CASE STUDY · AGENTIC AI DEVELOPMENT & ORCHESTRATION
Building a multi-agent system that replaced three disconnected tools.
- 15 hoursAbout 15 hours a week of duplicate entry removed
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised mid-market B2B services company. Name withheld under confidentiality.
Intake, validation and exception-handling agents coordinated by a rules-based orchestrator, with a shared audit log.
- 1Work moved by hand between a CRM, an ERP and a ticketing platform; duplicate entry and missed hand-offs caused avoidable delays.
- 2A deal closed in the CRM.
- 3Someone re-typed the order into the ERP.
- 4Someone else opened a ticket so fulfilment would start.
- 5Each step depended on a person noticing the last one.
02 · WHAT WENT WRONG
What went wrong, and how we fixed it.
Every engagement hits problems. These were the ones that mattered here, and what we changed so they did not come back.
- PROBLEM 01Early on, a network timeout caused the ERP adapter to retry, and one test order was created twice
No live data was affected, but it showed the gap. We added idempotency keys to every write, a duplicate check before create, and a test that replays failures on purpose. That test now runs on every release.
- PROBLEM 02That test now runs on every release
03 · THE PROCESS
How we ran it.
We ran this as a staged replacement, not a big switch.
- 01Phase 1, mapping
we traced one order through all three tools and listed every field copied by hand and every decision made along the way.
- 02Phase 2, shared record
we built a single order record and the audit log first, so every later step wrote to the same place.
- 03Phase 3, agents one at a time
intake first, then validation, then exception handling. Each ran alongside the old process for a short period before we retired the matching manual step.
- 04Phase 4, retire
we moved users off the old hand-offs once the orchestrated flow had matched them for a full cycle.
We held a weekly review with the people who did the work, and a short daily check while each agent first went live.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
Staff work from one view of each order instead of three.
- NOW 02
Exceptions arrive in a queue with context, so people spend their time on the cases that need judgement.
- NOW 03
Customer status questions are answered from the audit log.
- NOW 04
The client also keeps the orchestrator rules in version control, so changes to the process go through review like any other change.
Three tools consolidated into one workflow
- 15 hours
About 15 hours a week of duplicate entry removed
Near-real-time status visibility and automatic exception routing
05 · WHAT WE BUILT
How it works.
- The orchestrator is deliberately plain.It is a rules engine that moves an order through named states. It does not ask a model what to do next.
- Models are used only inside the agents, for reading unstructured text and extracting fields.
- Each state change is an event written to the audit log with the actor, the input and the outcome.
- Calls to the CRM, ERP and ticketing platform go through thin adapters with retries and idempotency keys, so a repeated call cannot create a second order.
- Agents hold only the permissions their step needs.
- Monitoring and controls
- Handover
- We ran two sessions with the client's team on reading the audit log and changing rules. We left a runbook covering failed calls, stuck orders and how to roll back a rule change. The client's engineer made the first rule change on her own before we left, with us watching.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- Work moved by hand between a CRM, an ERP and a ticketing platform; duplicate entry and missed hand-offs caused avoidable delays.
- What did Brandsmashers build?
- Intake, validation and exception-handling agents coordinated by a rules-based orchestrator, with a shared audit log.
- How was the work run?
- We ran this as a staged replacement, not a big switch.
- What were the results?
- Three tools consolidated into one workflow; about 15 hours a week of duplicate entry removed; near-real-time status visibility and automatic exception routing.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
