Using AI Agents to Process Documents Inside Proclaim, Actionstep and Tessaract: Chronologies, Summaries and Beyond

Every fee earner knows the task. A bundle of medical records, a run of correspondence, a stack of witness statements, all needing to be read, dated, and turned into something usable, whether that is a chronology for a clinical negligence claim or a summary for a case review. It is slow, it is repetitive, and it is exactly the kind of work that eats time a fee earner should be spending on judgement, not transcription.

An AI agent loop is a practical way to take the first pass of that work off someone's desk, without removing the human check that this kind of output always needs.

What an agent loop actually does

An agent loop is not a single prompt into a chatbot. It is a structured process: the AI works through a document page by page, or section by section, extracting what matters as it goes, then assembles the output once every part has been processed.

For a chronology, that looks like this.

• Each page or section is fed through in turn, rather than the whole bundle at once.
• The AI pulls out dated events (what happened, when, and which page it came from) and adds them to a running list.
• Once every page has been processed, the list is sorted into date order.
• A final pass turns the structured list into a readable narrative summary, alongside the underlying table of events.

The same approach works for case summarisation, correspondence review, or any large set of documents that needs reading, extracting and pulling together into something usable. Chronologies are the clearest example, but the pattern is general purpose.

Why the source reference matters

The part of this that firms should not compromise on is traceability. Every extracted event needs to carry a reference back to the page or document it came from, so the output can be checked against the original record before anyone relies on it.

This is not a nice-to-have. Without it, a chronology is just a summary someone has to trust blind, and that is not a safe way to work with case-critical documents.

The value of an AI-built chronology depends entirely on whether it can be traced back to the source. Without that, it is not a chronology, it is a guess with good formatting.

Connecting the loop to Proclaim, Actionstep or Tessaract

The most useful version of this is not a standalone tool that someone has to remember to use. It sits against records already held in the case management system and writes the finished output straight back into the matter.

• In Proclaim, documents are pulled directly from the matter's document store, and the finished chronology or summary is written back into the case file through Proclaim development work built for the purpose.
• In Actionstep, the same pattern runs against the matter record, and Actionstep's own MCP server gives a more direct route for an AI agent to read matter data and write results back without custom middleware.
• In Tessaract, the platform's open API and direct connection to its underlying cloud database give an agent loop the most unmediated route of the three, reading and writing matter data without a vendor-controlled integration layer sitting in between.
• Across all three, the fee earner never leaves the system they already work in. No uploading documents elsewhere, no downloading the output and reattaching it by hand.

Which route makes sense depends on the firm's platform and how its existing development work is structured, but the underlying agent loop is the same across all three.

Cost and data protection scale with the firm, not against it

One thing firms often assume, wrongly, is that adding AI into the case management system means a shared, opaque cost and a data protection headache that grows with volume. It does not have to.

• The firm can hold its own direct account with its preferred AI provider, whether that is Claude, ChatGPT or another, rather than routing usage through a third party.
• Costs then scale directly with what the firm actually uses, rather than sitting inside someone else's pricing model or being bundled into a per-seat fee that bears no relation to volume.
• Data protection sits on the same footing as the firm's own commercial terms with that provider, which the firm can review and negotiate directly, rather than relying on a vendor's blanket assurance further down the chain.
• This also means the firm keeps control over which provider it uses, and can change that decision later without redoing the underlying integration work.

For firms already thinking carefully about compliance and audit trail, this direct relationship is usually the cleaner answer, not a bolted-on complication.

What this does not replace

An agent loop is a first pass, not a final answer. The output should always go through a review step, typically a paralegal or junior checking the chronology against the source pages, before it goes anywhere near a claim or a client. Given the sensitivity of the underlying data, particularly medical records, firms also need to think through how documents are processed and confirm the data processing terms that sit behind whichever AI provider is used.

None of that is a reason to avoid the approach. It is a reason to build it properly, with the traceability, the review step and the data handling considered from the outset, rather than bolted on afterwards.

Getting started

Firms considering this are usually best served by a small proof of concept first: a sample bundle, run through the process against Proclaim, Actionstep or Tessaract, so the output format and accuracy can be judged before any wider rollout.