The short answer.

You have six routes for putting AI to work: do it yourselves, train your team, hire a developer, hire a consultant, switch on an AI feature in software you already run, or have a workflow built with a services company such as FullyPowered AI. Every route costs two things, a fee and the time of the people who know the work, and the fee is the smaller and more visible of the two. Once a workflow is running, the cost is the same whichever route built it: the ChatGPT, Claude, Copilot or Gemini subscription you already pay for, plus your own people’s time to run and check the work. Choose by the piece of work in front of you: do it yourselves when the procedure is written and someone has the time, have it built when the work is particular to your company and nobody inside has the time, and buy nothing new until your plan shows a gap your assistant cannot close.

The articles behind this guide each settle one of those choices. Read together, they establish something none of them says alone: every option is decided by the same few facts about one named piece of work. Which team does it, what records it starts from, who checks it, whether the method is written down, and who has the time to write it. Those facts decide whether a template is enough, whether a seat is needed, whether a vendor’s product earns its place, which sprint fits and whether an agent is ready. A company that writes them down once can answer every buying question in the same sitting, and a company that skips them pays for options it did not need.

Choosing among the options is part of the first two of three steps set out in How organizations plan their use of AI: plan it, develop the workflows, deploy them.

Every option costs a fee and your people’s time.

Every option for putting AI to work costs a fee, which can be zero, and the time of the people who know the work, which never is. Someone has to say how the work is done well, gather examples, judge the first results and keep the instructions current. Doing it yourselves moves the whole of that time onto your own people. Paying for help moves part of it onto someone else, and what you buy is mostly the expert’s time back, not the AI.

A third measure matters as much as the fee and the time: what each option leaves you with. Ask of every option what you will own, who runs it the next day and who changes it when the procedure moves.

The running cost does not depend on the route. A workflow runs in the assistant your company already pays for, so after setup you pay that subscription and the time of the person who runs the work and the person who checks it. What does it cost to put AI to work in your company? sets out our own fixed prices next to what your team does itself in each one.

Put the three measures together: the cheapest fee often spends the most expert time, and expert time is the scarcest thing a company has. The right option is the one that spends the least of that time for a result your people own.

Before you spend anything: take stock, write the plan, try a template.

Before you pay for any option, do three things that cost only a little of your own time: list the AI the company already pays for, write a one-page plan of the work each team repeats, and try a free workflow template on one piece of that work. Each step is cheaper than the one after it, and each one’s result is the brief for the next, so you never pay for a step you did not need.

The stocktake comes first because a company rarely knows what it already owns, and the column that matters names which piece of work each seat is used for. How to take stock of the AI your company already pays for has the three-column sheet and the three places to look.

The plan comes second: team by team, the work that repeats and the first piece to build. The template comes third, run on real cases where you know the right answer. If the template does the job, you may need nobody. If it falls short because the work depends on how your company does things, you now know exactly what to ask a developer, a consultant or us to build.

Before you spend: a checklist to copy. Fill it in for the company, then for the one piece of work you want to start with.

  1. What we already pay for, and the named work each one is used for.Seats, AI features inside our software, subscriptions on expenses: ______
  2. Our teams and the work each repeats.About five pieces per team: ______
  3. The one piece of work we start with.The team, the records it starts from, the person who checks it: ______
  4. Is the procedure written down?Written and agreed, in notes, or in one person’s head: ______
  5. Who knows the work best, and do they have time to capture it?Name, and yes or no: ______
  6. The template we tried, and where it fell short.______
  7. Why each miss happened.Missing instructions, missing data, or something the assistant cannot do: ______
  8. Does the work cross teams or systems?Which ones: ______
  9. The route these answers point to.______

Six routes, and when each one is right.

Each of the six routes is the right one for some piece of work, and four facts decide which: whether the procedure is written, who has the time, how particular the work is to your company, and who will keep the result current.

