Home Use Cases Work Insights About Contact
Back to Our Work

How a Growth-Stage Technology Company Built Request Handling That Answers Repeat Questions on Its Own

Administration1
Open the walkthrough

Five-Step AI Workflow with Human Oversight

Each step shows the workspace, what the AI completed, and what still needs human review.

Steps 3 and 5 need a person. Step 3 is waiting for approval right now. Click it to open the full walkthrough.

Estimates are directional and based on stated assumptions. All names, organizations, and identifying details have been anonymized in accordance with our confidentiality agreements.

The Transformation

How a Growth-Stage Technology Company Built Request Handling That Answers Repeat Questions on Its Own

Before
A person used to check every request form for missing fields and chase the requester.
A person used to read every request and answer the same questions by hand.
A person used to carry each approved request from system to system by hand.
After
everything that came in, whole, with the doubtful held
every ask named, the ones the record answers, answered
three packets, evidenced, waiting on your signature
approved steps run, and confirmed by the systems they changed
the week's record, and what should route differently
Executive Summary

Every internal request, from the moment somebody asks to the moment the work is done and on the record. The agents capture each request, read it, answer the ones the record already answers, run the approved ones across the connected systems, and hand a person only what is theirs: a signature where money, access or data moves, and the rules the whole thing runs by. Five stations, run as a loop: what station 5 learns about where requests went wrong changes how station 2 routes them on Monday.

What Was Broken

A person used to check every request form for missing fields and chase the requester
A person used to read every request and answer the same questions by hand
The real cost
A person used to carry each approved request from system to system by hand.

What We Built

Five stations and 13 subagents. Each subagent carries its own tasks and its own refusal.

A4
Request Sweep
Collect every request from the form, the shared channel and email
A3
Requester Check
Check each requester against the staff directory and their current role
A4
Request Reader
Read every request end to end, attachments included
A4
Record Search
Search past requests, policy pages and the connected systems for the answer
A3
Route Sort
Sort each remaining request onto the workflow its kind has always used
A1
Packet Builder
Draft the packet: the ask, the cost, and the systems the request touches
A2
Policy Check
Mark where the request sits against the spending and access limits
A3
Decision Record
Record each decision: who signed, when, and what was on screen
A4
Step Runner
Run each approved step across the connected systems, in order
A3
System Check
Check each changed record against what the request said would change
A4
Report Builder
Build the weekly report from the request record, per team and per kind
A2
Rule Proposal
Propose the routing or holding rule that would have stopped each repeated mistake
A3
Access Review
Compare each person's access against the requests that granted it

How It Runs

1

Request Intake

Nobody is asked anything here

Agents collect every request off the form, the shared channel and email, attach each one to its requester, merge the duplicates, and hold anything missing a fact the work cannot start without. Nothing waits on a person here. A held request is never started and never dropped.

2

Reading and Answering

Nobody is asked anything here

Agents read every request end to end, name the ask in one line, answer the questions past requests and policy pages already answer, and route the rest onto the workflow their kind has always used. Nothing waits on a person here. A request no workflow covers is labelled with its reason, not sent to somebody to work out.

3

Approval

A person answers here

Agents build one packet per decision: the ask, the cost, the systems it touches, what the policy says, and what this requester asked for before. A person signs, and nothing moves until they do. This is where a person works. Everything else exists so that three packets, not forty requests, reach this screen.

4

Fulfillment

Nobody is asked anything here

Agents run each approved step across the connected systems in order, confirm every step against the system it changed, and hold a run whole when two systems disagree about what is true. Nothing waits on a person here. A run whose systems disagree is held whole, with nothing undone quietly.

5

Reporting and Tuning

A person answers here

Agents build the weekly report from the request record, flag the numbers moving against their own pattern, propose the routing rules that would have stopped the week's mistakes, and list the access nobody uses. A person accepts or rejects each rule. How requests route, and who holds access, is not something an agent changes quietly.

Where a Person Decides

Step 3, Approval. Sign the laptop, the packet is ready. Give the expense raise a number, or decline it. your name is on the yes.
Step 5, Reporting and Tuning. Accept or reject the two proposed rules. Decide the 14 unused grants. rules and access are yours to accept.

Operating Model

This changes how work flows through the team.

Role
Responsibility
Approval owner
Sign the laptop, the packet is ready. Give the expense raise a number, or decline it. your name is on the yes.
Reporting and Tuning owner
Accept or reject the two proposed rules. Decide the 14 unused grants. rules and access are yours to accept.
Request Intake, when it goes wrong
Look here only if a requester says their request vanished.
Reading and Answering, when it goes wrong
Look here only if an answer reads differently than the policy you remember.
Fulfillment, when it goes wrong
Look here only if a requester says a step never landed.

