Legal teams are being pushed to do two jobs at once: move faster and think deeper.
That’s no small feat, but it feels colossal when the volume of data keeps rising, formats keep multiplying, and the questions that matter are rarely generic: they’re tied to a regulator, a contract population, a product line, a region, or a specific window of time.
It can be particularly painful when a regulator’s request hits an existing workflow that was never built for that specific fact pattern. Teams may be forced to squeeze a nuanced, evolving matter into a fixed template, or fall back to manual review just to keep the story straight. It’s a forced tradeoff: repeatable workflows that miss nuance and context, or manual review that captures nuance but doesn’t scale and burns time.
Custom analyses removes that tradeoff by extending, not replacing, the standardized workflows teams already trust. It keeps the discipline of predefined review patterns – consistency, quality control, defensibility – then adds an adaptable analysis layer tuned to the risk profile of each matter, which includes the “hard” data legal teams rely on but struggle to operationalize (think photos, scans, receipts, screenshots, and messy documents that don’t behave like clean text).
With AI-driven extraction and labeling, visual and semi-structured content becomes structured, searchable insight you can stand behind without abandoning your existing process. The result is control: you define what to extract, how to classify it, and how it flows downstream so your team can act with speed and confidence while keeping the backbone of your workflow intact.
Custom Analysis That Fits the Matter
Standard workflows create consistency, quality control, and defensibility, and they give teams a common starting point for every new matter. When teams stop forcing every matter into the same template and instead treat custom analyses as a layer on top, they can turn case theory and business context into operational choices: what to capture, what risk the data signals, and what should happen next.
That means moving beyond one-size-fits-all tags like “responsive/non-responsive,” not by discarding them, but by enriching them with matter-specific classifications and extractions that match how decisions are made. For example, using custom analyses, you can:
- Extract a regulator request ID and map documents to that request
- Identify document role (response, clarification, remediation plan)
- Classify document type using your own definitions
- Generate a short factual summary for triage
- Flag industry-specific risk indicators in controlled categories
The point isn’t customization for its own sake or building a completely separate path, but standardizing how you customize so expert judgment becomes reusable, testable, and consistently applied across reviewers, custodians, and time, all while staying anchored to the core workflow your team already knows.
Modern custom analyses makes this workable at scale because teams can define what they want in plain language, then apply it document-by-document across the full data set without abandoning existing review structures. Ideal implementations keep each insight single-purpose – one prompt, one objective – so outputs stay clean and don’t bleed into each other. That structure also supports disciplined iteration: validate on a small development set, tighten criteria, then scale with an audit trail that shows how decisions were made and how they interact with standard review steps.
This is most impactful in regulatory and investigative work where “relevance” is only the starting line and the real challenge is aligning to external expectations – organizing documents around the regulator’s requests and the organization’s responses in a way that stays consistent as the matter evolves and as new requests arrive. Teams often try to stitch this together with tags, keyword searches, and manual interpretation, which break when the same concept shows up ten different ways.
Custom analyses shifts the workflow from “read and type” to “apply matter intelligence at scale,” producing structured outputs that support tracking, prioritization, and fast pivots without giving up defensibility.
What This Looks Like in Practice
Say a review team receives a broad regulatory request covering a three-year period across multiple custodians. The standard workflow kicks in: ingestion, processing, first-pass review for responsiveness and privilege. That part runs fine.
Challenges surface in the next step. The regulator has organized its request into seventeen numbered sub-requests, each tied to a different business function. The team needs to know not just whether a document is responsive, but which sub-request it responds to, what role it plays (does it describe the practice, authorize it, or document a corrective action?), and whether it’s already been produced in a related matter.
None of that maps to existing tags. The options are to build a parallel manual tracking sheet – which breaks the moment a new custodian batch comes in – or fold the analysis back into review and hope reviewers apply it consistently across shifts and firms.
Custom analyses creates a third path. The team defines three targeted extractions: a sub-request mapping field, a document role classifier, and a prior-production flag. Each runs as a single-purpose prompt against the full data set. Reviewers still make the responsiveness call – that doesn’t change – but when they open a document, the sub-request mapping is already populated, the role classification is there for them to confirm or override, and the prior-production flag has surfaced anything that needs a cross-matter check.
The workflow doesn’t look different from the outside: same platform, same quality-control steps, same privilege log. What changed is that three decisions that previously required a reviewer to hold the entire regulatory map in their head are now structured inputs they validate rather than originate.
Throughput and consistency go up, and when the regulator asks for a production index organized by sub-request, it’s already there in a filtered export.
