How a pilot works

Test it on your own tickets before you commit.

A Typewise pilot is a time boxed test of typically 2 to 5 use cases, with a baseline and a success threshold agreed before it starts. It answers four questions: where is the work, does it hold up in a test, does it hold up live, and is there a business case. How each one is answered, from the data period to mock data, shadow mode and integrations, is agreed in the pilot plan. It ends in a written decision.

2 to 5use cases, typically
Agreed with you up front, so the pilot tests them properly instead of everything loosely.
Day onebaseline is set
Before anything is automated, so the result can be read at the end.
Agreedcontrols per channel and use case
Including when your team must approve replies or actions. Any reduction in supervision requires your approval.
In writingthe decision at the end
Go ahead, adjust the scope, or stop.
Step by step

Four steps, in the order you would ask for them.

Each step answers one question: where is the work, does it hold up in a test, does it hold up with real customers, and is there a business case. How each one is answered, and in what order, is agreed in the pilot plan.

  1. Before the pilot

    We start with your tickets, not a demo.

    A demo shows what the product can do. It does not show where the work is in your inbox. So the first step is an analysis of your past tickets. Which period, and in what form, is agreed in the pilot plan.

    • Tickets are clustered by topic, and each topic is sized by volume, handling time, complexity and the systems it touches.
    • Topics are ranked by the minutes they cost your team, not by how often they come in.
    • Typically 2 to 5 use cases are picked for the pilot, agreed with you: enough volume to measure, clear enough to test, low enough risk to start with.
    You bring
    An export of your past tickets or contact reasons, for the period agreed in the pilot plan. It can be anonymised.
    You get
    A ranked list of use cases and a first business case built on your own ticket data.
  2. Test phase

    It is tested before a customer sees it.

    The first AI specialists are set up for the chosen use cases and tested before a real customer is involved. Whether the test runs on past tickets, on mock data or in shadow mode is agreed in the pilot plan.

    • The AI is tested on past tickets and on realistic cases with a known right answer, including the awkward ones: missing information, conflicting policies, requests that are out of scope.
    • Where a use case needs an order or billing system, a simulated one can stand in for it. A mock test runs without connecting to or changing your production systems.
    • If you prefer real traffic, shadow mode is an option: incoming requests are processed by the AI, and nothing is sent or routed.
    You bring
    Your policies and knowledge sources, the data and technical input we agree the test needs, and a few people who handle this work today.
    You get
    A readiness review: where the AI is right, where it fails, and what has to change before go live.
  3. Supervised phase

    Go live is supervised.

    Real requests, with your team in control of what goes out. Supervision is where a pilot learns the most.

    • Before live testing, we agree the controls for each channel and use case, including when your team must approve replies or actions.
    • What reviewers change, and why, is recorded. Recurring changes point to a knowledge or policy gap, which is fixed and tested again.
    • A weekly review looks at quality, handover reasons and open decisions. Any reduction in supervision requires your approval.
    You bring
    The people who review and approve where the agreed controls require it, with time set aside.
    You get
    Evidence from real traffic, by use case, week by week.
  4. Final review

    The pilot ends in a business case, then a decision.

    The pilot closes on a date that was agreed at the start. The measured results are set next to the baseline and the threshold, and then extended into a business case.

    • Results are read by use case, on the same channel and the same mix of cases as the baseline. Human time spent on review, corrections and handovers is counted as well.
    • The business case compares today with the next 3 to 6 months and the next 1 to 2 years: incoming volume, work done by AI and by people, hours freed, and the costs.
    • Measured results and longer term scenarios are kept apart, with every assumption written down. Both sides then agree on the decision in writing: go ahead, adjust the scope, or stop.
    You bring
    The person who decides, in the room, and the figures your finance team works with, such as the cost of an hour.
    You get
    A business case with net value, ROI and payback, a written decision and, if it is a yes, the plan for production.

A pilot is time boxed. The dates for each phase, and for the decision, are agreed in the pilot plan before it starts.

Step one, up close

Count minutes, not tickets.

A simple question takes a person a few seconds to answer, and it takes the AI a few seconds too. A complaint or an exception can take a person fifteen minutes.

Rank your topics by ticket count and you automate the cheap ones. Rank them by the minutes they cost and the business case looks different, and so does the pilot.

Share of tickets
Share of handling time
  • Simple questions
  • Order and account changes
  • Complaints and exceptions
An illustration, not customer data. Your own split comes out of the ticket analysis.
Step two, up close

Do we have to open our systems for a pilot? Not for a test on mock data.

