# Sample: One AI Agent, Reviewed at 90 Days | Agent Blueprint

Source: https://agentblueprint.ai/sample-review

Sample / fictional company and numbers

# One agent, 90 days later.

This is what you get back from us. Calder Freight, a made-up company, launched an agent that sorts incoming IT tickets. Before launch we measured where they stood and wrote down the numbers it had to hit. At 90 days we measured again. Below: the summary for finance, the scorecard, then the agent's full page.

For finance

Fix

## Ticket Sorter: fix one gap, keep it running.

Calder Freight, IT service desk, 4,000 tickets a month. Reviewed at day 90.

**Staff time** (Saves 139 hours sorting and fixing tickets ($5,560 at $40 an hour, fully loaded). Running it takes 16 hours back: checking its work 10 h x $40, upkeep 6 h x $70 = $820.): +$4,740 a month

**New cash** (Hosting and connections. Plus the one-off $12,000 build and $7,500 review.): \-$80 a month

**Allocated AI usage** (Calder's agents use all 120,000 credits a month in its $144,000-a-year AI add-on; this one uses 20,000, a sixth. Already paid. Dropping it would let Calder renew on the 100,000-credit tier ($124,000 a year), saving about $1,700 a month from renewal.): \-$2,000 a month

Payback in cash

None yet

It adds $80 a month in new cash and saves none until the freed hours replace contractor time or a planned hire, or the add-on shrinks at renewal.

If staff time counts

4.2 months

$19,500 / ($4,740 - $80) a month

If staff time counts and its AI share is charged

7.3 months

$19,500 / ($4,740 - $80 - $2,000) a month

Decision: Fix. Keep it running, close one gap, and check again in 60 days (the recheck is included).

-   Whether freed hours become cash is Morgan Lee's call.
-   Confidence: medium. Reassignments come from the full ticket history. Hours come from ticket log timestamps, checked by timing 300 tickets by hand before launch and in week 12: 2 minutes to sort one, 5 to fix a misroute.

## Scorecard at day 90

Tickets sent to the wrong team first

Missed

**Before**: 22%

**Target**: 12%

**Day 90**: 14%

Staff hours a month sorting and fixing tickets

Met

**Before**: 206

**Target**: 90

**Day 90**: 67

-   Held: Major incidents sent to the wrong team: 0 before, 0 at day 90.
-   Held: Time to first touch on P1 and P2 tickets: unchanged.

[Get this for one of your agents](https://agentblueprint.ai/#contact)[Price and what's included](https://agentblueprint.ai/#offer)

## If your agent is already running

There's no pre-launch timing study, so the "before" number comes from history, and the decision comes in three weeks, not 90 days.

-   Before: the same measures from your ticket history for the 90 days before the agent went live.
-   No clean history: for two weeks, send a random 1 in 10 tickets around the agent and compare the two groups.
-   Cost: its share of the AI bill from the usage log, plus hosting and staff time, split into what you'd save by stopping it and what you'd only save at renewal.

The agent's page / stage: measured

## 1\. Purpose

**Does**: Reads each new IT ticket and sets its category, priority and team, with a one-line reason a person can check.

**Doesn't do**: Answer, resolve or close tickets. It never handles a ticket once a person has picked it up.

## 2\. Authority

**Can do on its own**: Set category, team, and priority P3 or P4 on new tickets. Enforced by: its ticket system role, which can edit only those three fields.

**Needs a person**: Anything that looks like P1 or P2. It flags the ticket and the on-call lead decides. Enforced by: the ticket system, which blocks P1 and P2 for its role.

**Must never**: Close a ticket or change one a person has touched. Enforced by: its role, which has no close permission and is locked out once a person is assigned.

**Runs as**: The service account ticket-sorter, with no admin rights.

**Limit on its usage**: 25,000 credits a month. Enforced by: a monthly credit limit on its account, which stops it and alerts Jordan.

## 3\. Proof

**Tested with**: 1,000 past tickets before launch. 88% matched how the team had routed them, and no major incidents were missed. The bar was 85%.

**Recorded in**: The ticket history, with the agent's reason on every change.

## 4\. Owner

**Accountable**: Jordan Ellis, service desk manager.

**Who can stop it**: Jordan or the on-call lead, by switching off the integration. Tickets fall back to the manual queue.

**Who checks its work**: The service desk checks 1 in 10 tickets it sorts, about 10 hours a month.

## 5\. Cost

Measured, average of months 2 and 3.

**Volume**: 4,000 tickets a month.

**AI usage**: $2,000 a month: 5 credits a ticket x 4,000 tickets = 20,000 of the 120,000 monthly credits in the $144,000-a-year AI add-on, so a sixth of its cost. From the platform's usage log.

**New cash**: $80 a month for hosting and connections.

**Internal time**: $820 a month: $400 checking its work (10 hours x $40) and $420 upkeep (6 hours x $70 for category changes and fixes).

**Mistakes**: No direct cash cost. Fixing misrouted tickets takes time, and that time is inside the sorting and fixing hours.

**Cost to build**: $12,000, one-off.

## 6\. Promise

Written before launch. Not edited since.

**The one number**: Tickets reassigned to another team at least once. Today: 22%. Target: 12% by day 90.

**Second number**: Service desk hours a month spent sorting and fixing tickets, not counting checking the agent's work. Today: 206 (133 sorting at 2 minutes a ticket, 73 fixing 880 misroutes at 5 minutes). Target: 90.

**Must not get worse**: Major incidents sent to the wrong team (zero today), and time to first touch on P1 and P2 tickets.

**The rule at day 90**: Keep if both targets are met. Fix if it's clearly better than today but short of a target. Stop if reassignments aren't at least a fifth below today's 22% (under 17.6%), or it misroutes a major incident twice. Planned escalations from first-line to second-line support don't count as reassignments.

## 7\. Result

Measured at day 90. The numbers are in the scorecard above.

**Where it goes wrong**: 6 in 10 of its misroutes are access requests sent to the network team. The request form has no option for access, so the ticket text is all it has to go on.

**Hours**: 67 a month: 20 sorting the 600 tickets it left for people (too little text) and 47 fixing 560 misroutes (14% of all tickets). Checking its work is extra and counted under cost.

**Who misrouted**: 520 of the 560 were the agent's: 15% of the 3,400 it sorted, against 12% in the backtest. Live tickets had more access requests than the test set. The other 40 were people's.

**Must not get worse**: No major incidents misrouted. Time to first touch on P1 and P2 unchanged.

**Setup still matches this page**: Yes. Checked against its ticket system role on day 90.

**Decision**: Fix, following the rule above: reassignments are well below today's 22% but short of the 12% target. Add an access option to the request form and the category list, and check again in 60 days. At 12% or lower it becomes a Keep and moves to monthly spot checks. Because it uses a sixth of the AI add-on, the result also goes into the renewal decision. Decided by Morgan Lee, IT director, on day 92.

## Want this for one of your agents?

It works the same whether the agent is still an idea or already running. Start with a free 30-minute call, or with our [free IT ticket triage blueprint](https://agentblueprint.ai/blueprints/it-ticket-triage) or the [blank one-page template](https://agentblueprint.ai/agent-one-page-template).

[Request a free 30-minute call](https://agentblueprint.ai/#contact)[See the price and terms](https://agentblueprint.ai/#offer)