Systems & Execution
Most guides on how to check AI generated content are about spotting whether a machine wrote it. This one is about a more expensive problem: your team used AI on purpose, the draft arrived in seconds, and then someone senior spent the afternoon checking and fixing it. The draft was fast. The accepted piece of client work was not.
The short answer: decide what acceptable output looks like before anything is generated, give the model a small set of authoritative inputs, and put human review at the decisions where an error would cost you, rather than spreading it evenly across every sentence. Then time the whole thing, drafting, checking, correcting and approval together, and compare it with doing the work by hand. Where AI still loses after a sensible redesign, stop using it for that task.
Below is the evidence behind that answer, the conditions that change it, a review path and acceptance rubric you can adopt, a worked example of reviewer time before and after, and the rules we use to decide when to pull AI out of a workflow.
This article continues a series built on Orbflo's study of 105 qualitative discussions about AI in service businesses, collected from public Reddit communities for agencies, consultancies, MSPs, recruitment, accounting and other service work between October 2025 and October 2026. It is a purposive sample of first-hand accounts, so it tells us which problems come up and how people describe them.
Within that corpus, what we call the correction tax ranked second of 20 implementation challenges, appearing in 20 of the 80 discussions that described a challenge. People described the same loop again and again: AI drafts, someone reads closely, fixes errors and tone, re-checks the facts, then sends. One agency commenter said colleagues "end up spending more time 'fixing' the AI's mistakes" than they would have spent writing the email. A consultant was "very sick of spending like half my day every day reviewing barely-disguised Claude output."
Larger surveys point the same way. Researchers at BetterUp Labs and the Stanford Social Media Lab coined the term workslop for AI-generated work that looks polished but lacks the substance to move a task forward, pushing the real effort onto whoever receives it. In their survey of full-time US employees, around four in ten said they had received it, and each instance took close to two hours to deal with. That is the correction tax seen from the receiving end.
Those are observations of a problem. The review path, rubric and thresholds later in this article are our proposed solution. Where we show reviewer minutes, they come from a worked example, and we label them that way. None of it is a measured outcome from a client engagement yet.
When you write something yourself, checking is cheap because you already know where every claim came from. When AI writes it, the job changes from composing an answer you understand to investigating an answer someone else produced. That investigation has its own costs, and most teams never measure them.
A broad request produces a long answer padded with extra statistics, examples, promises and recommendations. Every addition is something to verify or delete. A 1,500-word proposal with ten unsupported claims can take longer to approve than a shorter one written from an agreed outline.
The model rarely knows the client's scope, past decisions, exclusions or what your firm is allowed to promise. So a senior person adds that context line by line. We made the case in Your Business Context Is Your Biggest AI Asset that this context is where most of the value sits; when it is missing at the start, it gets paid for at the end.
NIST's Generative AI Profile describes how these systems can present false content confidently, including invented reasoning and citations. Polished wording lowers a reader's guard. In the frequently cited Harvard and BCG field experiment with 758 knowledge workers, people using AI on a task chosen to sit outside its capability were 19 percentage points less likely to reach the correct answer than those working without it. The draft read well. The conclusion was wrong.
"Make this better" has no finish line. The reviewer adjusts the tone, the model changes a fact on the next pass, and now that needs checking again. Without an agreed acceptance bar, each round reopens the whole document.
A general editor can catch clumsy wording but miss a flawed tax assumption. A Microsoft Research survey of 319 knowledge workers, presented at CHI 2025, found that higher confidence in AI was associated with less reported critical thinking effort, and that work shifted towards verifying and integrating AI output. It is self-reported and shows association rather than cause, but it matches what the corpus describes: an MSP technician worried that new hires "just ask ChatGPT" and "blindly trust it."
