Gemini, Microsoft Copilot, Perplexity & the AI Tool LandscapeChoosing tools and building your stack · Lesson 18 of 19
Cost, privacy and lock-in trade-offs
Video lecture
Cost, privacy and lock-in trade-offs
The narrated lecture is in production
Every chapter is scripted and ready. Browse the chapters and read the full transcript now — the video will appear here when it’s published.
Chapters
Transcript of the narration, chapter by chapter.
0:00 Cost, privacy and lock-in
Choosing an AI tool is not just about which one writes best. It is also about what it really costs once you include checking time, what happens to your data, and how hard it would be to leave. In this lecture you will learn to calculate the true cost of a tool, the questions to ask every vendor, and simple habits that keep you portable.
0:28 Why this matters
Why does this matter? Because AI subscriptions multiply quietly, one person at a time. Because data terms vary widely between tools, and even between plans of the same tool. And because every month you spend building prompts and workflows inside one product, it gets a little harder to leave. None of this is a reason to avoid AI. It is a reason to manage it deliberately.
0:57 True cost
Start with true cost. The formula is simple. Cost per useful output equals the tool cost, plus the human time spent checking and fixing, divided by the number of usable outputs. The sticker price is only the first term. Fixing time is usually the biggest one, and it is the one people forget.
1:20 The analogy
Think of a cheap printer with expensive ink. The box looks like a bargain, but over a year you pay far more. AI tools work the same way. A cheap tool whose drafts need ten minutes of fixing each can cost much more than a pricier tool whose drafts need three.
1:42 Worked numbers (illustrative)
Let's put illustrative numbers on it. Tool A costs twenty a month, and each draft needs ten minutes of fixing. Tool B costs forty a month, and drafts need three minutes. At sixty drafts a month and a staff cost of thirty an hour, Tool A costs twenty plus three hundred in time, so three hundred and twenty. Tool B costs forty plus ninety, so one hundred and thirty. The expensive tool is less than half the true cost. Use your own numbers, but always include the time.
2:20 Controlling costs
Now control costs. Audit overlaps and cancel duplicate subscriptions. Match model tiers to tasks, cheap and fast for routine work, top models only where they earn it. Use the AI bundled in tools you already pay for before adding new ones. Set spending limits and alerts on API accounts. And assign premium seats to heavy users only.
2:45 Vendor questions
Next, privacy and compliance. Ask every vendor seven questions. Is our data used to train models, and can we prevent it contractually? How long is data retained? Where is it processed and stored, and which sub processors are involved? What security controls exist, like single sign on, multi factor authentication and audit logs? Is there a data processing agreement, and support for our regulations, such as GDPR, the UAE and Saudi data protection laws? What admin controls exist for memory, connectors, agents and sharing? And how do we export or delete our data? Record the answers in an AI tools register.
3:29 Simple example
A simple example. A colleague loves a free transcription app and wants to use it for client calls. You ask question two, how long is data retained? The answer, buried in the terms, is indefinitely on the free tier. That settles it. The app can be used for public webinars, but not client calls. Five minutes of reading prevented a real confidentiality problem.
3:56 Staying portable
Finally, lock in. Keep prompts, instructions and source files in your own storage, a shared drive or repository, and copy them into tools, not the other way round. Export important outputs regularly. Favour open standards, MCP for tool connections and OpenAI compatible APIs, which many providers and local runtimes support. In code, wrap model calls in a thin layer so switching providers is a configuration change. And document workflows so they can be rebuilt elsewhere. Some lock in is fine for great integration, as long as you choose it knowingly.
4:35 Business example: an Abu Dhabi consultancy
A realistic example. A fifteen person consultancy in Abu Dhabi lists eleven AI subscriptions, including four overlapping assistants and two transcription tools. It calculates cost per useful output for the three most used tools, consolidates to one business assistant plan plus the Microsoft three six five Copilot it already partly licenses, keeps one research tool, and cancels six subscriptions. The vendor questions reveal one transcription tool kept recordings indefinitely on its free tier, so it goes. And all prompts and Project instructions move into a shared SharePoint library, ready for any future switch. The following year, their main assistant vendor changed its pricing. Because their prompts, Project instructions and workflows lived in their own SharePoint library, they tested an alternative with their saved test set in a week, and used the result to negotiate a better renewal. Portability gave them options, and options gave them leverage.
5:38 Common mistakes
Four common mistakes. Comparing sticker prices only. Using free tiers with unknown data terms for client work. Keeping your best prompts and workflows only inside one tool. And having no owner for the AI tools register, so it goes stale within a quarter.
5:57 The AI tools register
Pull all of this together in an AI tools register, a shared sheet with one row per tool. The owner, plan and monthly cost. The data classes it is approved for. Whether it trains on your data, the retention period, the residency, whether a data processing agreement is signed, and the next review date. It sounds like admin, but it answers client security questionnaires in minutes, shows exactly where the money goes, and turns renewals into deliberate decisions.
6:31 Watch me do it, part 1
Let me run the Abu Dhabi consultancy's audit. I list all eleven AI subscriptions with monthly cost and users. For the three most used tools, I add a cost per useful output column. The formula is the tool cost plus fixing time, minutes multiplied by the hourly staff cost, divided by the number of usable outputs. With their own illustrative numbers, the cheapest assistant turns out to be the most expensive per useful draft, because its outputs need the most fixing. Four overlapping assistants are highlighted amber.
