Lesson 10 of 38 · AI 101 - 12 min
AI 101: Choose a Starter Tool Stack
Assemble a deliberately small AI tool stack chosen by job-to-be-done, not by hype, so a business team learns fast, governs cleanly, and stops paying for three subscriptions that all do the same job.
Renamed from the best AI tools guide
Most wasted AI spend is not from buying the wrong tool - it is from buying three tools that quietly do the same job, none of them well, because nobody decided what each one was for. The instinct after a good demo is to subscribe; the discipline is to first name the job, then ask which tool you already pay for could do it. A starter stack is a governance object, not a collection: it should be small enough that a new team member learns it in a week, clear enough that a manager who never opens the tool can sign off what it is allowed to do, and anchored to where your data and your team already live. This lesson turns 'which AI tool is best?' into the sharper question a professional should ask - 'which one job earns this subscription, and what is it allowed to touch?'
Choose a lean starter AI tool stack
A branded walkthrough: pick tools by the job to be done, favour ecosystem gravity and total operating cost over hype, and start on flexible monthly plans with clear approval boundaries.
What to understand
- Choose tools by job-to-be-done, not by category. The real jobs are narrow: draft client follow-ups, summarise a contract, build a weekly numbers pack, research a market, generate launch images, run a multi-step agent task. 'Get an AI tool' is not a job and cannot be governed.
- Most jobs collapse onto one general assistant plus where your data lives. ChatGPT, Claude, Microsoft 365 Copilot, and Gemini in Google Workspace each cover writing, research, document work, and analysis competently. The biggest decision is usually which ecosystem your files, email, and identity already sit in - that is where the integration and adoption advantage is, not in a niche standalone app.
- One primary and one controlled backup per job is enough to start. The primary is the daily habit; the backup is a deliberate fallback for outages, second opinions, or sensitive comparisons - not a second daily login. Beginners who run four assistants in parallel learn none of them deeply and govern none of them at all.
- A tool can be genuinely capable and still be the wrong stack choice. If it cannot meet your permissions model, export your work, give an admin control, or get real team adoption, its raw capability is irrelevant. 'Best model on the benchmark' loses to 'the one already wired into our email and identity' in almost every business workflow.
- Compare total operating cost, not the sticker price. The subscription line is the smallest number. Add training time, review and correction time, the base licence an AI add-on requires (Microsoft 365 Copilot and Gemini sit on top of paid Workspace/365 plans), and the cost of unused seats. A $20 tool the team ignores is more expensive than a $30 one they actually use.
- Annual vs monthly billing materially changes the cost - and the lock-in. Most providers discount the annual commitment (ChatGPT Business and Claude Team both land near $20/seat annual but cost more month-to-month). Pilot on monthly billing, prove adoption, then move to annual; never sign annual for a tool you have not yet seen the team use.
- The approval boundary is part of the stack. For each job, write what the tool may read, what it may draft, what it must never touch, and what a human must approve before anything leaves the team. A boundary a non-user manager can understand is the difference between a governed pilot and shadow AI nobody signed off.
- Default to the ecosystem and the tier you can outgrow, not the biggest one. Start on an individual or small-team plan, learn the real usage shape for a month, and upgrade into headroom you have proven you need. Over-buying seats and top tiers on day one is the most common silent waste in a starter stack.
Deeper dive
The collapse test: how many tools do you actually need?
Before scoring anything, run the collapse test on your job list. Write every AI job you can think of, then strike through each one a general assistant you already pay for can do acceptably. For most business teams, writing, research, summarising, drafting emails, basic analysis, and brainstorming all collapse onto a single platform - ChatGPT, Claude, Microsoft 365 Copilot, or Gemini. What survives the collapse is the short list that genuinely needs a dedicated tool: heavy creative media, a specialist data product, a coding/agent workload, or a regulated workflow with hard compliance needs. The discipline is to buy for what survives, not for every job on the original list. A three-tool stack that survives the collapse test beats a seven-tool stack assembled one demo at a time, because every tool you add multiplies onboarding, billing, and the surface a manager has to govern.
Ecosystem gravity beats benchmark scores
The single most reliable predictor of whether a business AI tool gets adopted is not its model quality - it is whether it lives where the team already works. Microsoft 365 Copilot and Gemini win not because they out-reason ChatGPT or Claude, but because they sit inside Outlook, Excel, Gmail, and Docs where the work already happens and the identity, permissions, and files are already managed. A standalone tool that scores higher on a benchmark but forces copy-paste between apps, separate logins, and a new permissions model usually loses in practice. This is 'ecosystem gravity': the integration surface, the existing data, and the admin controls pull the decision harder than raw capability. The practical rule for a starter stack is to ask 'where does our data and identity already live?' first, and treat model leaderboards as a tie-breaker, not the deciding factor.