  • Do it yourselves.Right when the procedure is written, one person has the time and the interest to own it, and a wrong answer is cheap to catch because someone reviews the work. Costs no fee, only your people’s time, taken from their regular work. Leaves workflows written by the people who know the work, which they change themselves.
  • Train the team.Right when people need confidence across many kinds of task and you want broad individual ability. Costs a course fee and the time people spend in it. Leaves people who use the assistant more confidently, though not always a shared workflow the team runs the same way.
  • Hire a developer.Right when the job needs custom software, direct system connections or technical work beyond your team. Costs the developer’s fee plus your team’s time explaining the work and checking the result. Leaves software built to your requirements, which someone technical maintains when the procedure changes.
  • Hire a consultant.Right when the big question is which work to change, how responsibilities should shift or how several teams and systems fit together. Costs a fee that follows the scope, plus your leaders’ time in interviews and reviews. Leaves anything from a recommendation your team still carries out to working systems with support, so agree the handover and the tests before you choose.
  • Use a feature you already have.Right when the software your team works in all day already holds the information and produces the result your people need. Costs little beyond the time to test it on representative work, the awkward case included. Leaves a result shaped by how that tool works rather than how your company does.
  • Have a workflow built with FullyPowered AI.Right when the work is particular to your company, or crosses teams and systems, and nobody inside has the time to capture and test it. Costs a fixed price known before work starts, plus your people’s time in interviews and testing. Leaves a tested workflow your own people run in the AI you already pay for, which you keep whether or not you keep working with us. We are a better way to get the work done, not the only way.

Should we hire an AI consultant, and what does it cost? sets out each route in full, with what to ask before you hire anyone.

An options worksheet to copy. One sheet per piece of work. Fill in the facts first; the route follows from them.

  1. The piece of work.One sentence, named for what it produces: ______
  2. Team, owner and the person who checks every output.______
  3. How the method exists today.Written and agreed, loose notes, or someone who can show the job: ______
  4. Common or particular to us.Does a workflow template exist for it? ______
  5. Systems it starts from and writes back to.By kind: ______
  6. Route chosen, and why.______
  7. Fee, by kind.None, course, developer, consultant, or a fixed price: ______
  8. Our people’s time.Who, and for what: ______
  9. What we own at the end.______
  10. Who keeps it current after handover, and the decision date.______

Is there software for planning, or will a spreadsheet do?

Yes, software for planning how a company uses AI exists, and a spreadsheet does the job for one team and about ten rows. A planning spreadsheet strains as soon as a second or third team joins: one piece of work belongs to several teams, the same role appears in many rows, several leads edit the file at once, and the plan drifts away from the workflows it names. Is there software for planning how a company uses AI? gives the spreadsheet’s columns and what to look for in software.

FullyPowered AI Planner is our version of planning software. You set out your teams first, then their roles and the work between them, list a workflow under every team that takes part in it, and choose what to build first. Colleagues add their own teams. The prices, as the pricing page states them:

  • Up to 5 roles.$49 /year. Seven-day trial, everything unlocked.
  • Up to 20 roles.$199 /year.
  • Up to 50 roles.$499 /year.
  • More than 50 roles.Enterprise.

Priced by roles, never people. Unlimited colleagues. Teams never limited. The trial starts with a card; nothing is charged for seven days, and you can cancel before then. Every plan is on the pricing page.

The way a planning tool is priced shapes who writes the plan. Software priced per seat makes every invitation a purchase, so the plan gets written by the few people with a licence. A price that follows the kinds of job in the company lets everyone who knows the work join. Why the Planner is priced by roles, not by people counts one company both ways.

Should we buy another AI tool, or use the assistant we pay for?

Use the assistant you already pay for first, and buy a specialist product only when a test on real cases shows the work needs data that product holds, or something an assistant cannot do. The test is a workflow, not a loose question: the best person’s method written as inputs, steps, checks and output, run on last month’s cases, with every miss recorded and its cause noted.

The causes of the misses settle the purchase. A miss from missing instructions is fixed with better instructions. A miss from missing data is often fixed by a workflow step that reads an export from a system you already run, with a person approving anything that goes back. Only the third kind, something the assistant cannot do, points to buying. Should we buy another AI tool or use what we have? has five questions to send a vendor before the demo, and Your current AI platform, specialist software or a built workflow compares all three routes on one piece of work.

The test that settles a purchase also writes the first draft of the workflow, which you keep whichever way the decision goes. Why the seats a company already owns so often go unused is the subject of Why companies get so little from the AI they pay for.

