Unofficial. A job application for Cresta’s Forward Deployed Product Manager, AI Agent role, built from public pages. Not affiliated with Cresta.

One AI agent, from scoping to go-live and after.

Cresta’s posting for Forward Deployed Product Manager, AI Agent says: “you own each AI Agent and the outcomes they drive for our customers.” This is how I would own one. I built it in one day from Cresta’s public pages, with AI agents doing the reading. The customer in it is made up.

The bottom lineCresta already publishes how it scopes, tests and releases an agent. I would own the step between the test results and the customer’s decision.
So the order isAgree what counts before anything is built. Test by use case. Launch only what passes. Measure the way Cresta’s own playbook says to.
Done meansThe customer’s sponsor can say what the agent does, what it hands to a person, and which number decides the next phase.

Who is in the room, stage by stage

The posting names the lifecycle: pre-sale scoping, kickoff, launch and post-go-live optimization. This is one voice agent moving through it for Pinemoor Lending, a made-up lender. Pick a person to see what each stage asks of them.

Stage 1

Scope

Before the contract is signed.

A one-page scopeThe use cases that are in, and the ones that are out.
Each outcome measureOne sentence on how it is counted and what is left out.
A baseline for eachOr the words “none yet”.
The gate out of this stageThe sponsor signs the scope and names who owns the agent once it is live.
Stage 2

Kickoff

The first two weeks.

Interviews with every stakeholderTurned into one requirements list.
Each requirement a yes or no checkMarked must-have or nice-to-have.
The go or no-go rulesWritten down before the build starts.
The gate out of this stageRequirements and gate rules agreed in one workshop.
Stage 3

Build and test

Until the gate rules are met.

The agentBuilt with the Forward Deployed Engineer.
Test conversations by use caseSimulated callers, including bad lines and noise.
A weekly session on failing turnsOperators judge them, one at a time.
The gate out of this stageEvery must-have has a test, and every grader has been checked against people.
Stage 4

Go or no-go

Launch week.

A gate reviewDecided use case by use case.
A rollback drillRun before the first customer call.
A small launchA share of calls, one flow at a time.
The gate out of this stageGo, per use case. A use case with no data does not launch.
Stage 5

First weeks live

The first six weeks after launch.

The outcome numbersCounted the way the scope said.
Failures turned into testsSo the same one cannot ship twice.
The next use caseChosen from why calls still transfer.
The gate out of this stageA review with the sponsor: widen, hold or stop.

Pinemoor Lending and its people are made up. The stage names follow the posting. The practices I credit to Cresta come from its public pages. Sources are on the evidence page.

The gate rules

Five rules for the go or no-go decision. The first is from the planning board I built for my current program. The second is Cresta’s own.

No data shows a dash, never green.A green light nobody checked is worse than no light.
A failed must-have holds the use case.Whatever the pass rate. Cresta’s testing separates must-have from nice-to-have for this reason.
An unchecked grader counts as no data.If no person has checked the grader, its passes prove nothing yet.
The decision is per use case.What passes can launch while the rest waits. Scope shrinks before the date moves.
The rules are agreed at kickoff.Before anyone has a result to defend.

The plan, stage by stage

What ships, who it involves, and how I will know. Financial services is an industry Cresta names, and its own playbook for lenders is the base I built on.

Stage 1. Before the contract is signed

Scope one agent, and say what counts

Cresta’s pilot guide says a pilot with no written success criteria ends in debate, not a decision. Scoping is where that is prevented.

  • Pick the use cases from the customer’s own calls.Cresta’s Automation Discovery ranks call topics by how often they come up and how they resolve. The scope starts there, not from a wish list.SponsorOperationsSolutions architect
  • Define each outcome measure in one sentence.How it is counted, and what is left out. The list is further down this page.SponsorOperations
  • Write the baseline, or write “none yet”.If the lender cannot say today’s transfer rate, measuring it is the first task.OperationsIT
  • Name the owner and the stop rule.Who owns the agent in production, and what result would make us stop. Both are on Cresta’s own pilot checklist.Sponsor

How I will know. The sponsor can say in two sentences what the agent will do, what it hands to a person, and which number decides the next phase.

Stage 2. Kickoff, the first two weeks

Interview everyone, then write one list

This is how I start every program. On my current one, ten vendor teams planned in ten ways until I interviewed every lead and turned the answers into one script.

  • Interview every stakeholder before the workshop.The sponsor, the supervisors, the agents who take these calls today, the quality reviewers, compliance and IT. The same questions for each.Everyone
  • Turn the answers into one requirements list.Each requirement a yes or no check, marked must-have or nice-to-have. That is the form Cresta’s testing takes as its input.ComplianceOperationsEngineer
  • Agree the go or no-go rules in the workshop.Before the build, while nobody has a result to defend.SponsorCompliance
  • Publish the launch-readiness board.Every line shows a dash on day one. That is the honest starting state.Everyone

How I will know. Every must-have traces to the person who asked for it, and the sponsor has signed the gate rules.