Total operating cost: the four hidden lines
Sticker price is the smallest and most visible cost; the four that actually decide whether a tool pays off are usually invisible on the invoice. First, the base-licence line: Copilot and Gemini are add-ons that require a paid Microsoft 365 or Google Workspace seat underneath, so the true monthly figure is base plus add-on, not the add-on alone. Second, the learning line: the hours the team spends getting fluent - a powerful tool nobody can drive is pure cost. Third, the review line: the time spent checking and correcting output, which is highest exactly where the tool is least trusted or least scoped. Fourth, the dead-seat line: licences bought 'for the team' that three people actually open. A disciplined starter stack budgets all four, pilots on monthly billing to expose the learning and review lines early, and only moves to discounted annual commitment once real adoption is proven - never the other way round.
Anchor-platform comparison for a starter stack
Prices are per user and reflect the entry business/team tier each platform leads with. Copilot and Gemini are add-ons that require a paid base Microsoft 365 or Google Workspace licence - budget the base plan on top. Verify on the official pricing page before you buy; vendors change tiers and promotions frequently.
| Platform | Entry team price | Strongest first job | Lives best where | Watch-out |
|---|---|---|---|---|
| ChatGPT (OpenAI) - Business | ~US$20/seat/mo annual (≈$25-30 monthly, 2-seat min) | Prompting, project execution, Codex + agent tasks | Teams wanting the broadest model/agent toolkit independent of Office/Workspace | Codex and agent actions need an explicit approval boundary; usage is metered |
| Claude (Anthropic) - Team | ~US$20/seat/mo annual Standard (Premium ~$100 adds Claude Code) | Long documents, knowledge work, careful reasoning | Document- and policy-heavy teams; sensitive-file review with admin controls | Claude Code only on Premium seats; plan which seats need it |
| Microsoft 365 Copilot | ~US$30/user/mo annual (Business add-on ~$18 promo to 30 Jun 2026, then ~$21) | AI inside Word, Excel, Outlook, Teams you already use | Organisations already standardised on Microsoft 365 | Add-on requires a qualifying M365 base licence - true cost is base + add-on |
| Gemini in Google Workspace | Bundled in Business Standard ~US$14/user/mo (Starter ~$7 limited) | AI inside Gmail, Docs, Sheets, Meet you already use | Organisations already standardised on Google Workspace | Standalone Gemini add-on retired in 2025; full features need Standard+ |
Sources (as of June 2026): OpenAI - ChatGPT pricing · Anthropic - Claude plans & pricing · Microsoft 365 Copilot pricing · Google Workspace pricing
Anchor platform chooser
The four anchor platforms most starter stacks are built around. Match the row to your existing ecosystem before buying anything new - the full pricing detail is in the comparison table below.
Step by step
Name the jobs, then run the collapse test
List the exact AI jobs your team does in a typical week - phrased as tasks, not categories. Use statements like 'draft client follow-up emails', 'summarise policy changes', or 'build the weekly numbers pack'. Then strike through every job a general assistant you already pay for could do acceptably. Only what survives needs a dedicated tool.
HintIf the job is vague ('be more productive'), every tool will look useful and none will be accountable. Tighten it until one person could own the output. You are done when every surviving job is phrased so one person could own its output.
Score each candidate on four lenses
For each surviving job, score 1-5 on: workflow fit (does it do this specific job well), data risk (can it meet your privacy and compliance needs), integration surface (does it live where your data, email, and identity already are), and total operating cost (subscription plus any required base licence, training, and review time). A high integration score usually outweighs a small fit advantage.
HintA cheap tool that duplicates one you already pay for is still expensive. Score redundancy as a cost, not a saving. You are done when each surviving job has four scores and a defensible front-runner.
Anchor on an ecosystem, then choose primary and backup
Decide which platform anchors the stack - usually the one already holding your data and identity (Microsoft 365, Google Workspace, ChatGPT, or Claude). Pick one primary tool per job and one controlled backup for outages or second opinions. Write one sentence on why each is present. Pilot on monthly billing before any annual commitment.
HintThe backup is a deliberate fallback, not a second daily habit. If you find yourself using both daily, you have two primaries and no decision. You are done when each job has exactly one primary, one backup, and one sentence of justification each.
Write the approval boundary per job
For each tool, state what it may read, what it may draft, what it must never access, and what a human must approve before anything leaves the team. For agent or coding tools (Codex from OpenAI, Claude Code from Anthropic - the tools that can act on files rather than just chat) be explicit about actions: what it may run, edit, or send without a human gate.
HintThe boundary should be understandable to a manager who never opens the tool. If they can't sign it, it isn't a boundary - it's shadow AI. You are done when the boundary would survive being read aloud in a team meeting.
Cost the stack and set the upgrade trigger
Total the real monthly cost - every subscription plus required base licences and an honest estimate of seats that will actually be used. Start on the smallest viable tier and write the explicit trigger that would justify upgrading (e.g. 'when 80% of seats hit the usage limit two weeks running'). Mark any current tool the new stack makes redundant for cancellation.
HintBuy into headroom you have proven you need, not headroom you imagine. Over-buying seats and top tiers on day one is the most common starter-stack waste. You are done when the total includes base licences and at least one redundant tool is marked for cancellation or given a reprieve in writing.
Build a starter stack scorecard for one real workflow: name the job, run the collapse test, score the top two candidates on the four lenses, choose primary and backup, write the approval boundary, total the real monthly cost (including base licences), and set the upgrade trigger.