One assistant, and seats counted from the plan.

Standardize on one AI assistant as the company default, and give a role paid seats when the plan gives that role a workflow to run. One default means one set of rules, one place for shared workflows and one bill. Choose the assistant your company already has through the software it runs and that your people already open; the difference between ChatGPT, Claude, Copilot and Gemini matters far less than whether people reach for it without thinking. Should we standardize on ChatGPT, Claude, Copilot or Gemini? has a one-page decision sheet.

Workflows written as plain instructions move between assistants, so the choice is not a trap: on a switch you adapt the setup around the instructions and run every test case again.

Seats follow the same logic as the assistant. A seat bought before a role has a workflow waits for work; a seat bought with the workflow is used from the first day. Seats belong to roles, not people. Does every employee need an AI licence? has a seat list to copy, one row per role, filled from the plan rather than the payroll. The assistant, the seat and the workflow are three answers to one question: which role runs which named piece of work.

Template, built workflow or agent: what to develop first.

Start from a free workflow template where the work is common, have a workflow built where the work is particular to your company or crosses teams and systems, and leave agents until a workflow in daily use has earned less attention. The three are steps on one path, not rival purchases.

A workflow template from the Workflows Directory is a sample and a starting point; it becomes your workflow when you change five things (the inputs, the rules, the exceptions, the output format and who decides) and test it on your own past cases until its owner signs off every one without rewriting it. A template falls short in three situations: the work is your own method and nobody else does it your way, the work crosses several teams or systems, or the person who knows the work has no time to step away from it. Are AI workflow templates enough, or do we need workflows built for us? has a “make this template yours” checklist.

When you want the work built with you, our two sprints are the fixed-price versions of that route. They are alternative starting points; neither requires the other.

  • 3-Day Workflow Sprint, $750.Starting point: One written procedure, already agreed. What we build: One AI workflow from that procedure. Timing: Three days, once the procedure, examples, people and workspace are ready. Support: Five business days to correct delivery defects.
  • 2-Week Workflow Sprint, $3,500.Starting point: Loose notes, ad hoc processes or someone who can show us the job. What we build: We capture one team's procedures, write them down and build the priority workflow. Timing: About two weeks from the first interview. Support: Ten business days to correct delivery defects.

Fixed introductory prices, once per client. Further workflows and wider rollout are scoped separately. The question “is the procedure written down?” on the checklist above chooses between them: a written, agreed procedure suits the 3-Day sprint, and a method in notes or in someone’s head suits the 2-Week sprint.

An agent, meaning software that carries out steps by itself toward a goal, comes after workflows, not instead of them. A workflow a person starts and checks teaches the company the work’s inputs, its exceptions and what good looks like, and those are exactly what an agent is made from and checked against. AI agents or workflows: which should a small company build first? gives five signs that a workflow is ready to run with less attention, starting with a sustained period on real work and an exception log gone quiet.

What it costs to run, and how to tell whether it was worth it.

After setup, a workflow costs the AI subscription you already pay for and the time of the people who run and check it; whether it was worth it is measured against a baseline your own team writes before the first run. Six measures make the baseline: elapsed time, hands-on effort, review, rework, output quality and adoption, each with its source, and “unknown” written wherever nobody has the number yet. Take the same six again once people in the role use the workflow. What is an AI workflow worth? has the worksheet, blank and filled in. Freed time becomes money only when you decide what it goes to.

How a provider charges decides who counts. A share of the savings needs a baseline both sides accept, a running count of your people’s work and a reconciliation at every billing date. A fixed fee is agreed with the scope before work starts, and afterwards your people own the numbers and keep every gain. Why a fixed fee and not a share of the savings sets the two side by side. Our pricing page puts it in one line: “Every price is fixed and known before you start.” Larger work, across teams or to launch a new function, is listed as “Fixed fee after scoping”, and Keep it working, our offer for after the support window, as “Fee agreed once in use”.

The baseline is also the start of the rollout record. How to roll out AI across a company, and tell whether it is working carries the same record from the first team to the whole company.

One company choosing its options, piece of work by piece of work.

Illustrative example.

