Professional services

Delivery leaders could not see which engagements were at risk because status lived in email, slides and the memories of project managers.

A growing professional-services firm had plenty of AI tools. Partners used one copilot. Delivery used another. Finance used a third. None of them could answer whether an engagement was actually on track.

Anonymized. This page describes the operating problem and the work that was done. It does not report measured completion rates, cost or time savings.

Before

Status in email and slides
Late risk discovery
Tool sprawl across practices
No shared definition of “at risk”
AI used without a control model

After

Shared engagement context
Early, evidence-backed risk flags
Human judgment on client communication
Permission-aware commercial data
Weekly evaluation of the system

What the business was experiencing

Weekly status meant assembling a slide from five inboxes.
Risks surfaced late, after the client already felt them.
New project managers inherited folders, not operating context.
AI summaries of documents missed the commercial commitments in email.

Why existing technology was insufficient

Document search returned files. It did not return obligations.
A generic assistant had no permission model for partner-only commercial terms.
Automating status email without redesigning the workflow produced faster, still-incomplete updates.

What changed

GreyShacks mapped the delivery operating model: what “at risk” means, who must know, and which actions require a partner.
Engagement context was assembled from approved sources with role-aware access.
An advisory agent flagged risk. A partner still owned the client conversation.

Architecture

Engagement as the unit of context, not the document.
Advisory autonomy for risk flags. No autonomous client communication.
Evaluation against a defined risk rubric, reviewed by delivery leads.

Security and governance

Partner-only commercial fields excluded from broader retrieval.
No agent write-access to billing systems in the first release.
Audit log retained for every risk flag shown to a human.

Implementation

Started with one practice, not the whole firm.
Controls designed before any agent was given tools.
Operations cadence: weekly evaluation of false positives and missed risks.

Result

Risk conversations started earlier, with evidence.
Status stopped being a slide-assembly exercise.
Partners kept control of client-facing judgment.

Start with one workflow.

Show us a process where AI still depends on people to gather information, make routine decisions, move between systems or finish the work. We will assess what is preventing further automation and what it would take to improve it.