Portfolio/The business case
Across the four use cases

Which one to build first.

Type in how many people work at your company and the table sorts itself.

The numbers

Two numbers from you, and the table does the rest.

Use caseWhat you get backWorth a year
Get helpHours back for the IT team$136,500
Time offTime back for managers and staff$81,250
Policy searchHours back for the HR team$73,125
Parental leaveA wrong answer that never gets sentNo hours saved
Total a year$290,875
Minus what the AI costs to run−$2,500
Total, before you pay to build it$288,375

Get help comes out on top, and it probably would at your company too. There are simply more IT tickets than policy questions, and each one eats more of someone's day.

Time off looks like second place. Be careful with it. What it saves is fifteen minutes here and there for managers and employees. That time is real, and it almost never turns into money, because nobody's headcount changes when a hundred managers each get an afternoon back over a year.

Parental leave gets no number at all, because hours saved is the wrong way to value it. This is the question where a wrong answer costs the most. Tell someone in India she gets 16 weeks when the law says 26, and you have a legal problem and an employee who has planned her year around it. Rank it on time saved and it comes last every time, which is how a use case like this never gets built.

The total assumes you built all four and ran them for a year. It leaves out two things. The time-off row is that soft manager time, so the part finance can really count on is Get help plus Policy search. And it doesn't include what it costs to build any of this, which is the number that actually decides whether you do it.

Where these numbers come from, and how to change them

How often it happens, per person per year

Minutes it takes someone today

How much of it the assistant handles

What an hour of their time costs

These are starting points, not benchmarks. Volume and minutes move the order most, so they are the two worth chasing down properly. Where you cannot get a real number, put in the low end of what you would guess and see whether the ranking still holds.

Cost

What it costs to run isn't the problem.

The demo records what every message costs, so these are real numbers rather than estimates. I ran six of them through Claude, once each, so treat them as rough. Under them is the bill for a year, worked out two ways: everything cheap, and everything expensive.

What one conversation cost
A policy questionThe only AI call is the small model that works out where to send the question. The answer itself comes straight out of the policy documents.$0.0002
A parental leave answerTwo calls: work out who is asking, then read the global policy and the country one and settle the difference.$0.0342
A vacation booked, three turnsEight calls, across working out the dates, picking one, and filing the request.$0.3032
The AI bill for a year · at 50,000 conversations
If it were all policy questionsJust the routing call. Close to what policy search on its own would cost.$10
A realistic mixMostly the cheap kind, some of the expensive kind.$2,500
If every one were the expensive kindThe worst case. No real company looks like this.$15,160

Even the worst case is small next to the support costs in the table. The people cost far more: an engineer, someone to keep the policies current, someone to look after the tests. The pattern is worth noticing though. Cheap questions get answered by the cheap parts of the system, and only the conversations that plan something and write something cost real money, which is also where most of the value is. I also built a version that runs with no AI at all, so the 67 tests run free on every change and the public demo needs no key.

The method

How to do this for the tools you're looking at.

The table above is one worked example. The steps matter more than my numbers. Every vendor will give you a calculator. None of them will give you this.

  1. Decide what one of these is worth before you measure anything.

    What does one of these being handled actually save you? A ticket nobody has to answer. An hour of someone's day. A mistake that never happens. A form nobody has to fill in twice. Pick the wrong one and every number after it is neat and useless. The four above save four different things, which is why they can't really share a column.

  2. Work out which savings are actually money.

    If the IT team handles a thousand fewer tickets, that can turn into a role you don't have to backfill, and finance can act on it. If four hundred managers each save ten minutes a week, that's real time and it never shows up in anyone's budget. Count both, label which is which, and don't let the soft one carry your recommendation. In the table above, second place is the soft one, and that's the most common way these business cases get oversold.

  3. Don't score the safety one on time saved.

    There's usually one on the list that isn't there to save time. It's there to stop something expensive going wrong. Score it the same way as the others and it finishes last, so it never gets built. Say what the mistake would cost you instead, and put that next to the table rather than in it.

  4. Decide what counts as handled, and count it late.

    A vendor counts it the moment the assistant answers. Count it a week later, and only if nobody opened a ticket about the same thing. Most of the difference between a business case that holds up for a year and one that falls apart in three months is hiding in that gap.

  5. Give a range, not one number.

    Run it twice. Once assuming everything goes the cheap way, once assuming everything goes the expensive way. If the decision still makes sense at both ends, you have an answer. If it only works at the good end, you have a hope. The cost section above does this in one table.

  6. Bring the objections yourself.

    Whatever you leave out, someone in the room will bring up, and it sounds worse coming from them. Three that apply to almost anything like this:

    Calling it a win too early
    If a question counts as handled the moment the assistant replies, you're also counting every answer it got wrong. Those come back as tickets a week later, from someone who is now annoyed.
    Time that never turns into money
    Minutes saved for employees and managers are usually the biggest number on the slide. Nobody's budget changes because four hundred people each got nine minutes back.
    A percentage of a pile that keeps growing
    Handling 35% of tickets doesn't mean the pile gets smaller. It only shrinks if someone fixes the causes and writes them down, which is what Get help does. And an assistant that's easy to ask gets asked things people would never have bothered HR with. Good for them, and not a ticket you saved.

Doing this properly, with your real numbers and the vendors you're actually looking at, is the first half of an Evaluation Sprint.

The four use cases, each with the number it answers for, are chapters 01 to 04. How I'd know any of this is true: the evals →