If a five-minute draft needs 25 minutes of repair and ten minutes of senior approval, the draft time tells you nothing. People are poor judges of this. In METR's early-2025 trial with experienced software developers, AI access made tasks 19% slower even though participants believed it had sped them up. METR's February 2026 follow-up pointed towards gains with newer tools, but its own authors called the data an unreliable signal because of selection problems. Software is an adjacent field rather than client reporting, and the 2025 result is not a verdict on today's models. The lesson that transfers is to time real work rather than trust the feeling of speed.
None of this means AI always makes review slower. The strongest evidence shows real gains for bounded tasks. In a study of 5,172 customer support agents published in the Quarterly Journal of Economics, AI assistance raised issues resolved per hour by about 15% on average, with the biggest gains for newer and less skilled agents. The most experienced agents saw little speed benefit and a small drop in quality. Agents could ignore suggestions and stayed responsible for the conversation.
Similarly, Noy and Zhang's experiment with 453 professionals found ChatGPT cut writing time by about 40% and lifted graded quality by 18%. The researchers were clear about the limits: tasks had no company or client context, and nobody fact-checked the outputs. That is a good argument for testing bounded drafting, and a poor basis for assuming a 40% saving on a fact-heavy client report after expert sign-off.
Put together, the evidence suggests review stays cheap under some conditions and gets expensive under others:
If most of your rows sit in the right-hand column, the answer to "why is this taking longer?" is usually the workflow design rather than the model.
Use a short review path with a named owner. The order matters: you check whether the work is right before you spend time making it read well.
The task owner writes down the audience, the decision the output supports, required sections, length, the source pack and its date, exclusions and the reviewer. State what AI may and may not do. For example: "Draft the agreed scope from these notes; flag gaps; do not invent fees, dates, delivery promises or case studies." One consultant in the corpus said agreeing the approach before drafting "removes much of the rework."
Ask for the draft plus a separate list of material claims: anything that could change the client's decision, fee, rights, obligations or next action. For each one, the handover shows the claim, the source document and version, the page or record, the supporting passage, and whether it is a source fact, a calculation or a judgement. Unsupported items are marked unresolved. The model's list may miss things, so the reviewer still reads the draft, but now they know where to look first.
Client name, dates, amounts, totals, required fields, missing sections, word count. Use a spreadsheet, a calculator or a simple rule wherever the answer is exact, and do it before the senior reviewer opens the file. A correct total can still use the wrong inputs, so treat these checks as a filter rather than proof.
Open the sources and check the material claims. A citation is a lead, not proof. In the Stanford-led evaluation published in the Journal of Empirical Legal Studies, leading legal research tools that retrieve real source material still hallucinated in 17% to 33% of tested queries. Then look for omitted conditions, contradictory evidence and whether the recommendation follows from the facts. Only once those pass should anyone edit for tone.
Accept, return for a bounded fix, or reject and complete another way. A return names the failed gate, the affected passage, what must change and what must stay fixed. On the next pass, check the change and anything that depends on it: a corrected number may also move a total, a recommendation and the summary email.
The approver owns release. In Ayinde and Al-Haroun, the High Court in June 2025 dealt with false citations, at least some apparently AI-generated, that reached court documents, and stressed that professionals must check research against authoritative sources, including research someone else prepared. Naming a reviewer is not enough if nobody records what was checked.
Route by four things: the consequence of an error, how easy the claim is to verify, how novel the situation is, and whether the reviewer can detect a mistake. Length is a poor guide. A two-line email can contain a new commitment, and a familiar-looking template can hide a changed fee.
Give the expert a specific question rather than the whole document. For an accountant's commentary, ask them to verify the figures, the treatment and the implications, after a colleague has checked names, structure and tone. For an agency proposal, the delivery lead checks feasibility and exclusions and the commercial owner approves price and promises. A junior reviewer certifying a conclusion they cannot assess is a hidden cost, and their quick approval should never be counted as a time saving.
Use these seven gates for each client output. Pass means the test is met and the evidence is available. Fix means a bounded repair can meet it. Fail means the work has no sound basis or cannot be approved safely within the task. A factual error in a core recommendation, an unsupported authority, a confidential data breach or an unauthorised promise blocks release on its own. Mark a gate not applicable only with a reason.
