT is the chat the team uses to ask questions and request work. Actio is the agentic architecture that carries out that work with specialist agents, including in the background, on a schedule or on document events.
Does AIDA work with Microsoft 365?
Yes. AIDA integrates with SharePoint and OneDrive and adds automatic processing and structured archiving to their documents, as explained in the article on Microsoft SharePoint and OneDrive integrations with AIDA.
To see AIDA's AI agents at work on your documents, book a demo and bring a process you want to automate.
