How Much Human Review Does an AI Strategy Deck Need?

The review load does not disappear with AI drafting; it moves from producing slides to verifying claims. A three-tier protocol, source check, logic check and independent sign-off, takes roughly 6 to 12 hours for a 30 to 40 slide board deck: working estimates from practice, not survey data.

For a typical board deck of thirty to forty slides produced on an AI strategy platform, plan for roughly six to twelve hours of structured review across three distinct roles: an analyst or associate, the engagement owner, and an independent senior signatory. These are working estimates from practice, not survey data, and the next section sets out the assumptions behind them. The direct answer to the buyer's question, then, is three people and about a working day and a half of combined review time, distributed unevenly across seniority.

The review load does not disappear when the deck is machine-drafted; it moves. In a conventional engagement, much of the team's senior time goes into producing and checking every slide, and into the drafting iterations that follow. With an AI-generated deck, drafting is near-instant and the time shifts to verifying the load-bearing claims and testing whether the storyline holds. The hours are comparable in total; what changes is where they are spent and what a failure at each step costs.

• Tier 1, the source check: an analyst or associate opens every load-bearing figure at its cited source. Roughly three to six hours.
• Tier 2, the logic and so-what check: the engagement owner tests whether the storyline follows from the evidence. Roughly two to four hours.
• Tier 3, the independent sign-off: a partner or senior executive outside the engagement reads the deck cold and confirms the recommendation is defensible. Roughly one to two hours.

Whoever fills these roles needs more than spare capacity. The ICO's data protection audit framework states that human reviewers should have appropriate knowledge, experience, authority and independence to challenge decisions, and that non-meaningful human review is caused by automation bias or a lack of interpretability^[1]. The same framework asks that reviewers be assigned a manageable caseload with sufficient resource to give appropriate time to their tasks. A reviewer with the seniority to challenge but no time to open sources is reviewing in name only.

The protocol below is the checkable core of this article: three tiers, each with a named role, a bounded responsibility, a defined output and a failure mode. It is designed so that a head of strategy can assign it, an engagement owner can run it, and a partner can sign it. The fact-checking discipline behind Tier 1 is set out in more depth in a companion guide to fact-checking an AI-generated strategy deck.

The hour ranges assume a deck of thirty to forty slides with roughly fifteen to twenty-five load-bearing figures, sources that are reachable in one click rather than reconstructed by search, and reviewers who know the subject matter. They assume nothing about the tool's accuracy; they describe verification effort, not error rates. Where a deck leans on a dataset the reviewer already knows well, Tier 1 can sit at the bottom of its range; where it cites third-party market data the reviewer has never seen, it can exceed the top.

Each tier fails differently, and the failure mode decides what happens next. If Tier 1 finds a figure that does not match its source, the claim is corrected or cut and the list goes back to Tier 2; a claim with no source is flagged, not quietly fixed by the reviewer. If Tier 2 finds the storyline does not follow from the evidence, the deck is reworked before any sign-off is sought. If Tier 3 withholds its signature, the recommendation does not reach the board, whatever the schedule says. Among the ICO framework's stated expectations is that organisations maintain a log of when AI decisions are overridden by a human reviewer, including the reasons why^[1], which is the right model for the record of all three tiers.

How much of the protocol applies is set by the stakes of the decision, not by the tool that drafted the deck. The proportionality logic has a regulatory analogue: Article 14(3) of the EU AI Act, Regulation (EU) 2024/1689, states that human oversight measures shall be commensurate with the risks, level of autonomy and context of use of the high-risk AI system^[3][2]. That Article applies only to high-risk AI systems as classified under the Act, such as those listed in its Annex III, and imposes no duties on internal strategy decks. The useful point for a strategy leader is the principle, not the obligation: oversight effort should scale with what a wrong output would cost.

The pattern is that review effort tracks reversibility. A monthly update that turns out to be wrong costs a correction; a market entry recommendation that turns out to be wrong costs the capital. The protocol flexes downward for the first case and never drops below all three tiers for the last.

A fluent, well-formatted AI deck is reviewed less critically than a rough draft, and this is the best-documented risk in the whole workflow. Parasuraman and Manzey's review of complacency and bias in human use of automation, published in Human Factors in 2010, found that automation bias occurs in both naive and expert participants, cannot be prevented by training or instructions, and can affect decision making in individuals as well as in teams^[4]. Expertise does not exempt the reviewer; it can make the review feel unnecessary.

