AI Training or AI Enablement? How to Stop the Workslop at Source

In [The Rise of AI Workslop: Are You Throwing Slop Grenades?](/insights/ai-workslop-slop-grenades/) I looked at what happens when AI-drafted work lands on a colleague's desk half-finished. This piece is about stopping it being made in the first place. The usual response to sloppy AI output is to book some training. That helps, up to a point. A two-hour session on prompting is lively on the day and mostly forgotten by the following week. People go back to typing a one-line request into a chat window and sending whatever comes out. Camunda's *AI Process Gap* research, published this month, shows where that leads. In its survey of 5,000 employees at large organisations, 24% said they spend time fixing AI-generated work passed to them by colleagues, and 28% said they have to correct or redo AI output because the process wasn't set up for it. As I said last time, it's a vendor-commissioned enterprise survey, so treat the numbers as indicative. But the rework is real, and in a small team it lands on fewer, busier people. ## AI training versus AI enablement Training teaches people how the tool works. Enablement gives them what they need to use it well on real work, every time, without having to remember a course. It's the voice guide, the source material, the brief templates and the review rule that sit alongside the tool. You need some of both. Weight it towards enablement, because that's what's still there on a busy Thursday afternoon. ## Give it your voice once Most AI output sounds generic because nobody told the tool how your business writes. Fix that one time, not every time. Write a short brand voice document: how you sound, words you use and avoid, how formal you are with clients versus colleagues, and one or two examples of emails you'd be happy to send. Then save it where the tool can use it by default. Most mainstream AI assistants offer some way to store standing instructions or reference files, though the feature names and limits vary, so check what yours supports. Nobody on your team should be re-explaining your tone from scratch. ## Give it the facts, not just the task Much workslop is the AI filling gaps with plausible guesses. The technical term for the fix is grounding, which in plain English means handing the AI the actual material and telling it to work from that. Paste in the client email, the meeting notes, the quote, the policy. Then say: use only what's here, and if something isn't covered, say so rather than filling it in. This cuts down on invented detail considerably. It doesn't remove it entirely, which is why the next steps matter. ## Make it list its sources Ask for every source it used to be listed at the bottom of the draft. This turns checking from a vague re-read into a quick tick-through. It also goes at the trust problem directly. Camunda found that the biggest single reason AI hadn't saved employees more time was that they didn't trust the output, so they double-checked everything by hand (36%). The report's explanation is that when a process can't show how it reached an answer, people have no reason to believe it. A source list shows the working. One warning: open every source. AI tools can cite documents that don't exist or claim a source says something it doesn't. The list makes checking faster. It doesn't do the checking for you. ## Ask it where to look You can ask the AI to flag the parts of its draft that need careful reading: figures, names, dates, and anything it has inferred rather than taken directly from the material you gave it. Be clear with your team about what this is. It's a pointer, not a safety net. The model can miss its own errors, and it can sound equally confident whether it's right or wrong. Use the flags to decide where to start reading, not as permission to skip the rest. ## Tell it what good looks like A lot of slop is simply too long. Left alone, AI will give you five paragraphs where your colleague needed three lines. So define the shape before you ask. For an internal update, a brief like this works well: > Write an update for [colleague] using only the notes below. > Start with a one-sentence summary, then up to three key points, then any actions with who and by when. > No more than 200 words. > List the sources you used at the bottom. > Flag anything you've inferred or aren't sure of, so I can check it. That gives your colleague something they can read in a minute and gives you something you can check in two. Build a handful of these for the things your team writes most often, such as client updates, meeting follow-ups and quote covers, and keep them where everyone can find them. If you want a ready-made starting point, we've written about [defining what good looks like before you prompt](/insights/ai-rubric-what-good-looks-like/). ## Make review mandatory Nothing AI-drafted leaves the sender without a human read. That's the rule, and it needs to be stated plainly rather than assumed. Everything above exists to make that read quick. A short draft, in your voice, built from real material, with sources listed and uncertain bits flagged, can be checked properly in a couple of minutes. Scale the effort to the stakes: an internal note gets a careful read, while a client proposal or anything involving money or commitments gets the figures checked against the originals. ## Then train, on your own work Now training earns its place. Run it with your team's real examples, not generic demos. Bring a recent piece of workslop, anonymised, and rebuild it together using the voice guide, the source material and a brief template. People learn more from fixing one of their own documents than from an hour of slides. Involve them in building the kit too. Camunda found 69% of employees weren't fully consulted about how AI would be used in their day-to-day work. Your team knows which messages cause the most rework. Ask them which templates to build first. ## What you get for the effort A one-page voice guide, a few brief templates and a clear review rule add up to a good afternoon's work. What comes back is drafts that arrive shorter, sound like your business and can be checked quickly. Colleagues stop quietly fixing each other's output. And AI starts doing what it was bought to do, which is taking work off people's desks rather than moving it around. If you'd like this kit built with your team rather than for them, that's exactly what our [training workshops](/services/training-workshops/) are for — or [get in touch](/contact/) and we'll talk it through. ## FAQs ### What's the difference between AI training and AI enablement? AI training teaches people how an AI tool works, usually through a session or course. AI enablement gives them what they need to use it well every day: a brand voice guide, source material, brief templates and a clear review rule. Training is an event. Enablement is the set-up that makes good habits the default. ### What does grounding mean in AI? Grounding means giving an AI tool the actual material to work from, such as emails, notes or documents, and instructing it to base its answer on that material rather than on general knowledge. It reduces invented detail but doesn't eliminate it, so a human still needs to check the output. ### Can I rely on AI to flag its own mistakes? No. Asking an AI to flag figures, names and inferences it's unsure of is a useful way to direct your checking, but the model can miss its own errors and may sound confident when it's wrong. Treat its flags as a starting point for review, not a replacement for it. ### How long should an AI-drafted internal message be? As short as the reader needs. For most internal updates, a one-sentence summary, up to three key points and any actions, in under 200 words, is enough for a colleague to act on and for the sender to check properly before sending.

This article was written by Peter Lowe. The ideas and opinions are his own; AI was used to assist with drafting and editing.