--- id: KB-TL-029 url: https://app.codecontract.io/help/trackline/how-this-differs-from-an-ai-agent idioma: en categoria: trackline subcategoria: empezar audiencia: usuario nivel: intermedio actualizado: 2026-08-23 tambienEn: [es] relacionados: [KB-TL-015, KB-MA-013, KB-TL-030] citadoPor: [KB-TL-028, KB-TL-015, KB-MA-005] --- # How this differs from an AI agent _An agent writes and chases very well. What it cannot do is make the other side answer, or turn a generated answer into a document that counts._ **Responde a:** why not build an ai agent to request the documents · difference between an ai agent and trackline · can ai chase suppliers for me · automating document requests with ai **It is a fair question, and the short answer is that there is AI here too — what changes is where it sits.** An agent that drafts the request, decides who to nudge and summarises what arrives does real work. The problem is that the hard part of this process is neither drafting nor remembering: it is getting the document out of a third party's hands and in through a route you can later prove. ## What an agent solves and what it does not | Job | An agent does it | Something else is needed | | --- | --- | --- | | Drafting the request | Yes, and well | — | | Deciding who to nudge today | Yes | — | | Reading and summarising what arrives | Yes | — | | **Arriving through a channel they use** | No | Real multi-channel sending | | **Knowing who delivered it and when** | No | The system's own record | | **The team seeing the same status** | No | A shared case | > [!IMPORTANT] > **A model generates text; a case needs documents.** Everything else hangs off that distinction. Ask an agent for a company's certificate and what it can give you is what it knows or finds — never the document that company holds and has not handed over. And an answer that sounds like a certificate does not count as one for anybody: not a customer, not an auditor, not a court. ## The silent failure, which is the one to know about 1. **The agent assumes it already received it** — The classic failure: it believes a task completed and closes it with nobody confirming. 2. **And nobody notices until it matters** — The gap surfaces in the audit or the meeting, not on the day it happened. 3. **So status is not declared by the requester** — A document counts as delivered when it arrives, not when someone says it did. ## Where the AI is here **En corto** - **Building the process**: describe what you need and the template proposes itself, phases and documents included. - **Reading what arrives**: it extracts the data and flags what does not add up, instead of leaving you 40 PDFs to open. - **Answering questions** about your cases, through Marta, which consults what exists rather than inventing it. - **And it stops there**: AI speeds up the work around the document. The document, and the proof of who provided it, are not generated. > [!WARNING] > There is one case where your own agent is the answer: **when the process is entirely internal and does not depend on anyone outside replying**. Drafting, cross-checking data you already hold, watching expiry dates on what is already filed. That needs no channel, no identity for the responder and no delivery trail — the three expensive things to build, and the ones that make this not an agent. > [!NOTE] > What a system can attest about a delivery — and what weight that carries in a particular procedure — **depends on the applicable framework and is for your adviser to confirm**. What is explained here is operational: that it is on record who delivered what and when, and that the record does not depend on someone remembering to note it. **So you do not use AI?** We do, in three places: building the process, reading what arrives and answering questions. **Can I connect my own agent?** Yes, through the API. What you should avoid is letting it declare delivery status. **Does AI decide on its own whether a document is valid?** No. It proposes and flags; a person approves or rejects. ## Ejemplos **A team builds an agent that emails suppliers asking for certificates. Three weeks in it has sent 260 emails and cannot say how many documents came back.** - Lets the agent draft and prioritise - Moves sending and receiving into the case - Measures deliveries, not emails sent → The metric shifts from «emails sent» to «documents missing», which is the one that says whether the work is moving. **An agent summarises a supplier reply as «confirmed they will send the certificate». Two months later, in the audit, the certificate is not there.** - Stops accepting confirmations as status - Requires the document to arrive before closing the requirement - Reviews which other requirements closed the same way → Seven more requirements turn out to have closed on a promise rather than a document — in time to request them. **A company wants AI to validate its subcontractors' insurance and finds the hard part is not reading the policy but getting them to send it.** - Uses AI to read and check cover and dates - Uses the process to request and chase - Approves the doubtful ones by hand → Automatic reading stops idling while it waits for documents that never arrived. **A manager asks a general assistant for «this company's tax clearance certificate» and receives a very convincing explanation of what that certificate is.** - Separates explaining from obtaining - Sends the request to whoever holds the document - Stores what arrives → The explanation stops being mistaken for the document — the mistake that costs a week. **An accountancy firm automates reminders with AI and still has the same 20 % of clients who never answer email, however much they are chased.** - Switches that 20 % to WhatsApp or SMS - Leaves reminders where they work - Compares response by channel → The 20 % that was a channel problem, not a persistence problem, starts answering. **A technical team connects its own agent through the API to launch processes from their ERP.** - Lets the agent create and launch the process - Leaves status and trail to the system - Reads the result back into the ERP → The ERP triggers the work and receives the outcome, with nobody declaring by hand what arrived. **Someone asks Marta about a figure in a case and Marta answers citing the document and the date it arrived.** - Asks in plain language - Checks the source it cites - Opens the document if the detail matters → The answer can be verified in two clicks, which is what separates it from a generated one. **A management team considers replacing the process with an in-house agent and asks for a list of what they would have to build.** - Lists channel, responder identity and delivery trail - Prices each one - Decides with that list in front of them → The decision stops being «AI or not» and becomes which three pieces to build and what they cost.