7:09 Watch me do it, part 2
Next, the seven vendor questions. For one transcription tool, the answer to how long is data retained is indefinitely on the free tier. It has been used for client calls, so it is removed and those recordings are deleted where possible. I fill the tools register, with approved data classes, retention, residency and the data processing agreement for each tool I keep. Six subscriptions are marked cancel. Finally, portability. We copy every useful prompt and Project instruction into a shared library in SharePoint, so they live in our storage, not only inside a vendor's product.
7:50 Recap and try this now
Recap. Compare tools on cost per useful output, including fixing time. Ask every vendor the seven questions and keep a register. Stay portable with your own prompt library, exports, open standards and a thin wrapper in code. Try this now. Fill in a cost and privacy table for every AI tool you use for work, then identify one subscription to cancel or consolidate, and one data terms gap to fix this month.
True cost: cost per useful output
Subscription or API price is only part of the cost. Compare tools on cost per useful output:
Cost per useful output = (tool cost + human time to check and fix) / number of usable outputsIllustrative example: Tool A costs 20 per month and produces drafts that need 10 minutes of fixing each; Tool B costs 40 per month but drafts need 3 minutes. At 60 drafts a month and a staff cost of 30 per hour, A costs 20 + 300 = 320; B costs 40 + 90 = 130. The "expensive" tool is cheaper. (Numbers illustrative; use your own.)
Controlling costs
- Audit overlaps: list every AI subscription and seat; cancel duplicates.
- Match tiers to tasks: fast/cheap models for routine work, top models only where needed (applies to API usage especially).
- Use bundled AI you already pay for (Workspace, Microsoft 365) before adding tools.
- Set spending limits and alerts on API accounts and usage-based plans.
- Assign seats deliberately: heavy users get premium seats; occasional users get standard.
Privacy and compliance: questions to ask every vendor
- Is our data used to train models? Can we contractually prevent it?
- How long is data retained, and can we control retention?
- Where is data processed and stored (residency)? Which sub-processors are involved?
- What security certifications and controls exist (SSO, MFA, audit logs, encryption)?
- Is there a data processing agreement (DPA) and, where relevant, support for your regulatory regime (for example UK/EU GDPR, UAE PDPL and DIFC/ADGM rules, Saudi PDPL)?
- What admin controls exist for features like memory, connectors, agents and sharing?
- How do we export or delete our data?
Keep the answers in an AI tools register alongside the data classes each tool is approved for.
Lock-in and portability
Lock-in happens when your valuable assets (prompts, instructions, knowledge files, workflows, integrations) exist only inside one vendor's product. Reduce it by:
- Keeping prompts, instructions and source files in your own storage (a shared drive or repo), then copying them into tools.
- Exporting outputs regularly into your own document system.
- Favouring open standards: MCP for tool connections (supported across major assistants), OpenAI-compatible APIs (many providers and local runtimes support them), standard file formats.
- Abstracting API calls in your own code (a thin wrapper) so switching model providers is a configuration change.
- Documenting workflows so they can be rebuilt elsewhere.
Some lock-in is acceptable in exchange for integration benefits (for example, embedded copilots in your suite). Choose it knowingly.
Worked example: a consultancy's AI audit
A 15-person consultancy in Abu Dhabi lists 11 AI subscriptions across the team: four overlapping assistants, two transcription tools and several single-purpose apps. It calculates cost per useful output for the three most-used tools, consolidates to one business assistant plan plus the Microsoft 365 Copilot it already partly licenses, keeps one research tool, and cancels six subscriptions. The vendor questionnaire reveals that one transcription tool retained recordings indefinitely on its free tier; it is removed. Prompts and project instructions are moved into a shared SharePoint library so they survive any future switch.
Hands-on
Complete a cost and privacy table for every AI tool you pay for or use for work: cost, usage, cost per useful output (estimate), training/retention terms, approved data classes, owner. Identify one subscription to cancel or consolidate and one data-terms gap to fix.
A simple AI tools register
| Tool | Owner | Plan | Monthly cost | Approved data classes | Training on our data? | Retention | Residency | DPA signed | Review date |
|---|---|---|---|---|---|---|---|---|---|
| Business assistant | Ops lead | Business | [x] | Public, internal, confidential | No (business terms) | Admin-set | [region] | Yes | [date] |
| Research tool | Research lead | Pro | [x] | Public, internal | Check terms | [x] | [x] | [x] | [date] |
Keep it in a shared sheet. It answers client security questionnaires in minutes, shows where money goes, and makes renewals a deliberate decision rather than an automatic one.
Negotiating with vendors
For team and enterprise purchases, ask for: annual pricing with seat flexibility, a signed DPA, contractual no-training commitments, admin and audit features, data export on exit, and advance notice of material changes to terms or models. Smaller buyers can still ask; the answers tell you how the vendor treats customers.
Pitfalls
- Comparing sticker prices only.
- Free tiers with unknown data terms used for client work.
- Prompts and workflows that exist only inside one tool.
- No owner for the AI tools register.
How to measure success
Your AI spend is known and justified per tool, every tool has documented data terms and approved data classes, and your key prompts and workflows could move to another vendor within a week.
Key takeaways
- Compare cost per useful output, including human time to check and fix.
- Control costs by auditing overlaps, matching model tiers to tasks and setting spending limits.
- Ask vendors about training use, retention, residency, security, agreements and admin controls.
- Reduce lock-in by keeping prompts and files in your own storage and favouring open standards.
Check your understanding
Quick questions to lock in the lesson. They don’t count towards your certificate.
Put it into practice
Complete the cost and privacy table for every AI tool you pay for or use for work. Identify one subscription to cancel or consolidate and one data-terms gap to fix.
Enrol for free to save your progress
Reading is always free. Enrol to keep your place, take the final assessment and earn a verifiable certificate.