AI drafts a £12,000 research proposal from approved discovery notes. The draft promises six-week delivery, which appears nowhere in the notes. Source support and coverage both fail on that promise, and fluent wording does not offset either. The delivery lead supplies a feasible date, the commercial owner approves it, and the reviewer checks the scope section and covering email for knock-on changes.
If the notes also lack the client's required outcome, the brief goes back to the owner before anyone drafts again. That is a missing-input problem, and no amount of rewriting will supply the answer.
Review catches errors. It does nothing to stop the next one. The cheaper fix is to change the inputs and the shape of the task so fewer errors reach the reviewer. In the corpus, 14 discussions raised source-data problems, including AI pulling from outdated process documents, a failure that no amount of output checking will find if the reviewer trusts the same stale file.
Current scope, approved facts, client context, examples and rules, in one place. Remove superseded files. Put an owner and a review date on anything that changes. For recurring updates, pull fields from the live record rather than asking the model to recall account history. A larger prompt stuffed with old and conflicting files adds confusion.
First extract dates, decisions, actions and figures with their source locations, and check them. Then draft from the checked fields. Morgan Stanley's Debrief tool follows a version of this: with client consent it records meeting notes and action items, drafts a follow-up email for the adviser to edit and send, and saves a note to Salesforce. The adviser stays responsible for the message. That is a company description of the workflow, not a measured time saving. For advice work, let the expert set the conclusion before AI expands it into client prose.
Tell the model to flag missing or conflicting evidence and stop at the edge of the source pack. "Not stated in the notes" should count as a useful handover. Test this with cases where the right response is to withhold an answer.
Stable clauses, exclusions, policy statements and standard next steps should be approved wording that AI selects rather than rewrites. Where there is little variation, a fixed template or form often beats AI outright.
Log each error type and cause: wrong record, stale rule, unsupported claim, bad calculation, omitted condition, weak judgement, style mismatch. Change the input that caused it. A prompt reminder does little if the wrong policy is still in the source pack.
A second model can flag a changed date, a missing section or a sentence with no support in the supplied sources. Test it on known errors first and watch for false alarms. Agreement between two models is not independent evidence, since they can share blind spots, and research on self-correction found that models asked to fix their own reasoning without external feedback often failed to improve, or got worse, in the settings tested. Retrieval needs its own test too: Anthropic's contextual retrieval experiments cut retrieval failures substantially, but those are retrieval metrics, not client error rates. For a small firm, the first check is simpler: did the right, current document reach the draft at all?
The Deloitte report for Australia's employment department shows what happens when citation checks are skipped. AP reported fabricated references and a made-up court quotation; Deloitte agreed to repay the final contract instalment, and the revised report disclosed generative AI use, though Deloitte did not confirm AI caused the errors. The department's report page records a further correction in February 2026.
Here is a worked example for a two-page monthly client update, the kind an agency, consultancy or MSP sends every month. The figures are illustrative, built to show the mechanics, not measured from a client. All times are active human minutes per accepted update, and the same acceptance gates apply in every column.
The manual route uses checked account records and a normal approval pass. The broad AI route starts from a loose prompt and sends a long draft to the account manager. The redesigned route uses a fixed brief, approved source fields, a short draft, an evidence handover and a bounded correction pass.
Minutes per accepted client update (illustrative)
Look at the middle column. The broad AI process saves two minutes of total effort against manual work, which is why it can look like a win on a timesheet. But reviewer time more than doubles, from 23 to 48 minutes, and the reviewer is usually the scarcest and most expensive person on the account. That is the correction tax in miniature.
The redesign cuts reviewer time to 19 minutes and total effort to 37 minutes: 35 minutes, or about 49%, less than manual, and 33 minutes less than the broad AI route. It spends more time on inputs than either alternative, and that is the point. A few extra minutes preparing the brief removes far more repair later.