A starter stack scorecard: for each named job, your primary tool, controlled backup, the data it may touch, the monthly cost, and the approval boundary a non-user manager could sign off.
Production prompt examples
ROLE: You are a pragmatic AI procurement advisor for a small business team. You optimise for the smallest stack that does the real work, not the most tools. CONTEXT: - Our team size: [e.g. 6 people, mostly non-technical]. - Where our data and email already live: [e.g. Microsoft 365 / Google Workspace / neither]. - Tools we already pay for: [list, with monthly cost]. - Our top recurring AI jobs this quarter: [list 3-5, e.g. draft client emails, summarise contracts, build the weekly numbers pack, research competitors]. - Hard constraints: [e.g. client data must not train models; manager must approve anything client-facing; budget ceiling $X/month]. TASK: Recommend a starter stack of AT MOST one primary tool and one backup PER job. Prefer tools we already pay for or that live in our existing ecosystem. Do not recommend a new subscription unless a job genuinely cannot be done with what we have. FOR EACH RECOMMENDED TOOL, GIVE ME: 1. The exact job it earns its place for (one sentence). 2. Primary vs backup, and why. 3. Estimated monthly cost INCLUDING any required base licence. 4. The approval boundary: what it may read, what it may draft, what it must never touch, what a human must approve. 5. One risk to watch (data, lock-in, or adoption). THEN: Flag any tools we currently pay for that are now redundant, and any job where we are about to over-buy. CONSTRAINTS: - If two tools do the same job, pick one and say why - do not hedge. - Use ranges and 'verify current pricing' for any price; do not state a price as certain. - Output as a table I can paste into a one-page scorecard, followed by a 3-line summary recommendation.
- Filling the CONTEXT block with your real ecosystem is what makes the answer specific instead of a generic top-10 list - the model recommends into your existing data and identity.
- 'At most one primary and one backup per job' forces the collapse test and stops the model from listing every tool it knows.
- Requiring cost INCLUDING the base licence surfaces the hidden Copilot/Gemini add-on trap that headline prices hide.
- Asking for the approval boundary per tool turns the output straight into a governable scorecard a manager can sign.
- 'Use ranges and verify current pricing' keeps the model honest about fast-changing prices instead of asserting stale figures.
- 'If two tools do the same job, pick one' prevents the wishy-washy 'it depends' answer that leaves you with three overlapping subscriptions.
Common mistakes to avoid
- Buying three tools that all solve the same writing or research job because each had a good demo.
- Comparing demos and benchmark scores instead of comparing the actual weekly workflow and where your data already lives.
- Forgetting that Microsoft 365 Copilot and Gemini are add-ons that require a paid base licence - budgeting the add-on price alone.
- Signing an annual commitment before the team has shown it actually uses the tool day to day.
- Forgetting export and cancellation risk before a pilot stores useful work you'd lose if you switch.
- Letting personal preference for a favourite model override team adoption, integration, and governance needs.
- Over-buying seats and the top tier on day one instead of starting small and upgrading into proven need.
Source conflicts to review
- ChatGPT Business is reported near $20/seat/month on annual billing after a 2 April 2026 drop from $25, but month-to-month runs $25-$30/seat with a two-seat minimum - confirm the billing term before comparing platforms.
- Claude Team splits into Standard (~$20/seat annual, no Claude Code) and Premium (~$100/seat annual, includes Claude Code); a single quoted 'Team price' can mean either, so check the seat type.
- Microsoft 365 Copilot Business carries a promotional ~$18/user/month rate quoted as ending 30 June 2026 and rising to ~$21, while the enterprise add-on stays ~$30 - and all of these sit on top of a base licence, so headline figures understate true cost.
- Google retired the standalone Gemini add-on in 2025 and bundled Gemini into Workspace Business Standard and above; older guides still describe a separate Gemini add-on price that no longer applies the same way.
Key terms
- Starter stack
- A small set of tools chosen for named jobs, not a random collection of subscriptions.
- Approval boundary
- The line between what a tool can read or draft and what a person must approve before it leaves the team.
- Job-to-be-done
- The real task the learner needs completed, expressed in practical business language.
- Collapse test
- Striking out every job a tool you already pay for can do, so you only buy for what genuinely survives.
- Ecosystem gravity
- The pull of existing data, identity, and integrations that makes a tool adopt better than a higher-scoring standalone.
- Total operating cost
- Subscription plus base licence, training, review time, and unused seats - the real cost of a tool, not its sticker price.
- Shadow AI
- AI use nobody approved or supervises - tools, accounts, and prompts running outside any agreed boundary.
- Anchor platform
- The single ecosystem (e.g. Microsoft 365, Google Workspace, ChatGPT, Claude) the rest of the stack is organised around.
Resources
- articleOriginal best AI tools article
- siteAI tools directory
- docOpenAI - ChatGPT pricing (verify current tiers)
- docAnthropic - Claude plans & pricing
- docMicrosoft 365 Copilot - plans and pricing
- docGoogle Workspace - pricing
- siteAI Kick Start services
- siteAI Kick Start AI tools directory
- siteAI Kick Start news and guides
Checkpoint