The short answer.

Use the AI you already pay for first. Before you buy another AI tool, name the piece of work it is for, check whether ChatGPT, Claude, Copilot or Gemini prepares that work well when it is given written instructions, and buy only when your plan shows a gap the assistant cannot close. A purchase with no named work behind it becomes another unused seat.

A specialist product is the right call in two cases: the work depends on data that only that product holds, or the product does something an assistant cannot do. Everywhere else, the assistant you already own is the better buy.

Name the work before you name the tool.

The first step in any AI purchase decision is to write down the one piece of work the tool is meant to prepare. A piece of work has a name, a team that does it, the records it starts from and a person who checks the result. "Drafting the weekly client status update from the project tracker" is a piece of work. "Better client communication" is not.

Naming the work does three things for the decision.

  • It gives the purchase an owner.The person who owns the work owns the question of whether the tool helps. Without that person, nobody checks whether the seats are used.
  • It gives the demo a test.You can ask the vendor to run your work, not their showcase.
  • It puts the purchase on the plan.A company that plans its use of AI lists each team's repeating work and chooses what to build first. A tool that maps to a line on that list has a place to land. A tool that maps to nothing waits for someone to find a use for it.

Most companies arrive at this question after years of experimenting, with seats bought in one team, a trial in another and a demo on the calendar. Experimenting was the right way to start. Buying on purpose is the next step, and it starts from the work.

Test the assistant you already pay for.

Before you buy, run the named piece of work through the assistant your company already pays for, with written instructions, on real cases. One person asking the assistant a loose question is not a test. A workflow is the test: your best way of doing that work, written as inputs, steps, checks, the output and the point where a person decides.

  1. Write the instructions.Ask the person who does the work best what they start from, the order they work in, what they check and what good looks like. A free workflow template from the Workflows Directory gives you a starting point.
  2. Run it on last month's cases.Pick cases where you already know the right answer, so a miss is easy to see.
  3. Record every miss.Note what the assistant got wrong and why: a missing instruction, missing data, or something the assistant cannot do.

The record of misses answers the buying question. Misses from missing instructions are fixed by better instructions. Misses from missing data are often fixed by a workflow that pulls from the business systems you already run, with a person approving what goes back. Only the third kind, something the assistant cannot do, points to a purchase. How do we get more value from the AI we already pay for? covers this step in more depth.

When a specialist product is the right call.

A specialist product is the right call when the work depends on data that only that product holds, or on something the product does that an assistant cannot. Be fair to the vendor on both counts.

  • The data lives inside the product.Some work needs a licensed dataset, a large body of records the vendor maintains, or the history inside a system of record. If the product is where the data is, the product is where the work happens.
  • The product does something an assistant does not.Some work needs real-time processing, a specialized model trained for one narrow task, or a regulated process the product is built around. An assistant with good instructions does not replace those.
  • The product is the system your team already lives in.When AI arrives as a feature inside the accounting system or the CRM your team uses all day, turning it on for named work is a small decision. Name the work and test it.

Everything else, the drafting, summarizing, checking, comparing and preparing that fills most teams' weeks, the assistant you already pay for does well with written instructions. Your current AI platform, specialist software or a built workflow compares all three routes on one piece of work.

Five questions to take into a vendor demo.

Copy these five questions into the meeting invitation and send them to the vendor before the demo.

  1. Which piece of our work does this prepare?Name it in one sentence, with the team that does it and the person who checks it.
  2. What does this do that our ChatGPT, Claude, Copilot or Gemini, with written instructions, does not?The answer should be data the product holds or something the product does, shown on screen.
  3. Can you run it on our real cases during the demo?Bring three past cases where you know the right answer.
  4. Where does a person decide?Ask what the product sends, pays or changes on its own, and where a person approves first.
  5. Who on our side owns it, and how many people will use it every week?Name the owner and count the users before you count the seats.

One decision, worked through.

Illustrative example.

Take a 55-person property management company with a demo booked for a tenant-communication product. The operations lead writes down the work the product is for: drafting replies to tenant maintenance requests and a weekly summary of open requests for each building manager. The company already pays for Claude. The operations lead writes the most experienced building manager's method as a workflow and runs it on last month's requests. The replies come out well. The weekly summary misses requests logged by phone, because those live in the maintenance system and nobody gave the assistant that export. The fix is a workflow step that reads the export, so the summary goes on the plan as the next workflow to build. The operations lead sends the vendor the five questions. The vendor's answer to the second one is a portal tenants log in to, which the company does not need. The company keeps the demo, declines the purchase and puts both workflows on its plan with the building managers as owners.

What is different in a company of under 200 people.

Most buying advice for AI is written for large enterprises, with procurement teams, a platform decision and a pilot in one division. A company of under 200 people decides differently.

  • One person can see every purchase.The owner or the head of operations can list every AI seat and subscription the company pays for. Make that list before the next demo.
  • The person who buys knows the work.In a company your size, the buyer often runs the team the tool is for.
  • Each unused seat shows.A large company absorbs an idle subscription. A company of 40 people feels it. Buying on purpose matters more at your size.
  • You have one assistant to get right.Most companies your size have approved one of ChatGPT, Claude, Copilot or Gemini. Pointing that one assistant at named work is where 10X the ROI of your existing AI comes from.

When AI spending goes into next year's numbers, the same rule holds: each line names its work. How to put AI in next year's plan and budget shows how.

Where to go next.

A purchase decision is easy when the plan already exists. How organizations plan their use of AI sets out the three steps: plan it, develop the workflows, deploy them. FullyPowered Planner is where the plan lives: you set out your teams, their roles and the work between them, and choose what to build first. With the plan in front of you, every demo has one question to answer: which line on the plan does this close?