←
AI-Ready Foundation
Aware · M5 · lesson 5 of 5 · queued
Preview — browse every lesson free. Enroll to mark lessons complete, open partner links and save your progress. Login & enroll →
Your First AI Acceptable Use Policy and an Honest AI Readiness Assessment
📖
now learning

Your First AI Acceptable Use Policy and an Honest AI Readiness Assessment

15 min

"Are we ready for AI?" When a board member asks that, most CIOs reach for a vendor score and a policy nobody follows. An AI acceptable use policy is a short document that tells staff which data may go into which AI tools, what they may and may not use AI for, when a person must check the output, and who owns the rules. An AI readiness assessment is a structured check of whether your organization can execute its next AI initiative now. This lesson replaces weak versions of both with a one-page policy people can obey, a vendor terms check that makes "approved" mean something, and a readiness baseline you can defend to your steering committee.

"Are We Ready for AI?" What a Missing AI Acceptable Use Policy Costs

Picture the quarterly board meeting of an illustrative mid-market firm. A non-executive director puts down her pen and asks the question the CIO has been dreading: "Are we ready for AI?"

The CIO has two documents in the pack. The first is a readiness score of 72 out of 100 from a vendor questionnaire that one person filled in on a Friday afternoon, with no evidence asked for. The second is the AI policy. Or rather, the policies.

How many AI policies does your company really have?

Usually three, which means none. When a customer complains about an AI-drafted email, the team manager goes looking for the rule. She finds a 2023 email banning chatbots, which everyone ignores, a 14-page legal draft nobody finished, and "use AI responsibly" on the intranet. Each employee picks whichever suits them.

So the honest answer is "we do not know", and both documents hide that. The numbers suggest this firm is normal:

  • IBM's Cost of a Data Breach Report 2026 (29 July 2026) put the global average breach at USD 4.99 million, a record, up 12% from USD 4.44 million in 2025. Help Net Security's coverage of the report (30 July 2026) adds three AI findings: staff using unapproved AI tools, known as shadow AI, figured in 43% of security incidents, more than double the 2025 share; 92% of organizations reporting an AI-related incident were missing controls such as role-based access and multifactor authentication; and close to seven in ten breached organizations had no governance policy for managing AI or spotting unapproved use. These come from the gated, vendor-sponsored report via press coverage, so read them as a strong signal, not a census.
  • The Cisco AI Readiness Index 2025 (14 October 2025, more than 8,000 leaders in 30 markets) put about 13% of organizations in its top band, the "Pacesetters". It is a self-reported vendor survey.

Read the IBM findings as a pair: most breached firms had no AI policy, and almost every firm with an AI incident lacked basic access control. A policy with no control behind it is paper, so the workshop ties every rule to something a supervisor can check and every approved tool to terms someone has read.

One warning for your board paper. IBM's 2025 edition estimated that high shadow AI added USD 670,000 to a breach; the 2026 coverage compares shadow AI incident costs year on year instead. Do not write that shadow AI "adds" a 2026 dollar figure.

The legal cost has a precedent: the Air Canada chatbot ruling of February 2024, the story that opens the lesson on building an LLM golden test set, held a company responsible for what its AI tells customers. That is why policy sections 4 and 5, human review and accountability, are not boilerplate. They are the part a tribunal reads.

Why do AI policies and readiness assessments go wrong?

Policies are written to protect the author, and assessments let the person being graded hold the pen.

On the policy side, the practitioner guidance from Tines (a workflow automation vendor, so informed opinion, not a standard) lists failures you will recognize: bloated documents, vague wording such as "use AI responsibly", no named owner, staleness, blanket blocking, slow approvals, scope limited to one chatbot while browser extensions and embedded AI features go unmentioned, and "should" where "must" is meant.

Then there is the panic policy, like Samsung's 2023 ban-and-threaten memo that the shadow AI lesson takes apart. A policy written the week after an incident is a wall, a threat and a promise. Write yours before the incident.

On the assessment side, two practitioner critiques, both single-author opinion, are worth your time. A DZone article of 12 June 2026 describes most AI assessments as "self-reported, self-scored, and self-interpreted". Its sharpest example: "Do you have an AI governance policy?" returns "yes" whether the policy is an unread two-paragraph statement or an enforced framework. The second, from Usercall's blog, adds that a single blended score hides which weakness is causing the trouble.

