Preqin~3 min read
An AI routing engine anyone could overrule.
Three months, one PM, one engineer, and an operations team with zero trust in AI routing.
- Role: Product strategy and design
- Timeline: 3 months
- Team: 1 PM, 1 engineer, me
- Impact: Ticket turnaround down 75%, adopted by the whole team
- Platform: Web
Ticket turnaround
six months, down to about six weeks
Revenue
up from 10%
Renewals
up from 5%
The team adopted it fully, and the system managed 90% of the work. Read over the two weeks after launch. Turnaround is an operational measure of this tool. Revenue and renewals are business outcomes on a platform this tool served from the inside. I never wrote down how an internal operations tool produced them, and nobody instrumented it, so read them as context. Still missing: what the 75% and the 90% each measure, and who measured them.
Overview
Operations accepted AI routing only because they could overrule it.
Preqin sells data, and 60% of it was unhealthy at any given time. Accounts touched by stale records were three times more likely to churn, and a ticket took six months to turn around. People tracked the work in Excel.
We built the company’s first AI operations platform. It assigns tasks and priority scores from historical patterns, individual expertise, and team bandwidth.
Worked with: 1 PM, 1 engineer, operations staff, and Engineering.
What I owned
I owned
- Product strategy and design end to end
- User research and the operations journey map
- Framing the opportunity and the constraints
- The MVP definition, and the two directions that failed before it
- Rapid testing across the iterations
Shared with the PM and engineer
- What the routing model could infer from historical patterns
- The launch timing dispute, which neither of us settled: see the note
Research
Twelve steps updated one record, and only one of them was the update.
A job that should have been one step took twelve: find data, prioritize, create tasks, assign them, coordinate, wait on a data check, call Sales, wait for Sales to call the client, publish, and track it in Excel. Eleven steps were coordination.
Research found disjointed workflows, no visibility into the work’s state, and nothing tracked where the team could act on it. None of these were interface problems. Assigning well needs knowing who is good at what and who has capacity, across a queue nobody could see.
Before
Unhealthy data
Ticket turnaround
More likely to churn
From the deck’s problem slide. The churn multiple applies to accounts affected by stale data; the population it was measured across is not recorded.
Decision: routing over a faster interface
I changed the engine and left the surface alone.
A faster interface speeds one step and leaves eleven. I rejected a new interface and patching Excel, because operations were mid-collection cycle and learning a new tool competed with the work. The routing engine sits beneath the existing experience. The risk was capping improvement at the old workflow’s limits.
Decision: control
Assignment is automatic, always overridable, and every override is logged and fed back to the model.
Every objection was about control, and none was about accuracy. I rejected fully automatic and suggestion-only assignment. The log does not answer accountability. If a wrong route goes uncaught, it only records that nobody caught it.
Redesign
A split panel won over kanban, after two directions failed.
Kanban tested well on paper but was unfamiliar to non-technical staff. A priority-sorted list fit the team’s system but hid what teammates were doing. The split panel’s left queue is ordered by the routing score, and shows priority scores so people see why they got work.
Scope
Routing, priority scores, overrides, and a progress bar shipped. The data-input surface stayed in the old system.
Entering some work twice was the price of keeping the team where they were. Engineering called the progress bar scope creep. It shipped because the second iteration lacked visibility into work in flight. I do not know whether Engineering changed its view.
Results
Ticket turnaround fell 75%, from six months to about six weeks.
Operations never came around to AI routing, and the launch happened without agreement
They accepted a system they could overrule. The ops manager wanted to wait a quarter and the PM had promised Q3, so we piloted. What the pilot established was narrower than agreement. Who is accountable for an uncaught wrong route was open when I left.
