A smarter AI isn't the same as a trusted one. Here are five checks, with owners, that tell you whether an assistant is safe to rely on and ready to scale.
AI stays a demo until people trust it. The AI Trust Checklist has five checks (privacy, accuracy, fairness, brand consistency, human oversight), each with questions, evidence, a red flag and an owner. Pilot it on one use case, tell users what the assistant uses with a trust label, check single outputs in 60 seconds, and score readiness green, amber or red. The checklist is a working tool, not a validated instrument or legal advice.
Many AI assistants stall for the same reason, and it has little to do with how capable they are. Someone tries one, gets an answer that sounds certain and turns out to be wrong, or finds that a colleague asking the same tool got something different. From then on they check everything by hand or quietly stop using it. A smarter model doesn’t fix that. A trusted one does, and trust can be built in on purpose.
Capability is what an assistant can do on its best day. Trust is whether a person will rely on it on an ordinary Tuesday, with their name on the result. Our view at Future Factors is that you don’t scale AI by making it smarter. You scale it by making it trustworthy. That is a judgment from our work with teams, not a research finding.
The AI Trust Checklist is our working tool for it. It’s five checks, each with questions to ask, evidence to look for, a red flag and an owner, and together they tell you whether an assistant, or a single answer from one, is safe to rely on and safe to scale. It turns “do we trust it?” from a feeling into questions a named person can answer. Teams use it before they launch an assistant. Individuals use a shorter version before they rely on one output.
It isn’t a compliance exercise, although compliance questions sit inside it. A company can pass every legal review and still own an assistant nobody uses, because nobody believes what it says. The data on trust and checking points the same way. In a global survey of more than 48,000 people, 46% said they were willing to trust AI systems, and 66% said they rely on AI output without evaluating its accuracy.[1] People are using something they don’t fully trust, and not checking it, which is the exact condition that turns one wrong answer into a reason to stop.
Build the trust elements in before you scale an assistant, not after its first wrong answer.
A sales manager asks an assistant to summarise an account’s history before a renewal call. It comes back tidy, well organised and confident. One date in it is wrong, and nothing in the tone gives that away. That one moment shows the shape of the problem, which comes in three forms.
“Crisp slop” is our shorthand for output that looks finished and isn’t. It isn’t a cited term, and nobody should treat it as one. The underlying problem is well documented. NIST’s generative AI profile defines confabulation as the production of “confidently stated but erroneous or false content.”[2] OpenAI’s explanation is that standard training and evaluation procedures reward guessing over acknowledging uncertainty.[3] So the confident tone is built in, which is one reason a messy draft is easier to catch than a clean one.
The cost lands on colleagues. BetterUp Labs and the Stanford Social Media Lab use the word workslop for polished-looking AI output that lacks substance and pushes the work onto whoever receives it. In their survey of 1,150 US desk workers, about four in ten had received some in the last month, and each instance took about two hours to sort out.[4] That is one survey of US desk workers, so read it as an indication, not a global rate.
An L&D lead and a new manager ask the same assistant the same question about the expenses policy, and get two different answers. Sometimes that’s a real flaw. Often it’s the design. Microsoft’s documentation says its Copilot surfaces organizational data only where the individual user has at least view permissions,[5] so two people can be answered from different material. We use that as an example of the mechanism, not as a verdict on one product. Whatever the tool, inconsistency damages confidence faster than a single mistake, because one error can be explained and a pattern can’t.
The third form is quieter. Learners and employees get generic examples, an out-of-date policy quoted as current, or a tone that shifts from formal to chatty between tools. They don’t complain. They disengage, and the assistant becomes something people open once. Each failure has a different check that catches it.
| Failure | What it looks like | How users react | The check that catches it |
|---|---|---|---|
| Crisp slop | Polished, confident output with a wrong detail inside | Check everything by hand, or stop trusting any of it | Accuracy and Attribution (check 2), Human-in-the-Loop (check 5) |
| Inconsistent answers | Different people get different answers to the same question | Argue about which answer is right, then stop asking | Brand and Cultural Consistency (check 4), Privacy and Consent Boundaries (check 1) |
| Generic content | Wrong examples, outdated sources, shifting tone | Quiet disengagement | Accuracy and Attribution (check 2), Fairness and Harm Prevention (check 3) |
A Future Factors summary of the three failure modes in this section. The mapping to checks is our judgment.