Vendor tools also lean toward the vendor's strengths. Cisco's index gives Infrastructure its heaviest pillar weight, 25%, according to RCR Wireless (9 December 2025), and Cisco sells infrastructure. Microsoft's AI Readiness Wizard routes you to Microsoft's own adoption resources. That does not make them useless; it makes them brochures with a quiz attached.

How It Works: The Kitchen Card, the Bathroom Scales and the House Survey

You are on the last of the five lessons in Level 1, AI Aware. The previous lesson on AI literacy that meets EU AI Act Article 4 gave you a literacy plan and a training log, so most of your people already know the approved tool and the four data classes. This lesson writes those rules down and takes an honest measurement. Then Level 2 begins with fixing data quality one use case at a time, starting from the weakest area your baseline finds.

What is an AI acceptable use policy?

It is the short set of rules that tells every employee what they must and must not do with AI at work. Some firms call it a generative AI policy for employees. Think of the hygiene card on a restaurant kitchen wall.

No line cook has read the food-safety statute. What the cook reads is the laminated card by the sink: wash hands, raw meat on the red board, tell the head chef if you are ill. Each line is something you can see a person do or not do, and the card has an owner who replaces it when the menu changes.

Your policy is the card. The procedures, legal analysis and vendor contracts live elsewhere. If your draft is 14 pages, you have hung the statute by the sink.

The Tines guidance recommends six sections, which this program uses as its AI acceptable use policy template. It is vendor advice, not a formal standard; rename the parts if you like, as long as each question gets an answer.

SectionThe question it answersThe kitchen-card equivalent
1. Data classification and handlingWhich data may enter which AI systems? The most important section.Raw meat on the red board
2. Approved and prohibited toolsWhich tools, for which purpose and which class of data, on which contract terms?Buy only from inspected suppliers
3. Approved and prohibited usesWhat kinds of work may AI do here, and what is off limits?What this kitchen does and does not serve
4. Human oversightWhen must a person review output before anyone acts on it?Chef tastes before it leaves the pass
5. Accountability and enforcementWho owns an output once acted on, and what happens when rules are broken?Whose name is on the plate
6. Owner and review cadenceWhich named person owns this page, and when is it next reviewed?The head chef, and the date on the card

Four writing rules from the same guidance keep the card readable: use "must" for hard requirements, because "should" reads as optional; state observable behaviors, not principles ("do not enter customer personal data into unapproved tools" can be checked, "use AI responsibly" cannot); give department-specific examples; and move procedures elsewhere, which is how the policy stays at one or two pages.

What should you check in an AI vendor's terms before you approve a tool?

Five things, confirmed in writing: whether your inputs are used for training, how long data is retained, where it is stored, whether a data processing agreement is signed, and who carries the risk if something goes wrong. The kitchen card's "use these suppliers" is only worth something because the head chef inspected their cold chain. This check is this program's construct, not a published standard; legal sets the acceptable answers, and it lives outside the one-page policy.

CheckWhat to ask the vendorAcceptable answer (illustrative)
TrainingAre our prompts, files or outputs used to train or improve models?No, by contract, for the tier we buy. A settings toggle alone is not enough.
RetentionHow long are prompts and outputs kept, and can we delete them?A stated period, and deletion on request
Data locationWhere is data stored and processed, and by which sub-processors?Regions the privacy lead accepts; sub-processors listed
Data processing agreementIs there a signed agreement covering personal data?Signed before any personal data enters the tool
IndemnityIf output infringes someone's rights, or the vendor is breached, who pays?A position legal has read, accepted and noted

A tool that fails a check need not be banned outright; it can stay on the list for the data class it can safely take. And a vendor's terms change notice reopens the check.

What is light AI governance, and how much do you need at Level 1?

Light governance is the smallest set of people and lists that keeps the policy alive after publication: six things, no more. Committees with charters can wait for later levels.

