AI training for employees, built around the work they already do
Most teams do not need to be told that AI exists. They already use it, usually on a personal account, often on the tasks where a wrong answer costs the most. The gap is judgement: what to put in, what to check, and when to stop and ask someone.
Sessions are run by the person who does the implementation work, which is why the examples are real handoffs, real documents and the specific ways output goes wrong, rather than a slide deck about how language models are trained.
Training tracks by role
An AI workshop that treats a support agent and a marketing lead as the same learner will bore one and mislead the other. They reach for AI for different reasons and break it in different ways. Tracks are picked from this table, then narrowed once we see the team's real tasks.
| Role | What they try to use AI for | The failure mode this role hits | What the session covers |
|---|---|---|---|
| RoleManagers | What they try to use AI forSummarising threads, drafting board and team updates, reviewing others' work. | The failure mode this role hitsApproving output because it reads well. Fluent prose passes a skim that a clumsy but accurate draft would fail, so errors survive review precisely because the writing is good. | What the session coversReading for claims instead of tone, and deciding which tasks may reach a customer unchecked. |
| RoleOperations and admin | What they try to use AI forMeeting notes into tasks, reformatting spreadsheets, process notes, chasing missing form fields. | The failure mode this role hitsSilent data mangling. Ask a model to reformat 200 rows and it quietly drops, reorders or invents a few, because it produces plausible text rather than copying a table. | What the session coversWhich reformatting jobs belong in a formula instead, and checking row counts and totals before a file moves on. |
| RoleCustomer support | What they try to use AI forDrafting replies, summarising ticket histories, softening difficult messages. | The failure mode this role hitsPasting the whole customer email into a personal AI account. Name, address, order number, and whatever the customer volunteered about their health or finances. | What the session coversThe paste test below, fast redaction, drafting from approved knowledge-base text, and what to do when the answer is not in it. |
| RoleSales | What they try to use AI forFollow-ups, technical questions in proposals, discovery call summaries, tender responses. | The failure mode this role hitsLetting the model invent capability. Asked whether the product does X, it returns a confident yes with a feature name that does not exist, and that sentence lands in a written commitment. | What the session coversA capability sheet the model may draw from, marked roadmap versus shipped, and never answering a spec question from an empty prompt. |
| RoleMarketing and content | What they try to use AI forPages and posts, topic research, repurposing articles, product copy. | The failure mode this role hitsAccepting confident but unsourced claims: statistics with no origin, a "study" that resolves to a homepage, a quote nobody said, a competitor comparison built from stale training data. | What the session coversSource-first drafting, and verifying every figure and citation before publication. |
| RoleTechnical or ops leads | What they try to use AI forScripts and queries, debugging, documentation, assessing tools for others. | The failure mode this role hitsOver-trusting code that runs. It executes cleanly and does the wrong thing on the edge cases. Also pasting connection strings and keys into a chat window mid-debug. | What the session coversReviewing generated code against failure cases, not the happy path, and credential hygiene while debugging. |
AI upskilling needs differ far more by role than by company, which is why sessions are scoped by role rather than by headcount.
What a session looks like
The standard format is a half day, roughly three hours with a break, for eight to twelve people in the same or adjacent roles. The range is deliberate. The exercises depend on people cross-checking each other's output, which needs a few of them, and on being able to get round the room while they work, which caps the number.
Delivery is remote by default, on whatever conference tool you already run. On-site delivery is available across the UK, and in the EU where two or three sessions justify the travel. A company wanting an AI training course in London for several teams is better booking consecutive days than spreading them over a month.
People bring a laptop, access to the AI tool the company intends them to use, and two or three real tasks from their own week. The email they have been avoiding, the report they rebuild monthly, the ticket type that always takes twenty minutes.
They leave with a one-page reference for their role, a written paste rule, their session prompts saved where the team can find them, and a short list of tasks we agreed AI should not touch in their department.
Two shorter formats exist: a 90-minute manager briefing on approval rules and personal-account usage, and a single-workflow workshop where a small group builds one repeatable process end to end. An optional 60-minute follow-up runs a few weeks later, once the questions have become specific.
Exercises built on your own work
Generic exercises produce generic habits. Before the session you send a few real artefacts with sensitive detail removed, and the exercises are rebuilt around them.
- Rewrite a real replyAn agent redacts an actual ticket live, drafts from approved knowledge-base text, then swaps with a colleague who marks anything the source does not support.
- Rebuild the monthly reportTake last month's report and its source spreadsheet, produce the narrative, then check every figure back against the sheet. The checking is the exercise.
- Break a prompt on purposeAsk about a product capability where the correct answer is "not supported", and see whether the model invents one. This is the exercise aimed at sales and pre-sales roles.