Take an 80-person civil engineering firm with five teams: project delivery, bids and proposals, drafting, finance and the office. The operations manager owns AI and starts with the stocktake: company seats on one assistant, bought for the bids team, personal subscriptions claimed on expenses by two engineers, and an AI feature inside the project accounting system, switched off.

The firm’s email and documents suite already includes the assistant the bids team uses, and the engineers open it too, so it becomes the default. The operations manager moves the plan from a spreadsheet into the Planner once project delivery and finance join, because the handover from a won bid to the project team belongs to both bids and delivery.

Each piece of work then gets its own route. The weekly project status report is common work: the project managers start from a workflow template, put in the firm’s own headings and thresholds, and test it on last month’s reports, so they do it themselves. The method statement for a public tender is the firm’s own method, held by the senior bid writer and nowhere on paper, and the senior bid writer has no time to step away from live bids, so the firm books a 2-Week Workflow Sprint to capture it. A tender-search product holds a database of public tenders the assistant cannot reach, so the firm buys it for the bids team only. An agent that would submit routine documents by itself is put on the plan as “later”, to be checked against the five signs once the status report has run for a sustained period.

What is different in a company of under 200 people.

A company of under 200 people chooses among its options in one sitting, because the person who buys, the person who owns the work and the person who does it are often in the same room. Much published advice assumes procurement cycles and evaluation committees.

  • Expert time is the binding cost.The person who knows the work best usually does it every day. The question is less whether your team can capture a workflow than whether that person can step away to do it.
  • Every purchase is visible.One person reconciles the company card and can list every AI seat and subscription.
  • One assistant, one meeting.The owner of AI, the administrator and the team leads settle the default and the seat list together.
  • A fixed price fits the budget.A price known before work starts goes straight into this year’s numbers without a procurement round.
  • Speed is the edge.A small company can have a workflow in daily use while a larger competitor is still choosing a platform, and every approval step borrowed from a large company gives that edge back. How can a small company compete with larger ones using AI? shows where to aim the first workflows.

That is where 10X the ROI of your existing AI comes from in a smaller company: the assistant you already pay for, pointed at named work, with money spent only where your own facts show a gap.

Questions people ask.

For short answers across the whole subject of planning AI, Answers takes the questions leaders ask in order.

Do we need an AI consultant to get started?

No. A company can start by listing the AI it already pays for, writing a one-page plan of the work each team repeats and trying a free workflow template on one piece of that work. Bring in help when the work is particular to your company and nobody inside has the time to capture it, build the workflow and test it on real cases.

Is it cheaper to do AI ourselves?

In fees, yes; in total, it depends on whose time you spend. Doing it yourselves costs no fee beyond the AI subscription you already have, but someone has to write the procedure down, gather examples, test the result and keep it current. It stalls when the method lives in one busy person’s head.

Is there software for planning how a company uses AI?

Yes. A spreadsheet with one row per piece of work does the job for one team, and planning software earns its place when a second or third team joins, because the work that matters belongs to several teams. FullyPowered AI Planner holds your teams, their roles and the work between them in one plan, and you choose what to build first.

Should we buy a specialist AI product or use ChatGPT, Claude, Copilot or Gemini?

Test the assistant you already pay for first, with written instructions, on real cases where you know the right answer. Buy a specialist product only when the misses show the work needs data that product holds or something an assistant cannot do.

Should the whole company use the same AI assistant?

Yes. One default assistant on a company account gives one set of rules, one place for shared workflows and one bill. Choose the one the company already has and people already open.

Does every employee need a paid AI seat?

No. A role gets paid seats when the company’s plan gives that role a workflow to run, and then every person in that role gets one. Count seats from the plan, role by role, never from the headcount.

Are free AI workflow templates enough?

A workflow template is enough to start and to learn what a piece of work needs. It becomes your workflow once you put in your own inputs, rules, exceptions, output format and the person who decides, and test it on your own past cases. Have a workflow built when the work is your company’s own method, crosses several teams or systems, or nobody inside has time to capture it.

Should we build AI agents or workflows first?

Build workflows first. A workflow that a person starts and checks teaches the company the work’s inputs, exceptions and standard, which are exactly what an agent needs before it can act alone. A workflow earns less attention once it has run on real work for a sustained period, passes a written test set and its exception log has gone quiet.