The neutral reference is the NIST AI Risk Management Framework, a voluntary framework released on 26 January 2023 with four functions: Govern, Map, Measure and Manage. NIST says version 1.0 is being revised. The list below is this program's minimum-scale reading of the Govern function, a synthesis rather than a NIST checklist. The Level 3 lesson on AI governance with NIST AI RMF, ISO 42001 and the EU AI Act timeline grows it into a full model inventory.

  1. One named policy owner, usually the CISO, CIO or head of risk.
  2. A small group: IT or security, legal or privacy, and one business lead, meeting monthly.
  3. An approved-tools list, with the date each tool's terms were checked.
  4. A register of AI uses. You started it in Lesson 1.
  5. A fast path for requesting a new tool, with a target measured in days, which is why the terms check is a short form.
  6. An incident route: where problems are reported, and who picks them up.

What is an AI readiness assessment?

It is a structured check of whether your organization can execute its next AI initiative now. "Now" matters, because readiness is often confused with maturity.

Usercall's distinction is the clearest: readiness looks forward and asks whether you can execute now, while maturity describes capability you have already evidenced. A firm with a polished analytics team can be mature and still unready for its next project, because the data that project needs has no owner.

Two analogies for the honesty problem. The bathroom scales: the useful reading is the one someone else takes, because you will round down. The house survey: the seller's brochure says "well maintained", and the surveyor lifts the carpet and finds the damp. An honest assessment lifts the carpet.

What do you score? Cisco's index uses six pillars (Strategy, Infrastructure, Data, Governance, Talent, Culture) and Microsoft's wizard five drivers. Four areas are common to both, a usable AI readiness assessment framework:

  • Strategy: do you know what AI is for here, and in what order?
  • Data and technology: is the data for your next use fit and reachable, on tools you control?
  • Governance and security: are there rules, owners and controls that work?
  • People and culture: can your people use AI safely, and will they tell you the truth about it?

Then five honesty rules, this program's own construct, to counter what the critics describe:

  1. Evidence for every answer. No evidence, lowest score.
  2. Score what is deployed, not what is planned, the instinct behind Cisco's full, partial or no deployment scoring.
  3. Leaders and frontline staff score separately, then you compare.
  4. Keep four separate scores, never one blended number.
  5. Re-run every six months. Two readings make a direction.

Why does a readiness assessment need a prohibited-practices screen?

Because some uses of AI are banned outright in the EU, and a firm running one is not ready at any score. Article 5 of the EU AI Act, Regulation (EU) 2024/1689, has applied since 2 February 2025. It bans "the placing on the market, the putting into service or the use" of AI systems for a short list of practices. Note "use": the ban binds the firm that switches a feature on, not only the vendor.

For an ordinary employer, the practice most likely to hide in the register is emotion recognition in the workplace, banned by Article 5(1)(f) unless meant for medical or safety reasons. The Act defines an emotion recognition system as one that infers emotions or intentions "on the basis of their biometric data", meaning things like face and voice. So a call-coaching add-on that rates a rep's stress from their voice deserves a hard look, while a tool reading the sentiment of written text may fall outside the definition. Legal confirms which side a tool sits on.

The other points cover harmful manipulation or deception, exploiting vulnerabilities such as age or disability, social scoring, biometric categorisation that infers sensitive traits, untargeted scraping of facial images, and two practices that mostly concern policing. The Digital Omnibus on AI (Regulation (EU) 2026/1744) adds new bans from 2 December 2026 on AI that generates non-consensual intimate imagery of identifiable people or child sexual abuse material. The governance lesson linked above has the full timetable. Confirm every point against the Official Journal; this is not legal advice.

Clear up one word clash first: the Prohibited class in your Data-to-Tool Rule is about data that must never enter any AI tool, while Article 5 practices are about uses. Say "Prohibited data" and "Article 5 practices".

The Workshop: Write the AI Acceptable Use Policy and the Readiness Baseline in One Week

Two sessions of two hours, in one week. You already hold most of the raw material: the Shadow AI Register and Data-to-Tool Rule (Lesson 1), the AI Data Inventory (Lesson 2), the First Use Case Charter (Lesson 3) and the Literacy Plan and Training Log (Lesson 4).

