The rule we use

A job gets automated at Calibre when it passes three tests: it is a repetitive task, it has a definition of done a machine can check, and a mistake is cheap to catch. Fail any one of the three and the routine task stays with a person. That single rule has saved us from most of the expensive failures we see in client audits.

Invoicing chases pass the test. Client emails do not. Retouching triage passes. Final colour never will.

What AI automation means, without the vendor gloss

Strip the labels and it is this: artificial intelligence, mostly large language models with some machine learning and natural language processing underneath, now handles work that used to need a person's judgement, not just their keystrokes. That is the difference from traditional automation and classic business process automation, which follow fixed rules and break the moment reality shifts. AI-powered automation reads context, analyses large datasets quickly, and picks the trend out of the noise. The enterprise brochures call it intelligent process automation and illustrate it with supply chain management, predictive maintenance and inventory management. A studio's versions are humbler: the data entry nobody admits doing, reporting on marketing campaigns, first-pass triage of customer interactions. Same principle, closer to home.

What runs without us

A sample of the AI automations currently running our business operations, because a list of real jobs is more useful than a framework.

Search and site monitoring. Our indexes are checked daily, new pages get pinged to search engines the day they publish, and site errors are flagged in real time before a client stumbles on them. This layer exists because we once found a broken deploy a week late.

Reporting. Ad spend, traffic and lead numbers are pulled from existing systems, reconciled and written up in plain English on a schedule. Nobody at Calibre builds a weekly report by hand, and no client of ours should accept paying agency hours for one.

Production hygiene. File naming, delivery folder structure, backup rotation, the scaffolding that collapses when a studio gets busy. Machines do not get busy, and they do not make the copy-paste class of human error.

First drafts. Briefs to shot lists, transcripts to summaries, research to outlines. Drafts only. Nothing written by a machine leaves this studio without a person rewriting it, which is also the policy behind every word on this site.

What we keep human, on purpose

Anything a client sees before it is finished. Anything with taste in it: edit selects, colour, casting, the crop. Anything where the relationship is the point, a difficult email, a scope conversation, a no. The flagship example in every vendor deck is customer service: AI chatbots answering inquiries 24/7 with instant, personalised responses, response times down by two thirds. Fine for order status and opening hours. But a chatbot handling triage is not the same thing as a chatbot handling the relationship, and brands that confuse the two pay for it in customer experience. We deploy chatbots for the first sort of conversation and keep people on the second. We sell judgement; process automation exists so more of the week is spent on it.

The honest barriers

Four things slow AI adoption in real companies, and none of them is the model. Implementation cost, which is why we start with one small build instead of a platform. Data privacy and security, which is a governance decision to make before the build, not after: what the system can read, where it runs, who audits it. Integration, because wiring AI into existing systems is genuinely complex once real permissions and legacy tools are involved, and scaling past a pilot needs proper infrastructure behind it. And resistance to change, which is mostly rational: 65% of desk workers say AI will free up their time, but people resist tools imposed on them and adopt tools that remove work they hate. Start with the second kind.

How this shows up in client work

The same layer we run internally is what we build for clients under Intelligence. The pattern that works: we spend a day mapping where the hours actually go, pick the two or three business processes that pass the rule above, and ship those first, working with your IT team where systems need opening up. Working automations beat roadmaps. For the autonomous, long-running pieces, AI agents covers how we structure those and where we draw the line.

The sales pitch for all of this is that software runs 24/7 and cuts operational costs. True as far as it goes. Two honest numbers matter more. Most teams' first useful automation is live within a week, because the first job is always overdue and obvious. And roughly half of every engagement is not AI at all, it is cleaning the existing systems the automation has to touch: the naming, the permissions, the four places the same file lives. Anyone who quotes you an automation project without asking about your file structure is guessing.

Where images fit

For product and campaign work, automation now reaches into the imagery itself: batch clean-up, format variants, first-pass grading before a human touches the file. What survives contact with a real brand and what does not is its own post, AI product photography, and the standard we hold that work to is the same one we apply to retouching: if you can tell a machine did it, it is not finished.

If you are starting from zero

Do not buy a platform, and do not start with a dashboard full of predictive analytics. Pick the one repetitive task your team complains about most, automate that workflow, and measure the hours for a month. If it saved what it promised, take the next task. If it did not, you have lost a week, not a year's budget. When the list of working automations gets long enough that it needs an owner, that is the point where companies call us, and the honest ones we tell to wait another quarter.