Where the pilot plan includes a mock test, it runs without connecting to or changing your production systems, so it can start while your security review runs. Which data and technical input it needs is agreed with you in the pilot plan.

What a mock test uses

  • Past tickets, or realistic test cases written with the people who do the work
  • A simulated order or billing system, where a use case needs one
  • Your policies, described in plain language
  • Test accounts, so nothing can reach a customer

What it does not touch

  • Your real customers
  • Your production systems: nothing is connected, nothing is changed
  • Live traffic, unless shadow mode is agreed
  • Anything that is not in the pilot plan

When real systems come in

  • When the pilot plan says so, and not before
  • Usually read access first, for lookups such as an order status
  • Actions one by one, each with an approval rule you set
  • Access is agreed with your IT and security owners

What a customer service AI should be allowed to do →

Step four, up close

From measured results to a business case.

A pilot measures a few use cases over a few weeks. The business case takes those results and shows what they would mean over the next months and years, with every assumption stated.

Illustrative exampleTodayBaseline, measured before launchNext 3 to 6 monthsThe pilot use cases, from measured resultsNext 1 to 2 yearsScenario, built on stated assumptions
The work
Incoming requests per month10,00010,00010,000
Completed fully by the AI02,0004,500
Partly done by the AI, finished by a person01,0001,500
Handled by people10,0007,0004,000
The hours
Human hours per month, review and remaining work included1,000767513
Hours freed per month0233487
Value of that capacity, at $45 an hour$0$10,500$21,900
The costs
Resolution units billed per monthA full resolution counts one, a partly done one counts half, a ticket handed straight to a person is free: 2,000 + 1,000 at half = 2,500, and 4,500 + 1,500 at half = 5,250.02,5005,250
Typewise platform and usage, per monthThose units at $2.00, the published pay as you go rate, plus the $599 Growth plan fee.$0$5,599$11,099
Your team's time to run it, per month$0$900$1,350
The result
Net value per month$0$4,001$9,451
ROI on the monthly cost–62%76%

The one off setup effort of $7,200 is earned back in the second month, at the net value of the first 3 to 6 months.

An illustrative example with made up figures. It is not customer data and not a quote. Your own case uses your volumes, your handling times and your costs.

The cost lines use the published plans and rates. See the plans and rates →

The assumptions behind it

  • Volume stays at 10,000 requests a month. A person needs 6 minutes per request today.
  • A request the AI completes fully still costs 0.5 minutes of review. A partly completed one costs a person 3 minutes. A handed over one costs the full 6.
  • One hour of your team is valued at $45, fully loaded. Running the AI takes your team 20 hours a month at first and 30 later.
  • Setting it up takes your team a one off effort of 160 hours, or $7,200.
  • Resolution units are billed at $2.00, the published pay as you go rate, on top of the $599 Growth plan fee, so every cost line can be checked against the price list. Prepaid packages bring the rate down to $1.00, which makes this the upper end of the list. Your plan and rate are agreed in the pilot plan.

Hours freed are capacity, not cash

They turn into savings only if you decide so: fewer overtime hours, growth without new hires, or more time for the cases that need a person.

Measured and projected stay apart

The middle column rests on what the pilot measured. The right column is a scenario. If an assumption changes, the number changes with it.

Revenue can be part of the case

Where service drives sales, faster answers and a higher CSAT can lift conversion. CSAT is measured in the pilot from your own post contact survey, on the same channel and the same use cases as the baseline, and it enters the business case only where that lift can be measured.

What the pilot measures

If the baseline is not set before launch, nobody can read the result at the end, however good it feels.

What is measuredBaseline, before launchHow it is read at the end
ResolutionTickets the AI resolves fully, resolves partly, or hands to a person.How the same use cases are handled today.By use case, with the same definitions the invoice uses.
Handling timeMinutes, or touches, a person spends per ticket.From your helpdesk, or timed by hand on a sample if it is not recorded.Net of the time spent on review, corrections and handovers.
Response speedTime to the first reply.From your helpdesk reports, per channel.On the same channel, in the same period of the week.
QualityCorrect answers and actions, customer satisfaction, reviewer scores.Your current satisfaction score, where you have one.By use case and by how serious an error is, never as one average.
HandoversWhy a ticket went to a person.Not needed.As a to do list: what to fix, and what should stay with people.

Same channel, same mix of cases

A before and after only counts when both sides handle the same kind of work at the same quality bar.

Human work is counted too

An approved draft still cost someone a minute. That minute belongs in the result.

