Every renewal email in your inbox makes the same quiet claim. The higher tier pays for itself. The claim arrives with a feature list, a discount for paying a year ahead, and a deadline that lands before your team has time to test any of it.
Direct answer: Pay the premium only where the tool holds context your firm cannot rebuild at a fair cost. Run 4 checks before the renewal date. Test the context the tool holds, price the rebuild in hours, count real use over 30 days, and price the exit. A tier that passes all 4 earns its price. A tier that fails 1 goes back to the basic plan, or the job moves to a small build you control.
Money leaves a firm in 2 ways on this decision. The first is the seat price, printed on the invoice in CAD on the day the card runs. The second is the rebuild cost nobody prints. That second cost decides the question, and it sits inside your own records rather than inside a vendor pitch.
What does the premium actually buy?
Strip a premium tier down and it sells 4 things. It sells memory, in the form of saved projects, history and settings your team built up over months. It sells reach, in the form of connections to the other systems your firm runs. It sells room, in the form of higher limits on use. It sells cover, in the form of admin controls, retention settings and support terms.
Room is the easiest part to price and the least valuable part to overbuy. A limit your team never reaches has a value of zero, however large it looks on a pricing page. Cover has real value in a regulated firm, and the value shows up in an audit rather than in daily work.
Memory and reach carry the decision. Memory is the context the tool holds about your clients, your files and your past work. Reach is the set of live connections that feed that context each day. A tool with deep memory and live reach is a system of record in practice, whatever the contract calls it.
The audit in One prompt finds every AI tool you are wasting money on gives you the starting list for this test. That audit counts the seats and the duplicate tools. This test takes the survivors and asks a harder question about each 1. What would it cost us to walk away?
How do you test the context a tool holds?
Open the tool and list what it knows that your firm taught it. Count the saved projects. Count the custom settings. Count the templates your team built and the history a new hire reads to get up to speed.
Then ask the rebuild question in hours. A person on your team has to recreate each item inside a replacement. Price those hours at the rate you pay that person. Write the total down in CAD beside the annual premium for the same tool.
The comparison runs in a plain table, and the table is the kept asset from this piece. Build it once and reuse it at every renewal.
| Check | What you count | The premium passes when |
|---|---|---|
| Context held | Saved projects, settings, templates and history inside the tool | The rebuild takes more hours than the premium costs in a year |
| Reach | Live connections to your other systems | At least 2 connections feed the tool fresh data each week |
| Real use | Days in the last 30 with a real task run by a named person | Named users run real tasks on most working days |
| Exit | Export of your data in a form another tool can read | A full export lands in your hands inside 1 working day |
A tool that holds a week of rebuild work and costs a fraction of that week in premium has made its case. A tool that holds an afternoon of settings and charges a year of seats has failed it. Most firms find both kinds on the same invoice, which is why the test runs per tool and never across the whole stack at once.
Where does the build side win?
A build wins where the job is narrow, stable and close to your own data. A weekly report that reads 1 system and writes 1 page. A sorting job on records your firm owns. An intake form that checks 3 fields and routes the result to a named person.
Jobs like these change slowly. The rules were set a year ago and they still hold. A small build on top of a basic model plan handles them, and your firm owns the logic, the prompts and the records it produces.
The billing shape in Who pays when Claude runs inside 19 other tools? shows why ownership matters here. When a tool runs inside other tools, the meter runs in places a finance review can miss. A build you control has 1 meter, and it sits in an account your firm can read line by line.
A build loses where the job changes each month, where the vendor carries the compliance work, or where your team has no person who can maintain what gets built. A build with no owner becomes a risk the day the person who wrote it moves on and nobody else can read it.
How do you count real use before a renewal?
Vendors report seats. Seats tell you what you bought. They say nothing about what your team used.
Count tasks instead. Pick the last 30 days. For each premium tool, count the days a named person ran a real task that reached a client, a file or a decision. A login with no task counts as zero. A test prompt counts as zero.
The review discipline in Count accepted comments, not posted ones, before you renew your AI review tool makes the same point in a coding shop. Output is cheap to count and easy to inflate. Accepted work is the honest unit. The same rule holds for every category on your invoice.
Then divide. Take the annual premium in CAD and divide it by the count of accepted tasks in a year. That cost per accepted task is the number to compare against the build estimate and against the basic tier. Run the division on a full year where you can, because a launch month flatters every tool.
What does the exit cost, and when do you price it?
Price the exit before you sign, while the vendor still wants the signature. Ask for a full export of your projects, your history and your settings, in a form another tool can read. Run the export. Open the files. A promise of an export is a claim. A file on your own drive is a fact.
Canadian firms carry a second exit question, and it sits in the vendor’s answers rather than on its pricing page. Ask where the data rests, how long the vendor keeps it after you leave, and what proof of deletion you get in writing. The answers belong in the renewal file beside the price.
A vendor who answers these questions in plain writing has told you something useful about the next 3 years. A vendor who routes every question to a sales call has also told you something. Believe the second answer as much as the first.
How should a Canadian owner run this test this quarter?
Pick the 3 largest AI lines on your next renewal. Give each 1 the 4 checks in the table. Context held, reach, real use, exit. Score each check pass or fail in writing, with the count or the cost that decided it.
Renew the tiers that pass all 4. Drop the tiers that fail the use count to the basic plan for 1 quarter and watch the complaint level. Move 1 narrow, stable job to a small build with a named owner, and hold the build to the same cost per accepted task at the 90 day mark.
Recap: The premium earns its place only where a tool holds context your firm cannot rebuild at a fair cost. Price the rebuild in hours, count accepted tasks over 30 days, test the export before you sign, and get the data answers in writing. Run the 4 checks on your 3 largest AI lines before the next renewal date, and let the counts, in CAD, make the call.