- Redaction drillA stack of real inbound emails, five minutes, decide for each whether it can be pasted, must be stripped first, or should not go near an AI tool at all.
Where material is too sensitive to bring, the parallel gets built from our own reports, handoffs and support queues instead. Nothing in an exercise ever comes from somebody else's company. No anonymised tickets, no redacted contracts, no samples carried over from another session.
What not to paste into an AI tool
Long policies do not get consulted mid-task. People need a test they can run in three seconds, with a finger already over Ctrl+V.
Would you be comfortable if this exact text appeared in a screenshot on a stranger's blog, with your company name visible? If the answer is no, or if you have to stop and think about it, it does not go in.
The categories underneath the rule:
- Personal dataNames, addresses, phone numbers, dates of birth, NI or ID numbers, health details, anything about someone's finances or family. That includes the free-text box where a customer explained why they want a refund.
- Customer-identifying detailOrder numbers, account references, ticket IDs, company names in a complaint, and any combination of small details that identifies one person without a name.
- CredentialsPasswords, API keys, tokens, connection strings, .env contents. A key pasted into a chat window has been disclosed, and the fix is rotating it, not deleting the chat.
- Unpublished commercial termsDraft pricing, discount floors, margin, tender responses before submission, restructuring plans.
- Anything under NDA, and anything a client gave you that you could not forward to a third party.
Tooling matters as much as habit. A consumer chat account and a business tier are different products on different terms. Business and enterprise plans are normally offered with a data processing agreement, admin controls, and stated terms on whether inputs are used for training. A free personal account generally is not. That difference is contractual and invisible on screen, which is why staff must be told which account to use and given access to it. Read the terms for the plan you are on, and have your own advisers confirm they cover your processing.
Checking output before it goes anywhere
Checking is a different job from proofreading, and the order matters, because attention runs out.
- Identify the load-bearing claimsThese are the sentences that cause a problem if they are wrong: a price, a date, a contractual promise, a technical capability.
- Verify those first, against the source, one at a timeA second AI answer does not count as a source.
- Trace every number back to where it came fromIf a figure in the output is not in the source, it was invented.
- Open every URLBroken, redirected, or landing on a homepage instead of the cited page all mean the citation is decorative.
- Read for what is missingModels rarely announce an omission, so compare against your own checklist for that document type.
- Only then read for tone and grammar.
The failure worth drilling is the confident answer with nothing behind it. The tells are consistent: no hedging where a knowledgeable person would hedge; suspiciously round numbers; "a study" with no name, year or author; a named report that returns nothing when you search it; quotation marks around wording that reads like the model's own prose. Challenge it and you get an apology followed by a different confident answer, which needs checking as hard as the first.
One habit outperforms the rest. If the output contains a fact you did not already know and you cannot find its source in two minutes, delete the sentence. Do not soften it. Delete it.
Writing an AI policy people follow
Policies get followed when they fit on one screen and answer real questions. A fuller version, with each clause marked as legal obligation or house rule, is published in the AI policy template. The draft from a session usually covers six things: which tools and account tier are approved, the paste rule, which tasks need a named human to check output before it leaves the company, how to declare AI use where a contract or professional body requires it, what to do when something goes wrong, and who owns the document.
The last of those is the one to insist on. A policy with no named owner goes stale within a quarter, because the tools change and nobody is responsible for saying so.
Declaring AI use is the clause most likely to be a legal obligation rather than a house rule, and the trigger is narrower than most coverage suggests. What the EU AI Act actually requires of a UK business sets out where the line falls.
Two things kill a policy on arrival. Banning AI outright moves usage onto personal phones where you cannot see it. Rules so broad that nobody can tell whether their own task is covered produce the same result, more slowly.
What comes out of a session is a working draft written by someone who has watched your team work. Legal review, HR consultation, works council processes where they apply, and formal sign-off belong to your own advisers.
What we leave to your specialists
This is practical AI training for business use, and it stops where other people's professional responsibility begins. We do not certify anything as GDPR compliant, sign off DPIAs, or award accredited qualifications. Nobody leaves with a certificate that means anything to a regulator. A data protection question that needs a real answer goes on a list for your DPO or solicitor rather than getting an improvised answer from us.
Employment matters stay with HR: performance management, consequences for policy breaches, consultation obligations, and any decision about how AI adoption affects roles. Sector rules stay with whoever holds them. Financial services, healthcare, legal practice and education all carry obligations a general workshop has no business interpreting.
Procurement is yours too. We explain what separates account tiers and which questions to ask a vendor, but we do not resell licences or take commission. If a session surfaces real infrastructure work (private model hosting, network isolation, dedicated servers), that is a separate conversation with a different company. YNVAR does the hosting side.