Skip to main content Skip to footer

Interested?

Get in touch

One engine. Every mailbox.

Posted on: 17th November 2021

Case study · Legal and professional services

One engine. Every mailbox.

A UK legal firm's AI could already read client emails and open new cases. We helped take it from pilot to production, so a new mailbox now goes live with a configuration change, not a new build.

At a glance

The challenge

The firm's clients send new instructions by email. A pilot proved AI could read those emails and open cases, but rolling it out to every office, team and mailbox risked building a separate application for each one.

The impact

We built one shared AI pipeline that routes, classifies and validates every instruction. A new mailbox now goes live with a configuration change, not a new build, with consent checks and human review built in.

Key services

  • AI solution design
  • Email classification and extraction
  • Azure cloud architecture
  • Controlled rollout and testing

Technologies

  • Microsoft Exchange Online
  • Microsoft Graph
  • Azure Container Apps
  • Azure OpenAI
  • Azure Key Vault

The challenge

The pilot worked. Scaling it was the real test.

Email is how the firm's clients send new instructions. The pilot proved AI could read those emails and pull out what's needed to open a case. The question was how to support every office, team and client mailbox without building a separate application stack for each one.

  • Mixed mailboxesSome mailboxes only take instructions. Others carry everyday correspondence too.
  • Routing without guessworkRouting couldn't rely on the AI already knowing what a message meant.
  • Different rules, different boundariesSome work types need their own integrations, retention controls or data isolation.
  • Consent comes firstEach client's AI consent position has to be checked before anything is processed.
  • Validated, not trustedExtracted case data must be checked against authoritative sources before it's committed.
  • A safe route to peopleAmbiguous, low-confidence or non-consented messages need a clear path to human review.

The design principle

“A container is a processing engine, not a mailbox.”

A new mailbox, team or alias doesn't automatically earn a new container. We only add one when a flow needs materially different processing, or a genuine security or data-isolation boundary.

The solution

Route first. Then let the AI think.

We separated two jobs that are easy to blur. Simple, predictable rules decide where a message goes. Only then does AI work out what it means.

01 · Routing

Which approved environment should handle this message?

Dedicated instruction mailboxes are read directly through Microsoft Graph. In mixed mailboxes, Exchange Online rules use known details, like the recipient alias or sending domain, to divert only the instruction flow. Anything that can't be selected reliably stays in the normal process or goes to review.

02 · Classification

What does it mean, and what case information does it hold?

Inside the shared pipeline, Azure OpenAI uses the firm's instruction taxonomy, extraction schema and real examples to classify each email and pull out the case fields. It proposes structured data. Validation and business rules decide whether it's safe to use.

How a message flows: client mailboxes → Exchange routing rules → shared classifier pipeline → consent and validation → new case created, or human review.

One shared pipeline

Eight steps, every message, every time.

However a message arrives, it follows the same consistent path, from inbox to a validated case in the practice management system.

  1. 1IngestThe email, its metadata and supported attachments.
  2. 2ConvertSupported attachment formats turned into text.
  3. 3Classify and extractAzure OpenAI, guided by the firm's taxonomy, schema and rules.
  4. 4Check consentClient configuration and handling rules applied.
  5. 5ValidateIdentifiers checked against authoritative firm or external sources.
  6. 6StructureValidated JSON or XML, never free-form AI text.
  7. 7Create the caseThrough the approved practice management integration.
  8. 8Hand over when unsureAnything uncertain goes to human review.

Built on Microsoft Azure

Secure foundations, from inbox to case file

Microsoft 365

Exchange OnlineMailboxes, aliases and deterministic routing rules.

Microsoft Graph and Entra IDLets the approved app read new email and attachments.

Azure

Container AppsShared classifier, plus a second container only where justified.

Azure OpenAIClassifies meaning and extracts fields into a controlled schema.

Blob Storage and Key VaultRetention-controlled storage; secrets kept out of code, with managed identities where supported.

Monitor and Application InsightsLogs, audit evidence, performance and alerts.

Firm systems

Practice managementAuthoritative validation and case creation through an approved interface.

Human review queueManual handling for anything low-confidence, ambiguous or non-consented.

A controlled rollout

A test mailbox and staged routing let us bring each client or instruction flow in one at a time. Every flow is tested against known examples, reviewed by the firm's own people and monitored before it goes live, without touching unrelated mailbox traffic.

The results

Growth without the sprawl

The firm can now expand the service while keeping infrastructure in proportion to what it really needs.

  • One pipelineServes multiple mailbox feeds.
  • Config, not codeA new mailbox, team or alias is usually a configuration and permissions change.
  • Built-in oversightConsent, validation, human review, logging and alerting are part of the operating model.
Before and after
Measure Pilot Production
Mailbox feeds per processing engine Limited pilot Multiple feeds on one shared pipeline
Infrastructure for a new mailbox, team or alias Undetermined Configuration and access changes only
Time to register a new instruction as a case [Approved baseline] [Measured result]
Eligible instructions created automatically [Approved baseline] [Measured result]

“

[Add an approved quote from a client stakeholder]

[Name withheld], [Role], UK legal firm

What we learned

Don't use AI to decide where AI should look

A dedicated mailbox or alias already tells you a message belongs in the automated process. Save the model for the semantic work: understanding the instruction and extracting the case.

Managed identities and Key Vault keep credentials out of code. Human review and authoritative validation stop uncertain output from ever becoming a case on its own.

What's next

More mailboxes, more instruction types, sharper thresholds

We'll onboard further mailbox feeds through configuration, expand the instruction taxonomy and extraction schemas, and use production evidence to fine-tune confidence thresholds and review rules.

New containers will only appear where a documented processing, integration, security or isolation need calls for one.

Could your inbox open its own cases?

Let's talk about how a similar approach could turn client emails into validated cases in your organisation.

Get in touch

Could your inbox open its own cases?

Let's talk about how a similar approach could turn client emails into validated cases in your organisation.

Get in touch

Our vendors, suppliers, and partners