Less impressive than the demos, considerably more useful than the sceptics expect.
Less impressive than the demos, considerably more useful than the sceptics expect.
The reliable value is in the work nobody wants: summarising enquiries, drafting first-pass replies, extracting structured data from documents, tagging and routing incoming requests, turning meeting notes into actions.
None of that makes a good keynote. All of it gives hours back every week.
It goes wrong when the output is trusted without review, and when it is bolted onto a process that was already broken. Automating a bad workflow gets you the wrong answer faster.
It also goes wrong when nobody defines what "good" looks like, so there is no way to tell whether it is helping.
Pick one repetitive task with a clear input and a checkable output. Measure how long it takes now. Automate that one thing, keep a human reviewing it, and only widen the scope once it is boringly reliable.
Anything that reaches a customer unreviewed should be something you would be comfortable defending in public. For most businesses that means the drafting is automated and the sending is not.
The review step is not temporary scaffolding to be removed once things settle down. It is part of the design, and treating it as a phase you will grow out of is how organisations end up apologising for something a system said on their behalf.
Usage costs are easy to model and usually smaller than people fear. The real cost is maintenance: prompts drift as the underlying models change, providers deprecate things, and something that worked reliably in March behaves differently in September.
Budget for somebody to own it. Without an owner it degrades quietly, and the failure mode is not an error message — it is slightly worse output that nobody notices for a month.
Time the task before you automate it. Count how often it happens and how long it takes. Without that baseline you will end up with an impressive-looking system and no way to demonstrate it was worth building.
The measurement also tells you when to stop. Some tasks are annoying but rare, and automating them is a hobby rather than an investment.
An enquiry triage that reads the contact form, classifies the request, drafts a first reply and routes it to the right person. It does not send anything. It removes the fifteen minutes of reading and sorting that used to happen before anyone could answer, and the response time fell from two days to under four hours.
A document extractor that turns supplier PDFs into structured rows. The invoices arrive in nine different layouts from nine different suppliers, which is exactly the shape of problem that defeated the previous rules-based attempt and suits a model well. A human still approves the batch.
An internal search that answers questions from a company's own documentation with citations. The value is not the answer; it is that new staff stop interrupting senior staff, and the citations mean the answer can be checked rather than trusted.
Generating the public-facing copy for a professional services firm. The output was fluent and said nothing, which is worse than saying little clumsily, because clients recognise it. The differentiator of that business was its opinions, and a model averaged them away.
A support chatbot answering from a knowledge base that was two years out of date. The model was not the problem; it faithfully reproduced wrong information at speed. We fixed the documentation instead, and the chatbot became unnecessary.
A pricing recommendation engine on a dataset of two hundred rows. There was not enough data for any method to beat the owner's judgement, and dressing a guess in a model makes it harder to question, not easier.
Ask whether a competent new employee could do the task from written instructions in a few minutes. If yes, it is a strong candidate. If it needs years of context, judgement about relationships, or accountability for the outcome, it is not.
Ask what happens when it is confidently wrong, because it will be. If a wrong answer is caught by the next step, proceed. If a wrong answer reaches a customer, a regulator or an account, keep a human in the path.
Ask whether you can measure the before. Teams that cannot state the current cost of a task in hours cannot tell whether automating it helped, and that is how organisations end up maintaining automations that save nothing.