The checklist has five checks, and each one has an owner. That second part is the one teams skip. McKinsey’s 2026 survey of about 500 organizations found that those with clear ownership of responsible AI averaged a maturity score of 2.6, against 1.8 for those without.[6] It’s a survey on McKinsey’s own scale, so it shows a correlation and not that ownership causes maturity. It fits what we see as the sensible way to run this: no check is done until a role is named against it.
A check without a named role is a wish. Every check gets an owner before it gets a score.
| Check | Owner (a role, not a person) | The owner’s one-line action |
|---|---|---|
| 1 Privacy and Consent Boundaries | Data owner or privacy lead, with IT for system access | Tell users: here is what data we are using about you and why. |
| 2 Accuracy and Attribution | Knowledge owner (L&D lead, content owner or subject expert) | Label the data so the AI pulls the right sources: vetted expert material, not outdated LMS content. |
| 3 Fairness and Harm Prevention | People or HR lead, with the use-case owner | Set bias-mitigation rules and run equity checks on outputs. |
| 4 Brand and Cultural Consistency | Brand or communications lead, with the L&D design lead | Standardise how AI sounds across tools, on purpose and written down. |
| 5 Human-in-the-Loop and Auditability | Use-case owner or manager, with IT or security for logs | Give users a way to question, escalate and understand why an answer was generated. |
The five checks and their owners. Each check has four items (20 in total), scored later in this guide.
The five groupings are our working structure. They sit comfortably alongside published work: the trustworthy characteristics in NIST’s AI Risk Management Framework (currently being revised),[7] ISO/IEC 42001, the OECD AI Principles and Microsoft’s responsible AI principles all cover similar ground. The checklist isn’t a mapping to, or a certification against, any of them.
Below, each check gets the four parts: questions to ask, evidence to look for, a red flag and an owner. These are the same wording as the printable checklist.
Ask these questions
Evidence to look for
Red flag: ‘It uses everything in the CRM’ or users first learn what it knows about them when it quotes it back.
Owner: Data owner or privacy lead, with IT for system access.
The practical move is to start from the purpose and work backwards to the field. Where the GDPR applies, data minimisation (only what is necessary for a stated purpose) is the principle behind that question, and it’s a good habit even where it doesn’t. Role matters here: an HR lead worries about employee records, a sales lead about CRM notes, an L&D lead about learner histories, and each needs a different list.
Ask these questions
Evidence to look for
Red flag: Citations nobody can open, or a 2019 course sitting beside this year’s policy with nothing to tell them apart.
Owner: Knowledge owner (L&D lead, content owner or subject expert).
This check is about the system around the model, not clever prompts. A common gap is the source pile: everything in the shared drive gets connected, including the version of a policy from three reorganisations ago. For prompt-level habits, our guide to fact-checking ChatGPT covers the personal side. Here the question is whether the assistant can only draw on material someone approved.
Ask these questions
Evidence to look for
Red flag: A score or ranking about people goes straight into a decision with no human choice and no way to contest it.
Owner: People or HR lead, with the use-case owner.
The basic version of an equity check is to run the same request for different groups and read the answers next to each other. For decisions about people, the rules are getting specific. The GDPR gives people a right not to be subject to a decision based solely on automated processing where it significantly affects them, with safeguards that include human intervention and the chance to contest it.[8] New York City’s Local Law 144 requires a bias audit within one year of use for automated employment decision tools.[9] Rules differ by place, so treat these as prompts to ask your privacy lead, not as advice.
Ask these questions
Evidence to look for
Red flag: The same question gets a different policy answer in two tools, or the assistant’s tone changes depending on who asks.
Owner: Brand or communications lead, with the L&D design lead.
Different voices for different assistants are fine. Unplanned ones aren’t. The five-questions test is the quick win: pick five policy questions, ask every tool, and put the answers in one table. The mismatches tell you where to fix sources before you fix tone. For the voice guide itself, see our piece on keeping your brand voice as you scale with AI.