Bringing Images and Messy Artifacts into the Workflow
Next, there’s the data everyone dreads: images, scans, screenshots, and handwritten notes. This is often where the most human context lives, but also where traditional tools push teams back into slow, manual work that sits outside the main workflow, creating a separate, harder-to-track lane.
Vision-capable analysis changes the equation by letting you ask for exactly what you need and tie those outputs back into the same structured processes you use for text:
- A short factual description of what’s shown
- Extraction of visible text (even when angled, partially blocked, or in glare)
- Identification of logos or locations based on visible cues
- Classification of safety hazards into controlled categories (e.g., “critical” versus “non-critical”)
- Transcription and categorization of handwritten notes
When visual and semi-structured data become part of the same structured analysis framework as everything else, it stops becoming the “leftovers” pile that sits outside your main review track and humans get more time for higher-order judgment.
Prompt Design for Reliable Image Analysis
The difference between a demo and a dependable workflow is prompt design. “Summarize this image” is vague, hard to QA, and easy to over-interpret, especially when you need outputs that plug into existing review and reporting structures.
A practical prompt does three things:
- Defines a tight output format
- Draws a hard line between what’s visible and what’s inferred
- Adds guardrails for edge cases
A simple, repeatable structure outlines:
- Stated facts: only what is clearly visible (bullets, no interpretation)
- Potential inferences: reasonable conclusions (bullets), or “None” when unsupported
Add explicit rules for common failure modes:
- If text isn’t readable, say so
- If reflections appear, flag them and don’t treat them as evidence
- If the image is low quality, note limitations rather than guessing
This is how AI shifts from “clever narrator” to “dependable analyst”: clear instructions, predictable structure, outputs you can filter, search, and quality-check at scale, and results that can be slotted into existing tags, fields, and reports instead of creating a parallel system.
Text extraction from images is where teams feel the immediate payoff. AI can often pull usable text from real-world conditions that break traditional OCR – stylized branding, partial labels, glare, glass reflections. That extracted text becomes an indexable handle for investigation strategy: clustering similar images, prioritizing review, corroborating timelines, and connecting artifacts that would otherwise stay isolated, all while feeding the same search and filtering tools already in place.
You can also run targeted checks with disciplined outputs, for example:
- “Does this image contain visible reflections? Yes/No. Then describe in ≤10 words.”
Small prompts like this create reliable pipelines across hundreds or thousands of images, instead of one-off human interpretation, and they do so in a way that naturally layers into your existing quality-control and escalation processes.
Anomaly Detection: Use AI as a Spotlight
Anomaly detection is one of the most practical ways to cut review time while increasing confidence – especially for receipts and expense artifacts packed with dates, amounts, weights, and operational patterns. AI can flag “what’s odd” in a defined format, and the investigator decides what matters, using the same review lanes and approval paths already in place.
The operating principle is important: AI is a spotlight, not a judge. Use it to identify outliers and inconsistencies, then route flagged items into a tighter human review lane. Done well, this approach reduces noise while keeping accountability where it belongs and strengthens the existing workflow instead of sidelining it.
Custom analyses also helps when metadata is missing or unreliable. A street sign photo with no geotag can still become searchable by extracting street names; a venue photo can yield text from signage plus a factual description of what’s happening. That turns images into actionable records without pretending the system has perfect certainty, and it does so in a way that can be mapped back to standard matter fields and reports.
Notably, responsible use still matters. AI can be weak at math, can struggle with low-quality images, and won’t replace foundational investigative practices. The operating model that earns trust is the same one legal teams already know: start small, sample, iterate, validate, then scale – building a prompt library that becomes part of the standard playbook, sitting alongside existing checklists and templates rather than competing with them.
Conclusion
Winning teams don’t have to choose between standardization and customization – they can combine them. Build a repeatable backbone, then layer in custom analyses that reflects each matter, client, and risk, treating it as an extension of what already works instead of a wholesale replacement.
Operating this way will help turn your legal team’s expertise into a scalable asset: clearer outputs, faster triage, sharper prioritization, and a workflow that empowers you to direct and explore data while keeping the familiar guardrails in place.
Custom analyses also brings images and semi-structured artifacts into your core legal intelligence workflows instead of leaving them behind or handling them in a separate, ad hoc way.
Define the questions, the format, and the guardrails, and you get consistent, defensible outputs that plug straight into search, filtering, and decision-making – with humans firmly in control of what the data means and how it fits within the standard processes your organization already trusts.
Kelly O’Brien is a product marketing manager on the public sector team at Relativity.
