CASE STUDY · AGENTIC AI DEVELOPMENT & ORCHESTRATION
How a 12-person ops team automated 65% of routine decisions with AI agents.
- 65%65% of routine purchase orders decided without a person
- 18→6 hoursWeekly validation effort down from 18 to 6 hours
- 5 daysApproval time down from 5 days to 1 day
01 · PROJECT OVERVIEW
The client and the brief.
An anonymised mid-market B2B services company. Name withheld under confidentiality.
Three agents for intake, validation and exception routing, connected to the CRM, ERP and ticketing systems. People review policy exceptions, low-confidence cases and missing data.
- 1The operations team spent 18 hours a week validating and routing purchase orders, causing rework, inconsistent approvals and delayed fulfilment.
- 2The work was done by a 12-person team.
- 3Each person applied the approval policy a little differently.
- 4A purchase order that one person passed in minutes sat in another person's queue for days.
- 5Missing fields were found late, so orders bounced back to the requester and restarted the clock.
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 01In shadow mode the validation agent flagged far too many orders as exceptions
The cause was supplier names. The same supplier appeared under several spellings across the CRM and ERP, so the agent could not match them and treated each as unknown. We built a supplier alias table, checked it with the operations lead, and added a rule that unknown suppliers go to a person instead of being guessed.
- PROBLEM 02Exception volume fell back to a sensible level before we went live
- PROBLEM 03A second, smaller problem appeared in week 7
Two purchase orders for the same goods arrived a day apart from different requesters, and the agents approved both. Policy said the second should be queried. We added a duplicate check on supplier, item and value within a set window, and routed matches to a person.
- PROBLEM 04We also set a weekly report of agent overrides, so the lead could see where people disagreed with the agents
03 · THE PROCESS
How we ran it.
Three Brandsmashers engineers (AI, backend, QA) worked for 8 weeks.
- 01Week 1
we sat with the team, read the written approval policy and pulled a sample of past purchase orders with the decision taken on each. We then wrote down the rules the team actually followed, which differed from the policy in a few places.
- 02Weeks 2 to 4
we built the intake and validation agents in shadow mode. They read each order and recorded what they would have decided. People kept working as before.
- 03Weeks 5 to 6
we compared agent and human decisions every few days with the operations lead, fixed the misses and agreed which cases the agents could decide alone.
- 04Weeks 7 to 8
live on routine orders with the review queue open, then runbooks and handover.
We met the team twice a week for a short demo and review.
04 · RESULTS
What changed.
Measured against the way the work ran before the engagement began.
- NOW 01
The team no longer validates routine orders.
- NOW 02
They review the exception queue and a weekly sample of automated decisions.
- NOW 03
The operations lead owns the rules file and changes it through a short review, so policy changes are written down.
- NOW 04
The team also has a record of why each order was approved, which it did not have before.
- 65%
65% of routine purchase orders decided without a person
- 18→6 hours
Weekly validation effort down from 18 to 6 hours
- 5 days
Approval time down from 5 days to 1 day
05 · WHAT WE BUILT
How it works.
- The intake agent reads the order and normalises supplier names, units and cost codes.
- The validation agent checks the order against policy rules held in a versioned rules file, not buried in a prompt.It also checks the ERP for budget and supplier status.
- The exception agent routes anything uncertain to a named reviewer in the ticketing system with a short summary and the evidence.
- Each agent uses a scoped service account.The validation agent can read the ERP but cannot write to it. Only the final approval step writes.
- Every decision is stored with the input, the rule applied and a confidence band.
06 · QUICK ANSWERS
The short version.
- What problem did the client have?
- The operations team spent 18 hours a week validating and routing purchase orders, causing rework, inconsistent approvals and delayed fulfilment.
- What did Brandsmashers build?
- Three agents for intake, validation and exception routing, connected to the CRM, ERP and ticketing systems. People review policy exceptions, low-confidence cases and missing data.
- How was the work run?
- Three Brandsmashers engineers (AI, backend, QA) worked for 8 weeks.
- What were the results?
- 65% of routine purchase orders decided without a person; weekly validation effort down from 18 to 6 hours; approval time down from 5 days to 1 day.
YOUR TURN
Have a similar requirement?
Tell us what is stuck. We will come back with a plan and matched, pre-vetted engineers.