Ask these questions
Evidence to look for
Red flag: Nobody can say who reviews outputs, or a wrong answer cannot be traced back to its sources.
Owner: Use-case owner or manager, with IT or security for logs.
For an illustration of good oversight, the EU AI Act asks that people overseeing high-risk systems stay aware of the tendency to over-rely on output and can override it or stop the system. Most workplace assistants aren’t high-risk systems, so this is a model to borrow, not a rule that applies to you by default. The detail of writing review rules is in our guide to human-in-the-loop guidelines, and the operating routine is in managing the responsible use of AI.
A pilot is where trust gets built or lost, because it is the first time real people meet the assistant. An L&D lead choosing a first use case usually has three candidates: sales coaching, new-hire onboarding and manager support. Pick one. We’ll carry onboarding through the steps, because its sources are limited, its questions repeat and a wrong answer is usually fixable.
One use case, one owner, one group of users. The owner is whoever would have to explain a wrong answer to the person who got it. For onboarding that’s the L&D lead, working with HR operations. If you can’t name that person, the use case isn’t ready, and that’s a useful thing to learn before launch.
Write down the data used, the source rules, the tone and where humans step in. Then answer the 20 items in the checklist. Don’t try to get them all to Yes before you start. The point is to find out which ones are Unsure, and give each one a name and a date. The one-pager below is what the onboarding example looks like once it’s filled in.
| Field | Filled-in example |
|---|---|
| Use case | Answer new-hire questions about setup, benefits and policies in the first weeks |
| Owner | L&D lead, with the HR operations lead for policy content |
| Data used | Employee handbook, benefits guide, IT setup guide, org chart. No personal files, pay or performance data. |
| Approved sources | Four documents, each with a named owner and a review date. Older versions removed. |
| Rules | Answers cite the source. If the sources don’t cover it, the assistant says so and points to a person. No advice on pay, leave disputes or performance. |
| Tone | Warm, plain, on-brand. Written voice guide, one page. |
| Human steps | HR operations reviews a sample of answers weekly. Anything about pay or contracts goes to a person. |
| Escalation contact | People operations inbox, named role, same-day reply |
| Review date | First Monday of each quarter, and after any policy change |
| Go or no-go | Go for a small cohort once all four must-haves are Yes |
A filled-in example pilot one-pager. The blank version is page 5 of the free checklist PDF.
Workday’s survey of 3,200 employees at larger organizations found that nearly 40% of AI time savings were lost to rework: correcting errors, rewriting content and verifying outputs from one-size-fits-all tools.[10] Workday commissioned that research, so weigh it accordingly. It’s a good argument for starting narrow, because a pilot built around one job has fewer ways to be wrong.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Answers routine new-hire questions from the approved documents, with sources | Decides what counts as an approved source and when a policy has changed | HR operations samples answers weekly against the source documents |
| Drafts the plain-language wording of the trust label and launch note | Signs off what is promised to new hires about their data | Privacy lead reads it before launch and at each review |
| Flags questions it can’t answer from the sources | Answers those questions and fixes the gap in the sources | Escalation log reviewed at the quarterly review |
Where AI does real work, the split stays explicit. This is a Future Factors working example.
Tell people what’s happening behind the scenes: the data used, the safeguards, who to ask. People forgive a limit they were told about much more readily than one they discovered. The tool for this is the trust label in the next section, plus a short launch note that says what the assistant is for and where a human steps in.
Judge the pilot on three things, in order. Use: are people asking it real questions? Persistence: a month later, are they still using it and still opening the sources? Impact: did new hires get their answers faster and with fewer repeated questions to HR? A launch-week spike tells you about curiosity, not trust.
On who does what: training builds the capability to check an answer, and a workshop shortens that curve for a team. Workflow design gives the check a place in the work, such as a weekly sample on the HR operations calendar. Reinforcement is the manager who asks about it. Adoption is what you see when all three hold. A team can start this internally. Expert training helps people across roles apply it to their own work instead of generic tasks.
