My learning Protected course

Lesson 5 · Check an AI vendor

Check an AI vendor

Length
22 minutes across 7 sections
You will be able to apply
The Contract Floor · The Exit Rehearsal
You will produce
Commit Statement
You will work
5 gated questions

Personalize the practice

Apply this to your environment

These details adapt the application prompts and coach questions. They do not affect your score.

Reading

What does the contract say happens to your data when you stop paying?

Core question

risk tiering

Tier the vendor signals before you trust the badge

Required practice

Ruth Ainsworth and Jo Halvorsen are reviewing several vendor positions. Before reading the checklist, assign each signal the risk tier it deserves.

Complete each part before committing.

A vendor evaluation is not a technology assessment. It is a contract assessment that happens to involve technology. The tool might be brilliant. The pricing might be fair. The onboarding might take an afternoon. None of that matters if the contract gives the vendor rights over your clients' data that you never agreed to grant, because you did not read the clause that granted them.

Small businesses face a specific version of this problem. Enterprise buyers have procurement teams who read terms line by line. A twenty-two-person bookkeeping practice does not. The founder clicks Accept, the team starts using the tool, and the contract terms become the organisation's data governance posture whether anyone read them or not.

This lesson gives you two instruments. The first establishes the minimum contract position you must confirm before any data enters a vendor's system. The second rehearses your departure before you arrive, so that the cost of leaving never becomes the reason you stay.

Next: The pathology: Badge Trust

The pathology: Badge Trust

A security badge on a vendor's pricing page is not evidence of anything except that the vendor paid for an audit at one point in time against one scope. SOC 2, ISO 27001, GDPR-ready: each of these is a real standard, and each of them tells you nothing about the specific contractual terms that govern your data once it enters the vendor's system.

PrincipleBadge Trust is the reflex of treating a vendor's certification logo as proof that the contract protects your data. It skips the only question that matters: what does the contract actually say about your data, your clients' data, and what happens when you leave? The badge answers a different question, about the vendor's internal controls, and substitutes that answer for the one you needed. The tell is the sentence they are SOC 2 certified so we should be fine, spoken by someone who has not opened the terms of service.

The correction is not to ignore badges. It is to treat them as a precondition, not a conclusion. A vendor without credible security controls is disqualified before you reach the contract. A vendor with credible security controls still requires you to read the contract.

Next: The Contract Floor

The Contract Floor

Definition

The Contract Floor is the minimum set of contractual confirmations you require before any organisational data enters a vendor's AI system. It is not a checklist of features. It is a checklist of rights: what the vendor may not do with your data, what happens to your data when you leave, and whether you can disable any training or model-improvement use of the content you upload. If the contract does not confirm all three, the floor is not met and the tool does not proceed regardless of its capability.

When to Use It

Apply the Contract Floor before any trial that involves real data, not after. A free trial with real client documents is a production deployment with zero contractual protection. Apply it again when a vendor changes its terms of service, which happens silently and frequently. Apply it when a tool you approved for one use case is proposed for a second that involves more sensitive data.

How to Apply It

  1. Locate the current terms of service and data processing agreement, not a marketing summary.
  2. Confirm in writing that uploaded data is not used for model training or improvement unless you opt in.
  3. Confirm that you can export or delete all data upon termination, with a stated timeframe.
  4. Confirm the jurisdiction where data is stored and processed.
  5. Confirm that the vendor's sub-processors are listed and that changes require notice.
Contract elementAcceptableUnacceptable
Training on your dataDisabled by default or by written confirmationEnabled unless you find and toggle an opt-out buried in account settings
Data on terminationExported within 30 days, then deleted with written confirmationRetained indefinitely for service improvement
Sub-processor changesNotice period with right to terminateUpdated at vendor discretion without notification
JurisdictionNamed and contractually fixedWherever the vendor decides to process at the time
Interactive modelContract Floor verification sequenceflow · 4 elements
01
Training opt-out

Uploaded data is not used for model training unless you opt in.

Each check must pass before data enters the vendor system.
Training opt-out
Uploaded data is not used for model training unless you opt in.
Data on termination
You can export or delete all data within a stated timeframe.
Sub-processor transparency
Changes to sub-processors require notice and allow termination.
Jurisdiction fixed
Data storage and processing location is named and contractually committed.