Say the redesign takes four extra person-hours to set up. At a 35-minute saving per update, it pays back after 240 ÷ 35, or about seven updates. Add 30 minutes of source-pack upkeep for every ten updates, three minutes each, and effort rises to 40 minutes, the saving drops to 32, and break-even moves to about eight updates. For work you do twice a year, the setup may never pay back.
Mean human effort per accepted output = all human minutes spent on the task stream ÷ outputs accepted to the agreed bar. Include rejected drafts, abandoned AI attempts, manual rescues, source upkeep and later corrections. Keep setup separate and show how it is allocated. This is the same logic as the cost per completed outcome measure we set out earlier in this series, where human review cost usually decides whether the business case holds. Also track senior review time, first-pass acceptance and serious defects. Ten minutes moved from a junior to a scarce expert can lower total minutes and still make the workflow worse.
Stop the current attempt when evidence is missing, a critical claim cannot be checked, the output would exceed someone's authority, or a serious data or client risk appears. Route the work to the owner or an expert. Sometimes fixing the brief or narrowing the AI's role is enough; generating another full answer rarely is.
Set a repair budget from the manual alternative. If finishing the piece by hand would take 20 minutes, including its own check, and another AI pass needs a fresh five-minute check of whatever it changes, the AI repair has to take under 15 minutes to be worth trying. Time already spent goes in the log, but it should never justify one more unbounded round of prompting.
For pilots, we suggest these rules. They are management choices, not research findings:
Continue only where the workflow meets the quality floor and produces a real gain in effort, cost or lead time. A deliberate quality improvement can justify more time, but call it a quality investment rather than a time saving. Being honest about which workflows to drop is part of why most AI pilots never scale: teams keep expanding tools that never earned their place.
Pick one recurring output with steady volume and a stable brief. Name the owner and reviewer, agree the rubric, define a critical defect, and set a target before you start. A reasonable starting target is at least 15% less total effort per accepted output with no rise in serious defects.
In the first two days, collect manual baseline timings using the same stages and quality bar. From day three to seven, run comparable tasks through both routes, rotate staff, and where practical keep the final reviewer unaware of which route produced each piece. Twenty to thirty outputs per route is a practical starting sample if your volume allows; it is not a statistical power calculation and cannot prove safety against rare serious errors. In days eight to ten, compare quality at the same bar, total and senior time, rework and cost, broken down by task type and staff experience. Keep, narrow, rework or stop.
If you are still choosing tools for the pilot, our guide on how to choose AI tools for your service business covers the selection side.
Bring us the workflow that eats your review time
If one client deliverable keeps coming back for another round of fixes, bring a recent example to a free 30-minute working session. We map how it runs today and sketch a redesign: where plain automation beats AI, where AI can take responsibility, where a human stays in control, and the measures that would show it works. If AI is not the answer, we will say so.
Yes, as a filter for mechanical issues such as changed dates, missing sections or claims with no support in the supplied sources. Test it on examples with known errors before relying on it. It cannot provide independent assurance just by agreeing with the first model, and self-correction research suggests models often fail to fix reasoning errors without outside evidence. A named human still owns release.
It reduces the problem without removing it. The legal research tools tested by Stanford researchers retrieved real source material and still produced hallucinated answers in 17% to 33% of queries. Grounding in a current, owned source pack helps most when the reviewer also checks that each cited passage says what the draft claims.
Check your contracts and professional rules first. In our corpus of service-business discussions, a PR agency described telling clients about its AI use at onboarding, which can take much of the trust risk out of the relationship. Disclosure works best alongside a clear account of who checked the work and against what.
The AI Operating System Scorecard is a diagnostic tool that measures whether your business is structurally built to make AI compound, across nine dimensions including how decisions get made, how clearly your processes are defined and how your team is using and integrating AI.
The output is a clear view of where your biggest leverage gaps are and where to focus first.
One practical AI operating-system insight every week.
No fluff, no spam.
© 2026 AV Ventures Limited trading as Orbflo | All Rights Reserved
