Technologies
Publication
discovery call ↗
How to use AI at work without making things up: a review standard for small teams
- Treat AI output as a draft under review, not as a finished record, until a named person accepts it for the task at hand.
- A small-team review standard needs task types, risk tiers, required checks and a stop rule—not a vague “be careful with AI.”
- Separate internal drafts from customer-facing or system-writing actions. The second category always needs a stronger gate.
- Ban silent paste into CRM notes, contracts, quotes and public posts without a check for facts, commitments and tone.
- Train the workflow, not just the tool: who reviews, what they check, how disagreements are resolved and where approved text is stored.
How to use AI at work without making things up is less about model brands and more about review habits. Small teams get into trouble when a fluent paragraph is treated as verified, when a citation-looking sentence is left unchecked, or when generated text is pasted into a customer record because it “sounded right.” A review standard fixes the process before you scale the tool.
Bittel’s AI productivity consulting work helps teams set practical rules for everyday use. This entry is a standard you can adapt. It does not certify any model, invent accuracy rates or claim that review removes all risk.
§1What “making things up” looks like in real work
Fabrication is not always a dramatic false statistic. In small businesses it often looks like: inventing a policy the company never published; inventing a feature a product does not have; inventing a meeting outcome; inventing a source; or filling gaps in a customer history with plausible detail. The model is completing a pattern. Your job is to decide which completions are allowed to leave the draft stage.
Also watch for overconfident tone. A sentence can be wrong and still sound decisive. Reviewers should check claims, not vibe. If the task requires a number, a date, a legal term or a customer-specific fact, the reviewer must know where that fact came from—or mark it as unknown.
Shadow use is part of the picture. People will try consumer tools on work problems whether you approve them or not. A review standard that ignores shadow use pushes risk underground. Prefer an amnesty inventory: list tools in use, classify data that must not leave approved systems and offer an approved path for the legitimate jobs those tools were covering.
§2A review standard in one page
Write the standard so a new hire can follow it without a philosophy debate. Keep four columns.
| Tier | Examples | Minimum review | May write to systems? |
|---|---|---|---|
| A — Low risk internal | Brainstorm lists, agenda drafts, reformatting notes you already trust | Author skim for sense | No customer systems |
| B — Internal with facts | Process summaries, training outlines, internal FAQs | Second person checks facts against source docs | Internal wiki only after check |
| C — Customer-facing | Emails, proposals, website copy, social | Owner review for claims, commitments, tone | Send only after named approval |
| D — System / money / legal | CRM updates, quotes, contracts, filings | Dual check or prohibited without specialist review | Only with explicit authority |
Map your recurring tasks to tiers. If a task does not fit, default upward in risk until you decide. Publish the map where people actually work—not only in an employee handbook nobody opens.
Fluency is not evidence. Evidence is a source you can open.— The Bit’s review rule
§3The five checks reviewers actually run
Give reviewers a short checklist. Long policies get skipped under deadline pressure.
1. Claim check. Highlight every factual assertion. For each one: verified against a source, marked as estimate or removed. No “looks about right.”
2. Commitment check. Flag promises: dates, prices, SLAs, refunds, legal positions. AI drafts often invent helpful-sounding guarantees. Remove anything the business has not approved.
3. Identity check. Confirm names, account details and prior conversation context against the system of record. Do not let a model invent what the customer said last week.
4. Tone and audience check. Match the relationship. A stiff legal voice in a friendly local service email can be as wrong as a casual tone in a compliance notice.
5. Action check. If the next step writes to a calendar, CRM or payment tool, confirm the write is in scope and that failure handling is defined. Generated text that triggers an automation is not “just a draft.”
§4How to roll the standard out
Inventory the AI uses you already have
List tools and task types in use—official and shadow IT. You cannot govern what you pretend does not exist.
Assign tiers and owners
For each task type, name the tier and the person who may approve exit from draft. “The team” is not an owner.
Define stop conditions
Examples: unknown factual claim in customer copy; any legal interpretation; any change to pricing tables; any CRM field the reviewer cannot verify.
Practice on real samples
Take five recent AI-assisted drafts and run the checklist together. Disagreement is useful—it reveals missing rules.
Review the standard on a cadence
Update tiers when tools or products change. A quarterly review is enough for many small teams; high-change environments may need monthly.
When AI use touches company-wide authority and escalation, connect the productivity rules to a broader enterprise AI strategy. Productivity standards keep daily work safe; strategy decides what AI is allowed to do in the operating model.
§5What good looks like after ninety days
You are not aiming for zero AI use or for unchecked speed. You are aiming for predictable behavior: people know which tasks are allowed, drafts are labeled as drafts, customer-facing claims are sourced and system writes have owners. Spot-check a sample of outbound messages and CRM notes. If you find invented detail, treat it as a process defect—not only as a model defect.
Also protect learning time. A review standard that is so heavy nobody uses AI for legitimate Tier A work will be ignored. Keep low-risk paths light so the heavy gates remain credible where they matter.
Document examples of allowed prompts for Tier A and B work so people are not improvising under pressure. A short library beats a vague exhortation to “use AI responsibly.” Include what not to paste: full customer account dumps, secrets, unpublished pricing and anything covered by a client NDA unless your vendor and legal posture explicitly allow it.
When AI assists research, require a verification step before the research becomes a recommendation. A summary of a public help article is different from a summary that invents a regulation. If the task is “what does our own policy say,” retrieve the policy first and constrain the draft to that source. Retrieval without a source attachment is still a claim that needs checking.
Finally, align incentives. If staff are measured only on speed, they will skip review. If they are measured only on zero mistakes, they will avoid useful tools. Pair speed metrics with sample-based quality checks on Tier C and D work. Celebrate caught errors in review—those catches are the system working.
Install a review habit your team will actually use.
Bring two or three real AI-assisted drafts and the systems they touch. We can map tiers, owners and stop rules without inventing a bureaucracy.
QGood questions. Clear answers.
Can we let AI draft customer emails?
Yes as a Tier C draft, with a named reviewer checking claims, commitments and tone before send. Do not auto-send until that gate is proven.
What about AI that cites sources?
Treat citations as leads to verify, not as proof. Open the source. If you cannot, remove the claim or mark it unverified.
Should we ban AI for legal or financial content?
Many teams prohibit unreviewed AI for those domains or require specialist review. Decide explicitly; do not leave it to individual judgment under deadline pressure.
How do we handle staff who paste without reading?
Make the rule visible, sample work and coach. If the behavior continues on Tier C/D tasks, treat it as a performance issue—the risk is operational, not stylistic.
Does this replace an AI policy document?
It is the operating core of one. You may still need privacy, vendor and data-handling rules. A review standard without those adjacent policies is incomplete for regulated data.
discovery
call↗
Email subscriptions are not connected yet.
Book a discovery call ↗