What an Agency Owes You When It Uses AI

Every agency you talk to this year uses AI. Most will not say so unprompted, a few will say so as though it were the service itself, and almost none will tell you what actually changed about the work.

Here is what changed, in one sentence: producing a plausible deliverable went from expensive to nearly free, and checking whether it is true did not.

That single asymmetry is the whole story. It is why some agencies got much faster and much worse at the same time, and it is why the questions worth asking a prospective agency in 2026 are different from the ones that mattered in 2023.

The cost of being wrong did not fall

A confident, well-structured, entirely wrong strategy document used to take a week to write. That was an accidental safety feature. Nobody produced one carelessly, because carelessness was expensive.

Now the same document takes minutes. The writing is better than it used to be — genuinely, measurably better prose. What has not improved at all is the document's relationship with reality.

The failure mode is not gibberish. It is fluency. You get a report that cites a figure, attributes it to a plausible source, and draws a reasonable conclusion, where the figure is a subset that was never disclosed, or an average that quietly excludes the cases that would move it. Nothing about the writing signals the problem. That is precisely what makes it dangerous.

What we changed on our own site first

We applied this to ourselves before we wrote it down, which is the only order that makes the exercise honest.

In August 2026 we audited every claim on tivadigital.com.au and removed 22 that we could not substantiate. Among them: an average client ROI range, a customer satisfaction percentage, a project count, and five certification badges.

Some of those certifications are probably real. That was not enough. A badge is the single easiest claim on a website for a prospect to check, and the most damaging to be caught overstating — so the rule we settled on is that a claim we cannot evidence comes down and goes back up the day someone produces the certificate.

The client results were harder, and more interesting. We had numbers we could verify from the accounts we manage. We took them down anyway, because verifying a number and being entitled to publish it are different questions, and we had never asked the second one.

Three tests before anything gets published

Every figure that reaches a public page now has to pass all three. Most of the work is in the second.

• Reproducible — the number carries its source, the property or account it came from, and the exact date window, inline. Not "recent data" or "last quarter". A reader who has access should be able to re-run it and land on the same figure.
• Ours to publish — a client's performance belongs to the client. Trade plus city plus an exact metric is not anonymity; it is a small enough set of candidates to name the business. That needs written permission, not an assumption.
• Durable — a figure that will be wrong in three months either carries its window so it ages honestly, or it does not go up.

A number that fails any of the three gets cut, not softened. "Softened" is how an unsourced claim survives: it loses its precision, keeps its authority, and becomes impossible to check.

Verification is not a proofread

The instinct is to have someone read the output and see if it looks right. This does not work, for a reason that took us a while to accept: fluent output looks right. That is its defining property.

What works is asking a separate reviewer — one who has not been told the conclusion is correct — to try to break a specific claim. Not "check this", but "here is the assertion, find the reason it is false". The difference in what surfaces is not subtle.

In one review pass over six internal documents in August 2026, that approach returned 29 problems serious enough to stop publication. Every one of the six had been written carefully. Several of the problems were in the sentences that sounded most authoritative, which is exactly where you would expect them and exactly where a proofread does not look.

The most expensive error we found was not a wrong number. It was a correct number with a re-verification note attached, where the note only meant "this figure reproduces" and two separate readers took it to mean "this figure is safe to use".

What to ask before you sign

None of these require you to understand the tooling. They are about process, and a good answer is specific.

• Who checks the output, and are they the same person who produced it? If yes, that is not a check.
• Show me the source for one number in this proposal. Any number. The speed of the answer tells you more than the answer.
• What happens when a figure cannot be sourced — does it get removed, or reworded?
• Will my results appear on your website or in a case study, and do you need my written permission first? The correct answer is yes, you do.
• What of mine goes into a third-party AI service, and what does not?

The part that did get better

It would be a strange piece of writing that spent a thousand words on the risks and none on why anyone bothers. So: the gains are real and they are large.

Work that used to be skipped because it was tedious now gets done. Auditing every page of a site for a specific fault, checking every internal link, keeping a changelog that actually records what changed and when — these were always the right thing to do and were always the first thing dropped when a week got busy.

The right way to think about the change is not that AI replaced the expensive work. It made the boring work cheap, which moved the constraint to judgement and verification. An agency that understood that got better this year. An agency that did not got faster at being wrong.