The same literature identifies when the bias deepens. Goddard, Roudsari and Wyatt's systematic review of automation bias, published in the Journal of the American Medical Informatics Association in 2011, found that environmental mediators included workload, task complexity and time constraint, which pressurised cognitive resources^[5]. A reviewer handed a polished deck at the end of a crowded week, with a board meeting the next morning, is close to the worst case: high workload, high time constraint, and an output that looks finished.

• Protect the review time in the calendar as a distinct block, the way drafting time once was, rather than expecting it to absorb into the margins of the day.
• Tell the reviewer what to look for before they open the deck: the load-bearing figures, the three most consequential claims, the recommendation's dependency on flagged data.
• Treat fluency as a property of the formatting, not of the evidence. A well-structured deck with a wrong unit conversion is still wrong.

A reviewer cannot verify what they cannot trace. The protocol above is only feasible if the platform makes three things visible, and a buyer should test for all three in a demonstration before committing.

• Clickable claim-to-source links, so Tier 1 opens the cited source in one step instead of searching for it. Without this, the three to six hours in the table above roughly double, and some claims simply go unchecked.
• A visible list of unsourced claims, so the reviewer knows where the gaps are instead of assuming there are none. An unflagged gap is the one failure Tier 1 cannot catch, because the reviewer does not know the claim exists.
• A version history, so the record of what changed between review rounds survives and Tier 3 signs off on the deck that was actually reviewed.

These features are not conveniences; they are what keeps review meaningful. Traceability acts on both halves of the problem set out above: it shortens the verification path when time is short, and it makes the basis of each claim interpretable. An auditable deck, in which each figure carries a traceable source, is what makes that possible in practice.

Two exclusions matter as much as the role assignments. First, the person who wrote the brief should not be the only reviewer. They are anchored to their own framing: the hypothesis they posed, the scope they set, the sources they supplied. They will check whether the deck answers their brief, not whether the brief was right. Tier 2 or Tier 3 must sit with someone who can disagree with the framing itself, a division of labour explored further in a companion piece on where human judgement matters in AI management consulting.

Second, the platform's own confidence indicators are not a review. A confidence score is an output of the same system under review, produced by the same model from the same documents. Reading it back to the platform as evidence tells the reviewer nothing they did not already depend on the platform for. It belongs in the same category as the deck's formatting: persuasive, and evidentially empty.

• Reviewing formatting instead of claims. Font choices and chart styles are worth five minutes, not a tier.
• Sampling a few slides and extrapolating to the rest. Errors cluster around load-bearing figures, which are precisely not a random sample.
• Treating the time saved in drafting as time removed from review. The saved hours are the review's budget, not a saving to be banked.

Decisity is an AI-native strategy consulting platform, and the protocol above is close to how its review features are meant to be used. Decks it generates carry every claim, number and recommendation clickable to its source, and it lists unsourced claims for review through its AI Strategy Engine. That is what lets Tier 1 be completed by opening sources rather than searching for them, and it is why the hour ranges in this article assume one-click traceability.

The same traceability is what lets Tier 2 and Tier 3 see what was checked and what was flagged, rather than taking the engagement owner's word for it. Interpretability of that kind is what separates a review that is meaningful from one that is nominal. The platform's three-step workflow, from document ingestion to board-ready decks, is described on its how it works page.

The protocol, not the tool, sets the review load. The tool's job is to make the review cheap enough that it actually happens: sources one click away, gaps listed rather than hidden, versions recorded. The stakes of the decision set how many tiers run, and who signs. Nothing in an AI-generated deck changes that arithmetic, and nothing should be allowed to.

Related reading:

• Can We Reproduce an AI Strategy Analysis Six Months Later?
• How to Fact-Check an AI-Generated Strategy Deck

• ICO: Data protection audit framework, Artificial intelligence toolkit, Human review
• EUR-Lex: Regulation (EU) 2024/1689 (Artificial Intelligence Act)
• EU Artificial Intelligence Act: Article 14, Human Oversight
• Parasuraman and Manzey, Human Factors: Complacency and bias in human use of automation, an attentional integration (2010)
• Goddard, Roudsari and Wyatt, JAMIA: Automation bias, a systematic review of frequency, effect mediators, and mitigators (2011)