An average can hide a failure

A good overall number can sit on top of one use case that goes wrong. Each one is read on its own.

The pilot counts a resolution the way the contract does. What counts as a resolution →

If it misses the number

What happens when it misses the number.

The threshold is written down before the pilot starts, so a miss is visible to both sides. A pilot ends in a written decision: go ahead, adjust the scope, or stop.

Outcome 1

Go ahead

When
The use cases work, quality holds and the numbers support the business case.
What happens next
The plan for production is agreed: scope, owners and dates.
Outcome 2

Adjust the scope

When
The original use case does not carry the case, or one specific question is still open, such as an interface.
What happens next
A smaller or different test is agreed, with its own threshold and decision date.
Outcome 3

Stop

When
The evidence does not support going further.
What happens next
The pilot closes with what was learned written down.

Once you are live, the bill follows the result as well. Typewise charges per resolution: a ticket handed straight to your team costs nothing, and a reply you flag as wrong is not charged.

See how a resolution is counted →
Before it starts

Who to bring in before the pilot starts.

A pilot that works can still get stuck when procurement, security or the works council hear about it at the end. Bring them in from the first week.

And pick the weeks with care.

A pilot needs enough real traffic to learn from, so a quiet period tells you little. Your peak weeks are the wrong moment as well. Start so that the supervised phase falls in a normal stretch of the year.

  • The decision makerOwns the decision at the end and agrees the threshold at the start.
  • A project lead on your sideRuns the pilot day to day and keeps the people involved in the loop.
  • Reviewers from the teamApprove replies in the supervised phase and say what good looks like.
  • IT and securityDecide which systems and data are in scope, and when. Most reviews start from the trust center.
  • Procurement and legalHear about the pilot early, so approval does not hold up a good result. What to prepare for approval.
  • The works council, where you have oneHears about it early, not after the result.
Two ways to test

Small team? You may not need a pilot at all.

A structured pilot is for teams with several systems, several stakeholders and a review to pass. Its scope and terms are agreed separately from the free trial. If that is not you, the free trial is the faster test.

For small teams

Free trial

Connect your inbox and see the AI resolve real requests the same day. The right test when one person can decide.

  • 14 days, no base fee and no credit card
  • 100 resolution credits on your real tickets
  • You set it up yourself, in plain language
  • No sales call unless you want one
For larger teams

Structured pilot

Planned with you, for teams with several systems, several stakeholders and a compliance review to pass.

  • Starts with an analysis of your past tickets
  • Can be tested on mock data before it touches a real system
  • Baseline and success threshold agreed up front
  • Ends in a written decision on an agreed date

What buyers ask before a pilot.

Yes, on two routes that differ in what they cost you. The free trial is free: you connect your inbox and see the AI work on real requests the same day, which suits a small team. A structured pilot has its own scope and terms, agreed with you before it starts, and it is the route for larger teams: an analysis of your past tickets, a test phase, a supervised go live and a measured decision at the end. To compare vendors on the same tickets, use the open test kit.

Not for a mock test. It runs without connecting to or changing your production systems: the AI is tested on past tickets and realistic test cases, and a simulated system can stand in for your order or billing system. Which data and technical input the test needs is agreed in the pilot plan. Real systems are connected when the pilot plan says so, with read access first and each action approved by you.

A pilot is time boxed. The dates for the test phase, the supervised go live and the final decision are agreed in the pilot plan before it starts, and they depend on the use cases you pick.

Typically 2 to 5, agreed with you before it starts. The best first use cases have enough volume to measure, a clear process and low to medium risk. More are added after the pilot, based on what it showed.

With a baseline that is set before launch. Handling time is taken from your helpdesk, or timed by hand on a sample when it is not recorded. At the end the same use cases are measured on the same channel, and the time your team spent reviewing and correcting is counted in.

The threshold is agreed in writing at the start, so both sides see a miss. The pilot then ends in a written decision: go ahead, adjust the scope, or stop.

Not at first. In the test phase it does not reach customers at all. Before live testing, we agree the controls for each channel and use case, including when your team must approve replies or actions. Any reduction in supervision requires your approval. How a staged go live works.

An export of your past tickets or contact reasons for the agreed period, your policies and knowledge sources for the chosen use cases, a decision maker, a project lead who runs the pilot day to day on your side, and the reviewers the agreed controls require.

The free trial is free. A structured pilot has its scope and terms agreed separately, before it starts, together with the baseline and the threshold.

Start with your tickets.

Book a call and we plan the pilot around your past tickets: where the minutes are, which use cases to test first and what the pilot has to prove.