What Transfers, What Must Be True

What transfers
A person is in the loop wherever money, access or data moves in their name, or the rules the agents run by change. Two stations refuse to proceed on their own, station 3 Approval and station 5 Reporting and Tuning, and those are the two a person carries. Everywhere else the agents capture, read, answer from the record, run and confirm, and reach a person only when the yes is theirs to give.
Approval stops for a person, and your name is on the yes.
Reporting and Tuning stops for a person, and rules and access are yours to accept.
Every subagent says what it will not do. 13 of them do.
What must be true in your environment
The agent can read the systems your records already live in. This one reads 17.
Somebody owns Approval and has time for it.
Somebody owns Reporting and Tuning and has time for it.

Failure Modes

What breaks this pattern:

✗ Requests run for people who left

An employee quits and their open requests keep moving. The agent grants access or ships a change to an account nobody owns, and nobody notices until an audit finds it.

✗ The record gets guessed

The record does not hold the answer, so the agent fills the gap with something plausible. The requester acts on a made-up fact and the mistake carries their name, not the agent's.

✗ Money moves on a schedule

Approvals fire on a timer instead of a signature. A payment goes out because Friday came, and the person whose name is on the decision never saw it.

✗ A failed step cleans up quietly

A step fails mid-run and the agent undoes part of the work on its own. Half the change stays in place, the log says nothing failed, and the next request builds on a system in a state nobody chose.

Directional Outcomes

What the agent counts, and the station that counts it.

These counts are the tallies from one monitored run of the agents. They are not monthly or annual totals.

Requests screened
Counted at Request Intake
4,300
Whole requests held wrongly
Counted at Request Intake
0
Held requests later released
Counted at Request Intake
6 of 58
Requests read
Counted at Reading and Answering
9,800
Answers given from a stale page
Counted at Reading and Answering
0
Requests it would not route
Counted at Reading and Answering
41
Steps run
Counted at Fulfillment
18,700
Steps confirmed wrongly
Counted at Fulfillment
0
Runs it held for a person
Counted at Fulfillment
23
Our measurement policy: We do not publish precise ROI without baseline methodology. Every figure above carries its basis.

What Runs Where

Every step names the subagent that does the work, the record it writes, the thing that raises a question for a person, and what it is allowed to touch. This is drawn from the source, not from a diagram somebody kept in sync by hand.

1Request Intake
subagentrequest-arrivals
writesrequests/intake/<req-id>.json
raiseshold-incomplete-request
may touchrequests/intake/**, requests/held/**, the staff directory read-only
2Reading and Answering
subagentread-and-route
writesrequests/read/<req-id>.json, answers/<req-id>.json
raisesno-workflow-covers-it
may touchrequests/read/**, answers/**, the record read-only
3Approval
GATE
subagentpacket-and-decision
writesrequests/packets/<req-id>.md
raisesawait-your-decision
may touchrequests/packets/**, everything else read-only
4Fulfillment
subagentrun-the-steps
writesrequests/runs/<req-id>.json
raisessystems-disagree
may touchrequests/runs/**, the connected systems through their own APIs
5Reporting and Tuning
GATE
subagentreport-and-tune
writesreports/week-<date>.json, rules/proposed/<rule-id>.json
raisesaccept-rule-change
may touchreports/**, rules/proposed/**, the record read-only

Stack

Every system this agent reads or writes.

System
Read at
Stations
email
Request Intake
1 of 5
past requests
Reading and Answering
1 of 5
the HR system
Reading and Answering
2 of 5
the Monday report
Reporting and Tuning
1 of 5
the access grants
Reporting and Tuning
1 of 5
the finance system
Reading and Answering
2 of 5
the identity system
Fulfillment
1 of 5
the policy pages
Reading and Answering
2 of 5
the request form
Request Intake
1 of 5
the request record
Reporting and Tuning
1 of 5
the requester's history
Approval
1 of 5
the routing rules
Reporting and Tuning
1 of 5
the shared channel
Request Intake
1 of 5
the spending limits
Approval
1 of 5
the staff directory
Request Intake
1 of 5
the vendor portal
Fulfillment
1 of 5
the vendor records
Approval
1 of 5
Next Step

Want to see if this pattern fits your internal tools?

No build commitment·Real samples, not a demo·Estimate in writing