Worked example 1 of 3

Ruth Ainsworth at OmniCorp Ledger evaluated an AI tool that promised to automate bank-reconciliation matching. The pricing was competitive, the interface was clean, and the vendor's website displayed SOC 2 Type II and ISO 27001 badges prominently. Felix Nnamdi raised a concern about client documents.

Felix Nnamdi
Their marketing says enterprise-grade security. But I read the terms last night. Section 4.2 says they may use uploaded content to improve their models unless you contact support to opt out.
Ruth Ainsworth
Contact support. Not a setting. Not a contract clause.
Felix Nnamdi
Correct. And there is no written confirmation that opting out is permanent. The clause says they will use reasonable efforts to honour the request.
Ruth Ainsworth
That is not a floor. That is a preference they can revisit. We are not uploading client bank statements to a system that might train on them next quarter.

OmniCorp Ledger rejected the tool. They selected a competitor whose data processing agreement contractually prohibited training on uploaded content by default, with no opt-out reversal clause. The competitor cost twelve percent more per seat. The cost of a data breach notification to two hundred and forty bookkeeping clients would have been measured in lost engagements, not dollars.

Why This Works

The Contract Floor removes ambiguity by converting a feeling of safety into a verifiable position. A badge cannot be argued with because it is not a claim about your data. A contract clause can be read, confirmed, and held against the vendor if breached. The floor also creates a decision that is fast: five confirmations, each binary. A tool that meets the floor can proceed to functional evaluation. A tool that does not meet the floor is rejected in an afternoon rather than after three months of a trial you must now unwind.

Worked example 2 of 3Optional depth

Jo Halvorsen at OmniCorp Studio evaluated a design-generation tool for mood boards. The tool's terms stated that all uploaded images became part of a shared training dataset unless the enterprise plan was purchased. The enterprise plan required a minimum of fifty seats. An eleven-person studio could not meet the threshold, so the contractual floor could not be met at any price tier available to them. Jo rejected the tool and documented the reason: the contract made data protection contingent on a volume commitment the organisation could not satisfy.

Worked example 3 of 3Optional depth

Gina Marchetti at OmniCorp Trades assessed a scheduling-optimisation tool for field crews. The data processing agreement confirmed no training on uploaded data and provided a 14-day export window on termination. Sub-processor changes required 30 days notice. The floor was met. Gina moved to functional evaluation and ultimately approved the tool for a six-month pilot. The contract floor did not slow approval. It accelerated it, because the five binary questions resolved in a single reading session rather than weeks of back-and-forth uncertainty.

Edge Cases and NuancesOptional depth

Some vendors offer a toggle in account settings labelled do not train on my data. A toggle is not a contract clause. If the vendor's terms of service say they may train on uploaded content and the toggle is a product feature rather than a contractual commitment, the vendor can remove the toggle in a future release without breaching any agreement. The floor requires a contractual commitment, not a feature. Additionally, free tiers almost never meet the floor because the vendor's business model for the free tier is the data itself. A tool that meets the floor on its paid plan and fails it on its free plan has not met the floor until you are on the paid plan.

Knowledge check

A vendor displays an ISO 27001 badge and offers a do not train on my data toggle in account settings. Has the Contract Floor been met?

Answer first, then check.
Next: The Exit Rehearsal

The Exit Rehearsal

Definition

The Exit Rehearsal is the practice of testing your departure from a vendor before you depend on them. You identify what data you would need to extract, how long the extraction would take, what format you would receive it in, and what business process would break during the transition. You do this at the point of lowest dependency, before you have uploaded years of work, so that the cost of leaving is a known quantity rather than an escalating surprise.

When to Use It

Rehearse the exit before signing an annual contract, before migrating historical data into a new system, and every time the vendor changes pricing or terms. The rehearsal is cheapest when dependency is lowest. A business that rehearses its exit on day one of a trial has maximum leverage. A business that rehearses its exit two weeks before renewal has already lost it.

How to Apply It

  1. Export a sample of your data today and confirm the format is usable without the vendor's system.
  2. Identify which business processes depend on this vendor and what the manual fallback is.
  3. Calculate how many hours the transition would take at your current data volume.
  4. Confirm the contract states a data-export period and format upon termination.
  5. Document the total switching cost: export effort, re-entry effort, downtime, and retraining.

Worked example 1 of 3

Interactive modelExit Rehearsal review cyclecycle · 4 elements
01
Export test

Export a sample of current data and confirm the format is usable independently.