ItemSession 1: the policySession 2: the assessment
Time2 hours2 hours
In the roomPolicy owner, IT or security lead, legal or privacy, HR, procurement (who holds the contracts), one business leadThe governance group, with legal staying for the screen, plus 4 leaders and 4 frontline staff who score separately
ChairThe policy ownerThe CIO or transformation lead, who reports to the steering committee
BringContracts or terms pages for every approved toolEvidence files for each area

The running example is the illustrative freight forwarder from Lesson 1, not a real company. Its first chartered use case drafts replies to shipment-status emails, with a person checking and sending each one.

Session 1: how do you write the policy in two hours?

  1. Assemble the six sections on one page. Section 1 is your Data-to-Tool Rule, with its four classes: Public, Internal, Sensitive and Prohibited. Section 2 is the approved rows of your Shadow AI Register. You are assembling, not inventing. Give this 30 minutes.
  2. Rewrite every sentence as "must" or "must not", then add examples. Apply the supervisor test: could a team lead, standing behind someone's chair, tell whether this rule was being followed? If not, delete the sentence. Then write one "you may" and one "you must not" per department, and name the human review points, starting with your chartered use case.
  3. Run the vendor terms check on every approved tool. Procurement brings the contracts, legal and security answer the five questions, and each answer is recorded with a date. A tool that fails comes off the list or drops to the data class it can safely take.
  4. Name the owner and the light governance set. One person, a review cadence shorter than a year (vendor defaults change faster than annual reviews), the monthly group, the request path and the incident route.
  5. Publish, collect acknowledgements, and teach it where training is still happening. Most Lesson 4 literacy sessions have already run, so do not pretend the policy was part of them. Add a 15-minute policy segment to the sessions still booked, such as the night-shift repeat, and to new-joiner induction; have team leads walk through the page at their next team meeting; send everyone else the page with an acknowledgement link, recorded in the training log.

Step 2 is where the policy is won or lost. In the freight forwarder's session:

First draft sentenceSupervisor testRewritten
Staff should use AI responsibly.Fails. Nobody can observe "responsibly".Deleted.
AI output should be reviewed where appropriate.Fails twice: "should" and "where appropriate".A named person must read every AI-drafted customer email, check it against the transport system, and send it themselves.
New tools need approval.Half passes. From whom, and how fast?You must request a new AI tool through the service desk form. The governance group must answer within 5 working days.

Step 3 changes the list. The enterprise tier of the approved assistant passes all five checks. A free transcription app fails on training and retention and comes off the list. A translation add-on has no signed data processing agreement, so it may take Public data only until one is signed.

The illustrative finished page, compressed:

SectionWhat the freight forwarder's policy says (illustrative)
1. DataPublic data may go into any tool. Internal data must only go into approved tools. Sensitive data (customer names, consignments) must only go into tools approved for that class. Prohibited data (for example HR case files) must not go into any AI tool.
2. ToolsThe approved list on the intranet shows the data class each tool may receive, and covers chatbots, browser extensions and AI features inside licensed software. A tool is added only after IT security and legal confirm in writing that our inputs are not used for training, retention is stated, data location is acceptable, a data processing agreement is signed and the indemnity position is recorded. Personal accounts must not be used for anything but Public data.
3. UsesYou may use approved tools to draft, summarize, translate and reword. You must not use AI to make a customs classification, to make promises to a customer about refunds, liability or delivery dates, to infer colleagues' or candidates' emotions from their face or voice, or to create images of real people without their consent.
4. Human reviewA named person must read and send every AI-drafted message that leaves the company. Team leads must log every correction they make to a drafted reply.
5. AccountabilityWhoever sends or acts on an output owns it. "The AI wrote it" is not accepted as an explanation. Breaches go through the existing disciplinary procedure. Suspected incidents must be reported to the service desk within one working day.
6. Owner and reviewOwner: the head of IT security, by name. Version 1.0. Reviewed every six months, or sooner when an approved tool changes its terms.

Session 2: how do you run an honest AI readiness assessment?

Like a house survey: anything illegal first, then private scores, evidence and conversation. Steps 6 to 9 continue from Session 1.