The trust label is a short note, shown wherever people meet the assistant, that says what it uses, what it doesn’t, and who to ask. It’s a low-cost trust-building move, and one many teams skip. Gallup’s tracking shows that only 25% of US employees say their organization has communicated a clear AI strategy.[11] Gallup surveys employees, so that’s a perception, but it suggests the baseline is people guessing.
| Field | Filled-in example |
|---|---|
| What it is for | Answering your questions about starting here: setup, benefits and workplace policies. |
| What it uses | The employee handbook, benefits guide, IT setup guide and org chart. Last reviewed on the first Monday of this quarter. |
| What it does not use or do | It doesn’t read your personal files, pay or performance data. It doesn’t give advice on pay, contracts or disputes. |
| How it was checked | Tested with real questions that have known answers. HR operations samples answers every week. |
| How it sounds | Plain and friendly. If it isn’t sure, it says so. |
| Where a human steps in | Anything about pay, contracts or your personal situation goes to a person. |
| How to question or escalate | Click ‘Question this answer’ or email the people operations inbox. A named person replies the same day. |
| Last updated | The date of the last review. |
A copy-ready trust label, filled in with the onboarding example. Replace each field with your own. A blank version is page 4 of the PDF.
Write it for the person using the assistant, not for the legal file. Short, plain, and specific enough that someone could prove it wrong. If you operate in the EU, Article 50 of the AI Act sets transparency duties for some systems from 2 August 2026, including telling people when they’re interacting with AI unless that’s obvious.[12] Whether it covers your assistant is a question for your privacy lead or counsel. We cover the detail in our guide to Article 50 and California’s SB 942.
You won’t run a full checklist on every answer. Most of the time it’s you, an assistant and a draft that needs to go out in the next ten minutes. The 60-second check is the one-person version: six questions, answered in about a minute, ending in one of three outcomes.
| Question | What to do in about ten seconds |
|---|---|
| 1 Stakes | What happens if this is wrong: low, medium or high? |
| 2 Source | Can I point to where each key claim came from, and did I open it? |
| 3 Specifics | Numbers, names, dates and quotes checked against a primary source. |
| 4 Fit | Right data, version, region and audience? |
| 5 Voice and fairness | Would I say it this way to this person, and does it treat people fairly? |
| 6 Owner | Whose name goes on this, and who else should see it first? |
| Result | Use: stakes are low and 2 to 4 pass. Pause and fix: a source or a specific doesn’t check out, so fix it, then run it again. Escalate: stakes are high, or the output touches someone’s job, pay or assessment, so a named person decides. |
The 60-second check. A Future Factors working tool, not a validated instrument.
The first question decides how much of the rest you do. A low-stakes brainstorm needs a glance. A figure going into a board paper needs the source opened. In Microsoft’s 2026 Work Trend Index, 86% of AI users said they treat AI output as a starting point, not a final answer.[13] The check is how you make that true in practice. The same standard applies whichever tool or model you use. Anthropic’s own guidance on reducing hallucinations lists asking for quotes and citations and restricting answers to supplied documents, then adds that these techniques reduce hallucinations but don’t eliminate them, so critical information still needs validating.[14]
For the personal habits behind this, see our guide to using AI safely at work. The check here is narrower: one output, one minute, one decision.
Score each of the 20 items: Yes is 2, Unsure is 1, No is 0, for a maximum of 40. Four items are must-haves: 1.1 (we can list every system the AI can read), 2.3 (answers show their sources), 3.3 (a person makes any decision about someone’s job, pay or assessment, and can be challenged) and 5.2 (users can question an answer and escalate to a named person). A high total doesn’t cancel a weak must-have.
Unsure scores less than Yes, because ‘we think so’ is not evidence.
| Band | Score | Must-haves (1.1, 2.3, 3.3, 5.2) | What to do |
|---|---|---|---|
| Green | 34 to 40 | All answered Yes | Go to a wider pilot |
| Amber | 24 to 33 | Or any must-have answered Unsure | Small pilot only, a human reviews every output, fix the gaps first |
| Red | Under 24 | Or any must-have answered No | Do not scale. Fix first. |
Working bands set by Future Factors. They are a judgment, and a team should adjust them. A pilot with 14 Yes, 4 Unsure and 2 No scores 32, which is amber.