The rehearsal repeats at intervals so switching cost never becomes invisible.
Export test
Export a sample of current data and confirm the format is usable independently.
Cost assessment
Calculate hours for full migration at current data volume.
Dependency map
Identify every process that breaks if the vendor disappears overnight.
Schedule next
Set the date for the next rehearsal before closing the current one.

Felix Nnamdi at OmniCorp Ledger ran an exit rehearsal on a document-management system the practice had used for eighteen months. He exported one client's file set and found that the export format was a proprietary JSON structure that no other tool could ingest without custom parsing. Migrating 240 clients would have taken an estimated 180 hours of manual re-entry. Ruth Ainsworth used that number in the next renewal negotiation and secured a contractual commitment to CSV export within 14 days of termination. The rehearsal cost four hours. The leverage it created was worth the entire annual subscription.

Why This Works

Dependency that you cannot quantify is dependency you cannot negotiate. The Exit Rehearsal converts an abstract worry about lock-in into a concrete number: hours, dollars, or days of disruption. That number is the real price of the vendor relationship, and it belongs in the decision alongside the subscription cost. A vendor whose switching cost exceeds a year of subscription fees has priced you into staying. Knowing that number before you sign is the difference between a choice and a trap.

Worked example 2 of 3Optional depth

Hana Adeyemi at OmniCorp Provisions rehearsed exit from a demand-forecasting tool after six months of use. The export produced clean CSV files with timestamps and product codes. Re-importing into a spreadsheet took twenty minutes for the full dataset. Total switching cost: one afternoon of process documentation and staff retraining. Hana noted the low switching cost in her renewal file and used it as leverage to negotiate a month-to-month billing option instead of an annual commitment. The vendor agreed because they knew the lock-in was low.

Worked example 3 of 3Optional depth

Tomas Petrik at OmniCorp Clinic rehearsed exit from an appointment-reminder system that had been in place for two years. The export included patient names, contact details, and appointment histories in a standard format. However, the vendor's terms stated that data export was only available during the contract period and that data would be deleted 48 hours after termination with no extension. Tomas documented this as a risk: any contract dispute that led to sudden termination would result in data loss. He negotiated a 30-day post-termination retention clause before the next renewal.

Edge Cases and NuancesOptional depth

An exit rehearsal is not a threat to leave. It is an operational test, like a fire drill. You do not need to tell the vendor you are running it. Some vendors make export deliberately difficult, hiding the function or throttling download speeds during peak periods. That behaviour is itself a finding: a vendor that makes leaving hard is a vendor that knows its retention depends on friction rather than value. Also note that switching cost is not only technical. If your team has built muscle memory around a specific interface, the retraining time is part of the cost even if the data migrates cleanly.

Knowledge check

A team has used a vendor for eighteen months without running an Exit Rehearsal. The vendor now announces a sixty percent price increase. What is the team's position?

Answer first, then check.
Interactive modelBadge Trust versus Contract Floorcontrast · 2 elements
01
Badge Trust

Security certifications and marketing claims accepted as evidence of data protection.

The difference is whether assurance comes from a credential or from binding terms.
Badge Trust
Security certifications and marketing claims accepted as evidence of data protection.
Contract Floor
Specific contractual clauses confirmed in writing before data enters the system.

Decision point

Sam Oduya at OmniCorp Studio is evaluating a new AI copywriting tool. The vendor's website shows ISO 27001 certification and a GDPR-ready badge. The free trial allows uploading real project briefs. Sam wants to test it with three active client briefs to see if the output quality justifies the subscription. What should Sam do before uploading any client material?

Confidence before seeing the analysis
Commit, calibrate, and name contrary evidence first.

Decision point

Ruth Ainsworth at OmniCorp Ledger has used an AI categorisation tool for fourteen months. The vendor just sent a notice that terms of service are changing next month. The new terms add a clause allowing aggregated and anonymised usage data to be used for model improvement. Ruth has 240 clients' financial data in the system. The existing contract floor was met at sign-up. What action does Ruth take?