  1. Screen the register for Article 5 practices before anyone scores. Legal reads every register row and every planned use against the questions below. Any "yes" or "unsure" is marked "legal to confirm" and the use is suspended until legal clears it. This program adds one rule: while any flag is open, governance and security cannot score above 1.
  2. Score separately, in writing, before any discussion. Leaders and frontline staff each score the four areas from 1 to 5 on paper. If people hear the director's number first, you get the director's number eight times.
  3. Ask for the evidence behind each score: the document, the owner, and the last date it was used. No evidence, score 1. Planned but not deployed, no credit. A roadmap slide is a plan; a register with last week's date on it is evidence.
  4. Agree one score per area, pick the lowest as your Level 2 priority, and book the re-run. Record the gaps between the two groups first. The agreed score follows the evidence, not the seniority of the scorer, and the four areas are never averaged. Book the re-run, six months out, before people leave the room.
Screen questionArticle 5 pointWhere it tends to hide
Does any tool infer staff emotions from face, voice or other biometric data, other than for medical or safety reasons?5(1)(f)Call-coaching, interview and meeting tools with "sentiment" features
Does any tool use biometric data to infer race, political opinions or similar traits?5(1)(g)Photo and video analytics add-ons
Does any use score people on social behaviour, leading to unrelated or disproportionate unfavourable treatment?5(1)(c)Staff or customer "trust scores"
Does any customer-facing use manipulate, deceive or exploit vulnerabilities in a way that can cause significant harm?5(1)(a) and (b)Marketing personalisation, collections scripts
Does any image tool we allow lack controls against non-consensual intimate images or child abuse material?New 5(1)(ba) and (bb), from 2 December 2026Image generators without content controls

At the freight forwarder the screen finds one flag: a sales call-coaching add-on rates "rep confidence and stress" from voice. Legal marks it "legal to confirm", the sales director switches the feature off that afternoon, and the screenshot goes in the evidence file. The flag is closed, so the cap does not apply.

Anchor the 1 to 5 scale so that everyone means the same thing. This program suggests: 1 means nothing you can show; 2 means it exists for one use case or one team; 3 means it is in use across several teams with a named owner; 4 means it is in use across the organization and checked; 5 means it is measured and improved as routine. A firm finishing Level 1 should expect mostly 2s. That is not failure; it is an accurate reading.

Here is the freight forwarder's AI readiness scorecard.

AreaLeaders' scoreFrontline scoreEvidence shownAgreed score
Strategy31One chartered use case; no roadmap document2
Data and technology32Inventory for one use case; one red source; no shared platform2
Governance and security42Policy three days old; terms checked for every approved tool; one Article 5 flag closed; incident route not yet tested2
People and culture32Training log; 30-day spot check shows gaps; night shift not yet trained2

Read the middle columns first. Leaders scored governance 4 because they had just approved a policy; the frontline scored it 2 because the policy was three days old and the incident route untested. On strategy the floor had never seen a plan, and when asked for the roadmap, the leaders could not produce one either.

On a notepad, the leaders averaged 3.25 and the frontline 1.75. That distance is the finding of the session, and the one place an average earns its keep: as a measure of how differently two groups see the same firm. It never goes in the baseline as a score.

All four agreed scores are 2, a tie. The group picks data and technology as the Level 2 priority, because the first use case depends on it and the inventory still shows one red source and several amber ones.

What artifact do you leave with?

Two documents, the AI Acceptable Use Policy (one to two pages, with its vendor terms record) and the AI Readiness Baseline. Together they double as an AI governance policy template and an AI readiness checklist you can reuse.

The policy carries an owner, a version, an effective date and a review date less than a year away, plus an acknowledgement method. Behind it sit the vendor terms record (per tool: the five checks, who confirmed them, the date and the data class allowed) and the governance list (members, cadence, request path, incident route). The baseline has these fields:

Baseline fieldWhat to enter
Article 5 screenRows screened, flags raised, status of each flag and who cleared it
Leaders' score and frontline score1 to 5 for each of the four areas, recorded before discussion
Evidence citedDocument, owner, date last used
Agreed score and gap noteOne score per area, following the evidence; where the two groups differed, and why
Priority actionThe one thing to fix first in the lowest area
Assessment date and re-run dateToday, and six months from today

For a second opinion, run Cisco's free self-assessment or Microsoft's ten-question wizard afterwards, as a conversation starter only: no independent validation shows either score predicts AI outcomes.

