Self-assessment

Before you buy AI seats for the firm

Only 11% of organizations can forecast their AI spending to within 10%.

Mavvrik and Benchmarkit, “State of AI Cost Governance 2026,” July 29, 2026; 396 organizations.

Buying AI seats for everyone is the easy part. Knowing the total cost is harder. Settle these six questions first, and you’ll know what the seats are for, what they’ll cost at month twelve, and how to tell whether they’re working.

Each question ends with a prompt to try. Swap the brackets for your own details, and keep client data and proprietary information out until you’ve settled question four.

  1. Where does the time actually go?

    Pick three to five workflows that eat hours every week. For example: pulling numbers from PDFs into tax forms, or finding context buried in email. Time them now, before anything changes; a rough weekly estimate is enough to start. If one tool serves several workflows, split its cost roughly. Without that baseline, nobody can tell you later whether AI helped.

    Write down: each workflow, who does it, hours per week, the software it runs on and its cost, turnaround, and how often it’s reworked.

    Try this prompt:

    I run a [type of firm] with [number] people. Here is a rough list of what my team spends time on in a typical week: [paste the list]. Group it into workflows, and mark the ones that are repetitive and heavy on documents or email. Then give me a table I can fill in for two weeks: workflow, who does it, hours per week, software used and its monthly cost, turnaround, times sent back for rework.

  2. Which platform, chosen on the merits?

    Claude, ChatGPT, Gemini and Copilot aren’t interchangeable, and neither are the models behind them. Test the plan you would actually buy, not the free version, on a workflow from question one. Use made-up details, not client data, and give each platform the same context: the task and an example of finished work. Note the model you used in each, and keep it for both runs. Then check what the output won’t show: whether the platform can reach where the work lives, such as email and shared files, and whether you can control who sees what. A suite you already pay for doesn’t decide it: in one engagement, Microsoft’s own tools meant more cost and complexity for what the firm needed.

    Write down: the platform, plan and model, what it was tested on, and why it won.

    Try this prompt in each platform, with the same material:

    I work at a [type of firm]. The task: [one workflow from question one]. Here is finished work we were happy with: [an example, with made-up details]. Here is the material: [your template, filled with made-up details]. Do the task the same way, then list anything you weren’t sure about or assumed.

    Run it twice in each, since answers vary. Which got the facts right with the fewest edits?

  3. What will it cost at month twelve?

    Seats are the cost you see. Add-ons, any usage billed beyond the seat, and the hours people spend learning are the rest. Ask the vendor about minimum seat counts and annual commitments, what happens if someone hits a usage limit, and who sees that spend. Set whatever spend controls the platform offers before rollout. Buy the subscriptions in the firm’s name, so the firm owns the accounts.

    Write down: the twelve-month cost, the controls, and who approves more.

    Try this prompt:

    We plan to buy [number] seats of [platform and plan] at [price per seat per month, from the vendor’s pricing page]. Build a twelve-month cost table: seats, these add-ons: [list], and training time of [hours] per person for [number] people at [internal cost per hour]. Show the monthly and total cost, and what changes if we add [number] seats in month six. Show your working.

    Check the arithmetic before it goes to your partners.

  4. What goes in, and what stays out for now?

    An assistant is only as useful as the material it can draw on. Start with the firm’s own: templates, policies, how-to notes. Client matters can come later, once that works. Before any client material goes in, read the plan’s terms on training and retention, and check what your engagement letters promise clients about their information. Put the rules on one page people will actually follow.

    Write down: what’s in, what’s out for now, what to do if something goes in by mistake, and who can change the list.

    Try this prompt:

    Draft a one-page guide for our staff on using [platform] at work. What can go in: [list]. What stays out for now: [list]. Who to ask when unsure: [name]. What to do if something goes in by mistake: [steps]. Who can change the list: [name]. Plain language, one page, no jargon.

  5. Who goes first?

    Not everyone needs a seat on day one. Start with the people who do the workflows you measured. People learn fastest on their own work, so train them by role, not in one general session. Name one internal champion, the person colleagues go to with questions, and set aside hours for it. A champion with no time for it is just a title.

    Write down: the first group, the champion, and their hours.

    Try this prompt:

    Here are the workflows we timed and who does each: [paste the table]. Suggest a first group of [number] people to start with, and say why. Then write a one-paragraph brief for our internal champion, who has [hours] a week set aside: what they do each week, and what they report to the partners.

  6. How will you know it worked?

    Time the workflows from question one again at six weeks, then monthly. Report hours, turnaround, rework and any software you’ve dropped, not logins. Time saved on a first draft that then needs heavy fixing isn’t time saved. If a workflow didn’t move, say so and drop it.

    Write down: the date of the first read, and who reports it.

    Try this prompt:

    Here are our workflow times from [date] and from [date]: [paste both tables]. For each workflow, show the change in hours per week, in turnaround and in rework. Say plainly which moved and which didn’t. Finish with a three-line summary for the partners. Show your working.

    Check the numbers against your own tables before you share them.

Can’t answer all six?

These six questions follow the work of AI Enablement for Knowledge Work. If you can answer all of them, you may not need us. If you can’t, that’s the work.

The prompts here are starting points. In the engagement, we build a working prompt library around your firm’s own work.

4–6 weeks. Fixed fee. No hourly billing.

Twenty-five minutes. Tell us where the firm is spending its time, and we’ll tell you honestly what AI can and can’t do about it, including when the answer is “not much, yet.”