Confidence before seeing the analysis
Commit, calibrate, and name contrary evidence first.
Next: Common Failure Modes

Common Failure Modes

Failure modeThe badge conclusion. What it looks like in the moment: a team member points at the SOC 2 logo on the vendor's website and says we are covered, and no one opens the terms of service. The conversation moves to features and pricing. The cost when this happens: the organisation's data governance posture is whatever the vendor's default terms say it is, and those defaults favour the vendor. A breach notification goes to clients who assumed their data was protected by something more than a logo on a website. The correction: treat badges as a filter that removes vendors without basic controls, then read the contract for every vendor that passes the filter.
Failure modeThe toggle fallacy. What it looks like in the moment: someone finds a setting labelled do not use my data for training, enables it, and reports that the issue is resolved. No one checks whether the toggle is a contractual commitment or a product feature that can be removed in the next update. The cost when this happens: the vendor ships a new version, the default resets, and three months of data has been training a model before anyone notices. The correction: confirm that the commitment not to train is in the data processing agreement, not in the product interface.
Failure modeThe sunk-cost trap. What it looks like in the moment: a team discovers that the vendor's contract terms have changed unfavourably, but no one raises it because we have been using this for a year and moving would be too disruptive. The discomfort of confronting the switching cost prevents any action. The cost when this happens: the organisation accepts terms it would never have agreed to at sign-up, and the switching cost grows every month the data stays in the system. The correction: run the Exit Rehearsal quarterly so the switching cost is a known number, not an imagined catastrophe.
Failure modeThe free-tier drift. What it looks like in the moment: a team starts using a free tier for internal experiments, then gradually uploads client data because the tool is so useful and we will upgrade soon. The free tier's terms permit training on all uploaded content, but nobody checked because it was just a trial. The cost when this happens: client data has been training the vendor's model under terms the organisation never consciously accepted, and there is no mechanism to un-train it. The correction: apply the Contract Floor to free tiers before any real data enters, and treat we will upgrade soon as an admission that the current tier does not meet the floor.
Failure modeThe outdated diligence. What it looks like in the moment: someone asks whether the tool has been vetted and hears yes, we checked it last year. No one verifies whether the terms have changed since that check. The vendor updated its terms silently three months ago. The cost when this happens: the organisation operates under a contract floor assessment that no longer describes the actual contract, and the gap widens with every silent update. The correction: re-apply the Contract Floor every time the vendor notifies a terms change, and set a calendar reminder to re-check quarterly regardless of notification.
Next: The floor end to end

The floor end to end

Ruth Ainsworth at OmniCorp Ledger needed to evaluate a new AI tool that promised automated expense categorisation for the practice's 240 bookkeeping clients. The vendor offered a compelling product demonstration and displayed three certification badges on its landing page.

Ruth applied the Contract Floor. She located the terms of service and data processing agreement. The DPA confirmed that uploaded data would not be used for model training. Data export was available in CSV format within 14 days of termination. Sub-processor changes required 60 days notice. The jurisdiction was contractually fixed. Five confirmations met. The floor was satisfied.

Ruth then ran the Exit Rehearsal. She uploaded a sample of 50 test transactions, not client data, and immediately exported them. The CSV was clean, with all fields intact and timestamps preserved. She estimated that at current data volume, a full migration would require approximately 12 hours of staff time across the practice. She documented the switching cost: 12 hours of export and re-entry, plus one afternoon of process retraining for the team.

The known switching cost gave Ruth two things: confidence that the dependency was manageable, and a negotiating position. She signed a twelve-month contract with a contractual right to export at any time during the term, not only at termination. Six months later, when the vendor proposed a 40 percent price increase at renewal, Ruth quoted the switching cost, demonstrated that she could leave in under two days, and negotiated a 15 percent increase instead. The rehearsal paid for itself before the first anniversary.

Mark the level that describes you today. Nothing is submitted.

BehaviourReadyDevelopingNot yet
Applying the Contract Floor
Running the Exit Rehearsal
Responding to changed terms
Next: Commit

Commit

Commit Statement

Complete every line in your own words, then sign and date it. Attach the contract floor checklist results for one vendor you currently use.

WindowField application
Days 1 to 7Apply the Contract Floor to one AI vendor you already use. Locate the actual DPA and confirm or deny each of the five elements.
Days 8 to 21Run the Exit Rehearsal on that same vendor. Export a sample, calculate switching cost in hours, and document it.
Days 22 to 30Set a quarterly calendar reminder to re-check terms for every approved vendor. Re-apply the floor to any vendor that has sent a terms-change notice since your last check.

Checking a vendor before you sign protects the data you already hold. The next capability this program develops is giving employees clear, practical guidelines for which data they may use with approved tools and which data they may not, so that a sound contract is not undermined by unsound daily practice.

DisclaimerGeneral guidance only. All organisations and people named in this lesson are fictional. Regulated organisations should confirm requirements with a qualified professional before relying on this material.
Required practice must be complete.

Learner feedback

Did this change what you can do?