Failure Modes and the Metrics to Track

Both documents decay in predictable ways, each with a tell you can check in five minutes.

Failure modeHow to spot it
Policy as statuteMore than two pages; procedures and definitions mixed into the rules
Vague verbs"Should", "responsibly" or "where appropriate" appear; no sentence could be audited
Orphan policyNo named person as owner; review date missing or passed
Narrow scopeMentions one chatbot; silent on extensions and embedded AI features
Unchecked approved listA tool on the list with no terms record, or a check older than the vendor's last terms change
Unscreened registerNo record that anyone read the register against Article 5
Flattering assessmentScores with no evidence column; leaders scored alone; one blended number reported; credit for a roadmap item that is not deployed

To test for an orphan, ask three people who owns the policy; a department instead of a name means orphan. To test for flattery, ask for the evidence column; if there is none, you are holding a brochure.

Which metrics show the policy and the baseline are working?

For the policy, track counts anyone can verify:

  • Policy exists and is under two pages (yes or no).
  • Percentage of staff who acknowledged it: acknowledgements divided by headcount.
  • Number of approved tools, and the share with a current terms check.
  • Number of registered AI uses, and open Article 5 flags (target zero).
  • Time to approve a tool request, in working days. A path that takes weeks teaches people to stop asking.
  • Date of the last policy review.

For the assessment, track the leader and frontline gap per area and the change in agreed scores at the re-run. A narrowing gap means leadership and the floor see the same firm; a score that rises with evidence can go in a board pack without blushing.

Now picture the re-run. Six months on, the illustrative freight forwarder's request path answers in four working days, the night shift is trained, and correction logs show which drafts still need work. When the director asks "Are we ready for AI?" again, the CIO hands her one page of rules and four scores with evidence. That beats a vendor's 72.

Frequently Asked Questions

What should be in an AI acceptable use policy?

Six things: data classification and handling, approved and prohibited tools, permitted and prohibited uses, when a human must review output, who is accountable, and a named owner with a review date. Keep it to one or two pages, write each rule as an observable behavior using "must", and keep procedures such as the vendor terms check elsewhere.

Should employees be allowed to use free AI tools for work?

Only for Public data, unless the tool has passed your vendor terms check. A tool that cannot confirm in writing that your inputs are not used for training, how long data is kept and where it is stored has not earned Internal or Sensitive data, whatever its features.

How do you know if your company is ready for AI?

Score four areas (strategy, data and technology, governance and security, people and culture) from 1 to 5 with evidence for every score, after screening your AI uses for practices the EU AI Act bans. Leaders and frontline staff score separately, the four scores stay apart, and a re-run in six months shows whether you moved.

What is the difference between AI readiness and AI maturity?

Readiness looks forward and asks whether you can execute your next AI initiative now; maturity describes the capability you have already shown. An organization can be mature in one area and still unready for its next project, for example when the data that project needs has no owner.

Does the EU AI Act ban any workplace uses of AI?

Yes. Since 2 February 2025, Article 5 has banned using AI to infer employees' emotions from biometric data such as face or voice, unless it is for medical or safety reasons, along with social scoring and inferring sensitive traits from biometric data. The ban covers firms that use such systems, not only vendors, so screen your register and have legal confirm borderline tools.

Key Takeaways

  • "Are we ready for AI?" deserves an answer with evidence, not a one-person vendor score and three conflicting policies.
  • An AI acceptable use policy is a kitchen card, not a statute: one to two pages, six sections, every rule a "must" that a supervisor could check.
  • "Approved" must mean the terms were read: training on inputs, retention, data location, the data processing agreement and indemnity, rechecked when terms change.
  • Light governance at Level 1 is six things: a named owner, a small monthly group, an approved-tools list, a register of uses, a fast request path and an incident route.
  • Screen the register against EU AI Act Article 5 before you score. A banned use, such as emotion recognition on staff, is a stop, not a low score.
  • An honest AI readiness assessment lifts the carpet: evidence for every score, deployed not planned, leaders and frontline scored separately, four scores kept apart, re-run in six months. The gap between the groups is a finding.
  • Level 1 ends with five artifacts and a baseline. Level 2 opens by fixing the amber and red sources in your inventory for the chartered use case.