Three time horizons

Conversations about AI mix up three different things: what works today, what is emerging, and what is speculation.

Keeping them separate makes planning much easier. This article takes each in turn.

Available now

These are in Odoo today and businesses are using them.

Document reading. Scanned vendor bills, receipts and purchase orders are read and turned into draft records. The vendor, dates, amounts and line items are extracted. A person reviews and posts.

Text drafting. Product descriptions, email replies, quotation notes, knowledge base articles. Drafted in the screen you are already in.

Record summarising. A long email thread or a customer’s history condensed into something readable before a call.

Assisted search and questions. Asking about your own data in plain language rather than building a report.

Automation of repetitive steps. Suggesting the next action, filling fields, routing records.

The pattern in that list: every one is a task that is repetitive, high volume, and easy to check. That is not a coincidence. It is where AI is genuinely good.

Emerging

Real but less mature. Value depends heavily on your data.

Demand forecasting. Predicting what you will need based on history. Works with years of clean data and stable patterns. Produces false precision on new products or volatile markets.

Lead scoring. Ranking opportunities by likelihood to close. Needs enough closed-won and closed-lost history to learn from — a few hundred deals is not enough.

Anomaly detection. Flagging unusual transactions. Promising, and prone to false positives early on while it learns what normal looks like for you.

Payment prediction. Estimating which invoices will be paid late, so collections effort goes where it matters.

For all four, the honest position is the same: useful as one input alongside judgement. Not as a decision.

Speculation

Worth knowing about, not worth planning around.

Autonomous agents. Systems that carry out multi-step tasks on their own — investigating, deciding and acting. Genuine progress is happening here. In a business system where money moves, the exceptions are expensive, and most companies will want a person in the loop for some time yet.

Fully conversational ERP. Running your business by talking to it. Parts of this work now. All of it, reliably, does not.

Self-configuring systems. Software that sets itself up for your business. Configuration is a series of business decisions — your chart of accounts, your approval rules — and those are yours to make.

The structural advantage

There is a reason AI inside an ERP behaves differently from AI attached to a single tool.

Context.

An AI on your email sees emails. An AI on your accounting package sees transactions. An AI inside Odoo sees the customer’s orders, their payment history, their open tickets, their credit position and their delivery record — because it is all one database.

For most useful questions, that context is the difference between a plausible answer and a correct one.

It also matters practically. AI in a separate tool means another subscription, another integration and another thing that breaks at upgrade time. AI in the system you already use, working on data it already has, is a much lower-friction proposition.

What actually decides your outcome

Not which features you switch on. Three other things.

Your data quality

AI amplifies what is already there. Companies with clean master data get value in weeks. Companies with duplicate customers, wrong product costs and inconsistent units get confident nonsense.

This is the single biggest predictor, and it is unglamorous work.

Where you keep humans

Draft-and-review works. Automatic posting without review is where companies get hurt — not often, but expensively when it happens.

A sensible rule: wherever money moves or a customer sees the output, a person checks.

Whether you measure it

Before switching something on, write down what it should improve and how you will know. Hours saved on bill entry. Days off your collection cycle.

Without a baseline, you cannot tell whether it worked, and you will keep paying for things that do not.

A sensible plan

Four steps, in order.

1. Fix your data. Duplicates merged, costs correct, units consistent. Nothing else works properly without this.

2. Start with document reading. Highest volume, easiest to verify, fastest payback. Measure the hours before and after.

3. Add drafting where it saves real time. Descriptions, replies, summaries. Keep the review step.

4. Pilot prediction carefully. One area, alongside your existing judgement, for a few months. Compare its calls against what actually happened before you rely on it.

What not to do

Do not buy AI as a category. Buy a specific capability that addresses a specific cost you can name.

Do not automate a broken process. It will run faster in the same wrong direction.

Do not remove the human from anything financial because the demo was impressive.

Do not assume it will improve your reporting. It reports what is there. If what is there is wrong, you now have a faster route to a wrong conclusion.

The short version

The near future is unglamorous and genuinely valuable: less typing, faster drafting, better-ordered attention.

The far future is interesting and not yet something to plan around.

The businesses that will benefit most are not the ones adopting fastest. They are the ones with clean data, a clear process, and a person still checking the things that matter.