Work Out What Your Own AI Model Would Cost to Run

Turn a list of the jobs you want an AI model to do into an estimate of hours, a monthly running cost, and a straight answer on whether you need your own model at all.

--- SCOPE ---

SITUATION: I run or manage a [DESCRIBE THE BUSINESS, e.g. eight-person accountancy practice, family support charity, engineering firm] and I am working out whether to run an artificial intelligence model on our own server instead of using a subscription or a connection to one of the large providers.

Here is what I want the model to do:

- Job one: [DESCRIBE THE JOB, e.g. read the week's case notes and write a summary report]
- How often job one runs: [E.G. ONCE A WEEK ON A FRIDAY EVENING, OR TWICE A DAY]
- Roughly how much text goes in each time: [E.G. 20 CASE NOTES OF ABOUT A PAGE EACH]
- Job two, if there is one: [DESCRIBE IT, OR WRITE NONE]
- How often job two runs: [E.G. DAILY, OR LEAVE BLANK]
- Roughly how much text goes in each time: [E.G. 40 EMAILS, OR LEAVE BLANK]
- Does anyone need to ask the model a question and wait for the answer: [YES OR NO]
- The most sensitive information involved: [DESCRIBE IT, e.g. client health records, payroll data, nothing sensitive]
- Whether our own client contracts restrict where data is processed: [YES, NO, OR NOT SURE]

CHALLENGE: I need to know whether running our own model is justified here, and if it is, roughly what it would cost each month. I do not want a technical specification. I want a number I could put in a budget and defend, and an honest answer if the whole idea is wrong for us.

EVALUATE: Work through this in order.

1. SHOULD I DO THIS AT ALL. Look at the sensitivity of the information and any contract restrictions I have named, and tell me plainly whether running our own model is justified or whether a subscription or a paid connection to a large provider would do the job. Running your own model controls where data is processed. It does not save money and it does not improve quality. If those are my reasons, say so and stop there.

2. BATCH OR LIVE. Tell me which of the two shapes my jobs actually are. Work that runs on a schedule with nobody waiting is far cheaper than work where a person asks a question and waits for an answer. If I have said somebody needs to wait, say clearly that this changes the price and treat it as a separate decision rather than folding it in.

3. HOURS OF MODEL TIME. For each job, estimate how long one run would take and multiply by how often it runs, to give me hours per week and hours per month. Show the arithmetic in plain numbers so I can argue with it. State every assumption you had to make about speed, and say which assumptions would change the answer most if they turned out to be wrong.

4. WHAT THAT COSTS. Give me a monthly figure for two options. First, a server sized to run the model all the time whether it is working or not. Second, a smaller server that rents a graphics card only for the hours you calculated above. Use publicly published cloud hosting prices in pounds, name the assumptions behind each, and show the point at which the second option stops being cheaper than the first.

5. WHAT WOULD BREAK IT. The three things most likely to make the real bill higher than this estimate, and one sentence each on what to put in place to catch them early.

6. WHAT TO ASK A SUPPLIER. Five questions I should put to any hosting provider or software supplier before I commit, written so I could paste them straight into an email. Cover which model, whose machines, which country the data is processed in, how long it is kept, and whether my data is used to train anything.

Write it in plain English for someone who runs a business and does not work in technology. Use pounds. Where you are estimating rather than calculating, say so in the same sentence. If the honest answer to question one is that I do not need my own model, tell me that first and keep the rest short.