AI for business: what actually changes, and what it costs

Where AI actually changes a business

Not evenly, and not where the demos suggest. The work that moves has three properties: it repeats, the inputs are text or data rather than judgement about people, and somebody can tell quickly whether the output is right.

Run your own processes against those three and the list shortens usefully. What survives, in roughly the order businesses see returns:

• Inbound enquiries. The same twenty questions, asked constantly, answerable from documents you already have. Highest return, lowest risk, and the answer quality is immediately visible.
• Administrative movement. Copying an enquiry into the CRM, the calendar, the invoice, the notes. Nobody's job satisfaction depends on it and everybody's week does.
• First drafts. Replies, reports, proposals, summaries. The draft is the expensive part; the edit is fast.
• Monitoring and recurring reporting. Work that ought to happen weekly and happens when someone remembers.
• Research. Reading a lot and producing something short — competitors, suppliers, prospects, a market.

And what does not move: pricing decisions, hiring, anything where being confidently wrong is expensive and hard to verify, and relationships. Those stay yours.

Telling a real opportunity from an expensive one

Four properties make a workflow a good first candidate:

• You can put a number on today. Volume, hours, response time, enquiries missed. Without a before, there is no after.
• It is bounded. A beginning and an end, not "handle the inbox".
• Errors are visible and cheap. You find out quickly and fixing it costs an apology, not a customer.
• It is already documented, or could be in an afternoon. If nobody can write down how it works, an agent cannot follow it either.

Three properties make an obvious-looking one a trap: it depends on tacit knowledge nobody has written down; it touches money or identity; or nobody has time to review the output, which means the approval boundary will be abandoned in week two. Where to start works this through properly.

What it costs

Three costs, and businesses routinely plan for one:

• The platform or infrastructure. The visible line item, and usually the smallest.
• Model usage. Consumption-based, and it moves with how much work you hand over — which is why hard spending caps matter more than a low headline price.
• Attention. Somebody has to choose the process, own the integration, decide the boundary, review the output for the first few weeks, and adjust it when the business changes. This is the real cost and the one that sinks deployments.

That third cost also explains the adoption gap between large and small firms. It does not scale down: a big company can make it someone's job, a small one lands it on an owner who already has three. It is the main argument for a managed engagement over a build-and-hand-over.

Measuring whether it worked

Use numbers you already track, not vendor metrics. Three that are hard to argue with:

• Response time on inbound, including out of hours. Measurable this week, and AI either changes it or does not.
• Hours your most expensive person spends on work that does not need them. The actual return, denominated in attention.
• Things that used to fall through. Unanswered enquiries, uncontacted leads, reports that did not get written.

Ignore messages handled, tokens consumed, and time saved in the abstract. Measuring AI ROI goes further.

The governance question you cannot skip

The moment AI acts rather than suggests, four things need deciding: what it may do unsupervised, what it does when unsure, how you reconstruct what it did, and who is accountable when it is wrong. These are decisions, not features — and they are reusable, so the second workflow is much faster than the first.

Get them right and unattended work is comfortable. Skip them and you have an unaudited actor in your business.

A reasonable first ninety days

• Weeks 1–2. Pick one workflow. Write down what it costs today. Write down how it is supposed to work.
• Weeks 3–4. Deploy it narrow: reads autonomous, anything outbound gated. Review every output.
• Weeks 5–8. Widen the boundary where the evidence supports it. Capture the procedures you keep correcting.
• Weeks 9–12. Measure against your before number. Then either scale to a second workflow or stop — both are legitimate results.

Then the compounding starts, because the governance decisions and the written procedures carry over. Improving efficiency with AI is the arithmetic; AI agents is what is doing the work.