Read the result as a conversation starter and a go or no-go call, not a certification. The AI Trust Checklist is a Future Factors working tool, built from published frameworks and practical experience. It hasn’t been validated or tested as a measurement instrument, the thresholds are our judgment, and we haven’t found any study showing that a trust checklist raises adoption. What the surveys do show are correlations: organizational factors outweigh individual ones in Microsoft’s data, and clear ownership goes with higher maturity in McKinsey’s. Treat the score as a way to see gaps and assign them.
Rules vary by region and by use, and this guide isn’t legal advice. For EU-facing uses, note that the AI Act’s high-risk rules for the employment uses listed in Annex III now apply from 2 December 2027 after the Digital Omnibus changes, and most workplace assistants won’t fall in that category.[12] Ask your privacy lead or qualified counsel what applies to you.
If you want the checklist as a printable, fill-in document, the free AI Trust Checklist PDF has the 20 items as a table with Yes, No and Unsure boxes, owner and evidence columns, a blank trust label and a pilot one-pager. Download the free AI Trust Checklist. It costs nothing, and you just enter your email to get it. It sits well alongside The AI Adoption Checklist, which covers the wider rollout.
Your next step is small. Pick one assistant your team already uses, name an owner for each of the five checks, and answer the four must-haves this week. The three you can’t answer tell you where to start.
A working document, not a policy. Where a policy states intentions, the checklist asks for evidence: a list of data sources, an approved-source register, a test log, a named escalation contact. A first pass on one assistant is usually a single working session with the owners in the room, and the useful output is the list of items marked Unsure, each with a name and a date.
No. Compliance sets a floor that depends on where you operate and what the system does, and it’s a question for your privacy lead or counsel. A trust checklist asks the further question of whether people will rely on the assistant. An assistant can meet its legal obligations and still lose its users through wrong answers, shifting tone or silence about what data it uses.
There are three common reasons. Permissions: tools grounded in company data answer from what each user can access. Context: the prompt, the conversation history and the connected sources differ. And variation built into how models run. One engineering study found that the same prompt at temperature 0 produced 80 unique completions across 1,000 runs, which the authors trace mainly to varying server load changing batch sizes.[15] So consistency comes from grounding answers in approved sources and writing rules down, not from hoping.
Each check has its own role, and one person is accountable for the whole assistant: the use-case owner, who would have to explain a wrong answer. Data owners and privacy leads take check 1, knowledge owners take 2, people leads take 3, brand leads take 4 and managers take 5. For a small company, one person can hold several roles, as long as the roles are written down. Senior sponsors can support, but they shouldn’t be the only name on the page.
Yes, with the 60-second check, which is built for one output. If the stakes are high, borrow two items from the full checklist: whether the answer shows sources you can open (2.3), and whether a person decides anything about someone’s job, pay or assessment (3.3). If either fails, escalate instead of relying on the answer.
Every statistic in this guide was checked against its primary page on October 5, 2026, including the KPMG and University of Melbourne trust study, the BetterUp Labs workslop research, McKinsey’s 2026 AI trust survey, Workday’s rework research, Gallup’s AI indicator, Microsoft’s 2026 Work Trend Index, NIST’s generative AI profile, OpenAI’s explanation of hallucination and the Thinking Machines nondeterminism write-up. Workday and Microsoft are vendors, and their research is attributed as such. ‘Crisp slop’ is Future Factors shorthand, not a cited term. The five checks, the trust label, the 60-second check and the readiness bands are our own working tools, informed by published frameworks, not validated measurement instruments, and not a mapping to or certification against NIST, ISO/IEC 42001, the OECD AI Principles or Microsoft’s principles. We found no study showing that a trust checklist improves adoption, so we don’t claim it. Survey findings are correlations. We left out a benchmark of model hallucination rates because it doesn’t reflect real-world error rates, and a Gartner claim about trust being the top adoption barrier because we couldn’t find it on a primary page. EU dates reflect the European Commission’s AI Act timeline after the Digital Omnibus on AI. Nothing here is legal advice. The scenes in this guide are illustrations, not client stories.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.