Stage 3. Build and test

Build with the engineer, test by use case

The posting pairs this role with a Forward Deployed Engineer. The engineer builds. I keep the scope, the customer and the evidence straight.

  • Test conversations grouped by use case.Simulated callers drawn from the lender’s own conversations, then the same calls on a bad line with noise and interruptions.Engineer
  • Check the graders against the customer’s own reviewers.A grader nobody has checked is not evidence. Cresta scores each one against human review.OperationsEngineer
  • A weekly session on failing turns.Show one turn and ask one yes or no question. It is Cresta’s own published lesson, and it keeps operators in the room.OperationsCompliance
  • Check the numbers myself.On one of my builds, an agent reported 682 tests. A stale copy had inflated the count. The real number was 384.Me

How I will know. Every must-have has at least one test, every grader has a score against people, and I have read the failing transcripts, not only the pass rates.

Stage 4. Launch week

Go or no-go, then a small launch

I have run go or no-go decisions through build, test, acceptance and cutover. The rule I carry over: a gate with no data shows a dash, never green.

  • A gate review, use case by use case.What passes can launch. What fails, or has no data, waits. The date holds for the rest.SponsorComplianceOperations
  • A rollback drill before the first customer call.Cresta’s releases roll back in one step. The drill proves the people know how.EngineerIT
  • Launch on a small share of calls, one flow at a time.Cresta’s financial services playbook asks for this: a small change you can watch closely, and you always know what changed.EngineerOperations
  • Pick the day from the billing calendar.The same playbook warns against the Monday after a Friday or Saturday billing cycle.Operations

How I will know. The decision is recorded per use case with the evidence beside it, and supervisors have watched the first live calls.

Stage 5. The first six weeks live

Measure honestly, then widen

The posting says to own the outcomes and to find more to automate. Both start with numbers the customer trusts.

  • Report resolution beside containment.Cresta’s playbook calls containment the easiest number in a contact center to inflate. After-hours calls come out, and a hang-up is not a win.SponsorOperations
  • Turn every production failure into a test.A flagged conversation becomes a test case, so the same failure cannot ship twice. Cresta’s testing does this in one step.OperationsEngineer
  • Choose the next use case from why calls still transfer.The transfer reasons are the map for the next phase.OperationsSponsor
  • Write it into the playbook.What I would do again, what I would not, and what the platform could do for the next team. The posting asks for this by name.ProductForward deployed team

How I will know. The sponsor reads one page, recognizes every number on it, and decides to widen, hold or stop.

I have put no total length on this plan. One Cresta customer wrote that its first agent took four months, and Cresta says its newer tools cut a first deployment roughly in half. The real figure comes from Cresta’s own deployments, which I cannot see.

What I would cut

  • A second channel at launch.Voice first, chat later. One flow at a time is Cresta’s own advice.
  • Hardship and dispute conversations.They go to a person by design. Cresta’s pilot guide says the same about high-emotion calls.
  • Nice-to-have requirements that threaten the date.They stay on the list and off the gate.
  • New reporting.Cresta’s dashboards exist. The board in my prototype is one page for one decision.

What would change this plan

  • If the agent cannot reach the lender’s systems in time,the first launch answers questions only, and payments wait. One Cresta customer wrote that integrations took real IT work.
  • If Cresta already shows customers a readiness view,I use it. The rules matter more than the page.
  • If the customer will run the agent itself,the plan becomes a hand-over. Their team runs the gate with me watching, then without me.
  • If the baseline shows the calls are not what the sponsor thinks,the scope changes before the contract does.

How I would measure it

Seven measures, each defined at scoping and counted the same way every week.

I can see no baseline for any of them from outside, so there are no targets on this page. Setting the baselines is stage 1 work. The definitions follow Cresta’s financial services playbook where it gives one.

  • Containment.The share of conversations the agent handles without passing to a person. After-hours calls are left out, because nothing can transfer then.
  • Resolution.The share where the customer’s issue was finished. A hang-up is not a resolution.
  • Transfers, with reasons.How many conversations go to a person, and why. The reasons choose the next use case.
  • Wait after a transfer.How long a customer waits once the agent hands them on. Cresta’s playbook counts transferred calls that wait under 90 seconds.
  • Minutes handled by people.Total talk time for the human team. The playbook prefers it to handle time per call, which rises when the easy calls leave.
  • Customer satisfaction.Asked of the customer, or predicted from the conversation. Reported next to containment, never instead of it.
  • Must-have pass rate on live calls.The same checks that gated the launch, run on real conversations.