Deliverable QA before it reaches the client: automation as a floor

Deliverable QA is the quality argument in our automation ranking, not the speed argument, and it is the one we would make hardest to a sceptical owner — because it is the automation that makes your work better rather than merely faster, and it does so by taking over the part of review nobody was doing well anyway.

The checkable layer

Every deliverable that leaves an agency has two layers. One requires taste: is the strategy right, is the writing good, does this serve the client. The other requires only attention: is there a broken link, is the client’s name left over from the template it was cloned from, does the figure in the executive summary contradict the figure in the table, is there an image without alt text, is there a claim with no source attached.

Nothing in that second list needs judgement. All of it is embarrassing in front of a client, and the embarrassment is disproportionate — a wrong client name in a header can undo the credibility of twenty correct pages, because the client reads it as nobody looked at this.

Here is the uncomfortable truth about human review: senior reviewers are bad at the checkable layer, and they are bad at it because they are good at the judgement layer. Attention spent verifying totals is attention not spent on whether the recommendation is right, so reviewers skim the mechanics to save themselves for the substance — which is the correct trade, made silently, with the mechanical errors as the accepted cost.

What the automated pass does

Before any deliverable is marked ready for review, a checking pass runs: links resolved, names and dates consistent with the brain’s record of the client, figures cross-checked between summary and source tables, template remnants hunted, claims matched to sources, the house checklist for that deliverable type applied. What comes back is not a grade — it is a short list of specific findings, each one pointing at a place in the document.

The reviewer then does the only review worth a senior hour: reading the work for whether it is any good. The floor is guaranteed; their attention goes to the ceiling. The human stays in front of everything that ships — that is permanent — but what the human is for changes from proofreading to judgement.

The by-product nobody expects

Run automated checks for six months and you acquire something most agencies have never had: evidence about your own error patterns. Which mistakes actually occur, in which deliverable types, from which processes, how often.

Every agency has strong opinions about its quality problems. Almost none has data. Ours surprised us — the errors we were sure we made rarely appeared, and a category nobody had named (figures that drift between a document’s summary and its body during revision cycles) turned out to be the most common failure by a distance. That one finding changed how we structure revisions, which is a process improvement no amount of “be more careful” would have produced.

The error log also answers a question clients increasingly ask: how do you keep the AI honest? “Every deliverable passes an automated check, a named person reviews every finding, and here is the kind of thing it catches” is a better answer than any policy document.

What not to automate here

Resist the temptation to let the checker fix what it finds. A system that silently corrects figures is a system that will someday silently “correct” a right number into a wrong one, and nobody will have seen it happen. The checker flags; a person changes. The whole design depends on the finding being visible on its way to being fixed.

And do not let the checklist grow unbounded. Every check added is noise added for the deliverables it does not apply to, and the same fatigue mechanics that ruin monitoring apply to QA findings. Twenty checks that are always worth reading beat eighty that train reviewers to scroll past.

Where to start

Take your last ten shipped deliverables and check them manually against a list of the ten mechanical errors you fear most. The ones you actually find are your starting checklist — not the ones you guessed. Automate those checks on one deliverable type, run it for a month, and let the findings log tell you what to add. Like everything early in the sequence, the point is a boring, reliable win that the sceptics on the team cannot argue with: fewer embarrassments, better use of the senior hour, and a record to prove both.