The short answer: AI reliably automates high-volume text and data work — reading documents, drafting responses, triaging messages, summarising and translating, assembling reports — when it is integrated into real workflows with human review at the right points. It does not reliably automate judgment, responsibility or anything your business has not documented.
We build AI automations for clients, so we see both the successes and the demos that die quietly after the pilot. The difference is rarely the model. It is whether the task was suited to automation in the first place.
What works today
Reading documents so people don't have to
Supplier invoices, purchase orders, application forms, CVs, delivery notes: models extract structured data from messy PDFs, scans and email bodies with accuracy that — combined with a human approval step — beats tired humans retyping at 5 pm. This is the most common first automation we build, because the volume is measurable and the baseline error rate is known.
Answering the questions you have already answered
If the answer exists in your documentation, price lists or policies, a retrieval-grounded assistant can deliver it instantly, in the customer's language, at 2 am. The key phrase is retrieval-grounded: the assistant quotes your actual documents rather than improvising. Ungrounded chatbots are how companies end up apologising on social media.
Triage and routing
Support tickets, incoming email, reviews, applications: classifying by topic, urgency and sentiment, then routing to the right person, is boring and enormously valuable. Nobody's talent is wasted when this gets automated.
First drafts of recurring documents
Weekly reports, proposal sections, product descriptions, meeting summaries, translations: AI produces the draft, a human owns the final. Time saved is real and measurable; accountability stays where it belongs.
Copying data between systems that don't talk
Strictly speaking this is automation with AI assistance rather than AI — but the combination of workflow tools and models that can read unstructured input finally kills a lot of swivel-chair work.
What does not work
Decisions with consequences. Credit approval, medical anything, legal commitments, hiring rejections, pricing exceptions. Models make errors confidently, and the errors are not evenly distributed. Automate the preparation, keep the decision human.
Anything undocumented. AI cannot automate a process that exists only in one employee's habits. If you cannot write it down, you cannot delegate it — to software or to people.
"Replace the whole department." The pitch is common; the result is a pilot that impresses in a demo and cannot survive contact with edge cases. Automations succeed narrow-and-deep, then multiply.
Set-and-forget. Models change, your business changes, inputs drift. Automations need monitoring, measurement and an owner — which is why every workflow we ship has a dashboard, logs and an accountable human.
How to choose your first automation
Score candidate tasks on four questions:
| Question | Good sign |
|---|---|
| Volume | Happens dozens+ of times per week |
| Clarity | Success is checkable — extracted correctly, routed correctly |
| Tolerance | An occasional flagged error is recoverable, not catastrophic |
| Data | The inputs are already digital (or can be) |
The best first automation is usually the most boring one on your list. It builds the integration plumbing — connections to your ERP, CRM and inboxes — that makes every later automation cheaper, and it produces a number ("11 hours per week returned") that settles the internal debate about whether this AI thing is real.
The honest economics
Model fees for document and correspondence automations are typically cents per task. The real cost is engineering: connecting systems, handling errors, building the approval flow. That cost is one-time and quantifiable — which is why we start every AI engagement with an audit that ranks your candidate tasks by payback, including the ones we recommend against automating.
The businesses getting real value from AI in 2026 are not the ones with the boldest vision statements. They are the ones that automated invoice entry in March, ticket triage in May, report drafting in August — and measured each one.