How I work out what AI a business needs

“We want AI” tells me almost nothing. Sometimes it’s all a business owner says when they first get in touch, and it can mean three completely different jobs.

If I guess which one and guess wrong, the business pays for the wrong project, or gets promised something nobody can deliver. So before anyone talks about tools, I work out three things: which job it is, where each part of the business is starting from, and what the owner wants AI for. Here’s how I work out each one, so you can do the same for your own business.

The three jobs hiding inside “we want AI”

The first job: they don’t know what they need yet. You’re in this one if something is costing you time or customers, but you couldn’t say what the fix is.

This is custom work, and it starts with mapping how the work really runs. I sit with the people who do it and go through their actual day using recent real cases, instead of how the process is supposed to work. Owners are usually sure they know how it runs. They usually don’t. Ask two people on the team to walk through the same case and you’ll often get two different answers.

Out of that comes a map of how everyone’s work connects, where it breaks, and where people think it breaks but it’s actually fine. Plus a baseline: how long things take now and what that costs. Then every task gets one decision. Remove it, simplify it, use a tool the business already pays for, automate it without AI, give someone an AI assistant, or hand it to an agent with a person approving what it does. A lot of the value is the list of things that don’t need automating at all, because every automation is one more thing someone has to keep working.

The business should walk away with one working pilot and a plan. A 60-page strategy deck goes on a shelf and nobody’s day changes.

The second job: they have a common problem I’ve already solved. You’re in this one if you can name the job, like following up enquiries or chasing invoices, and you know it keeps slipping.

Following up leads, turning a WhatsApp enquiry into a booking, chasing unpaid invoices, a daily summary for the owner. These come up at business after business, so about 80% of the fix is already built. The other 20% is fitting it to this business.

And that 20% is almost never the AI. It’s the data: field names that don’t match, phone numbers typed five different ways, the same customer saved three times. If it messages customers on WhatsApp, it needs the WhatsApp Business Platform, with Meta’s verification and approved message templates, because a personal WhatsApp number can’t be automated properly. Anyone selling you an “80% ready agent” that needs no setup for your business is selling you the demo.

The third job: they want the whole business running on AI. You’re in this one if you want the whole company working differently, not just you, and your team can give it the time.

That means choosing or cleaning up the systems where the real records live, writing down how things are done, connecting everything and deciding who can see what. Then an assistant goes on top, people get trained, and every part gets a named person who looks after it.

Most businesses don’t have their processes written down. Getting them out of people’s heads and onto a page is the real work, and it takes weeks of sitting with the team. Nobody gets “a company brain” in two weeks.

Each part of the business starts at its own level

The first move depends on where the records live today. There are three levels, and it’s worth being strict about which one you’re at, because the first move is different at each.

No system. Paper, WhatsApp and whatever people remember. You’re here if a new enquiry lands on someone’s phone and nowhere else. The first move is to pick one place for each kind of record, write the five most important processes down and move customer chats into one shared inbox. AI helps at this level by capturing things, like turning a voice note into a written process. It’s too early for agents, because an agent needs something written down to follow and somewhere to record what it did.

Scattered tools. Excel, WhatsApp and an app or two, with the same details typed into several places. You’re here if two people can give you two different numbers for the same thing. The first move is to decide which system is the one to trust for customers, orders and invoices, remove the duplicates and retire the side spreadsheets. Then connect them. Connect them first and you just copy the disagreement faster.

Connected. A CRM, accounting software and apps that share data. You’re here if one system can tell you where a customer stands without anyone asking around. The first move is to check who can see what. Then the AI reads the data, then assistants come in, and anything that changes a record or sends a message comes last, with a person approving it. Connected doesn’t mean clean, so I test the data anyway.

The catch is that this is per area, not per company. A restaurant group can have a connected till system, finance in Excel and HR running on WhatsApp. So I look at sales, operations, finance and people separately, and give each one its own level.

If someone wants the whole business on AI and most areas are still at no system, the first job is getting the records into one place. That comes before anything else, and it’s a bigger project than it sounds, so it has to be planned and budgeted as one.

What they want AI for is a separate question

This one is about the owner, not the systems.

Some owners are thinking long term. They don’t need a quick return. They want the business to run differently in two years, and that usually points to custom work.

Some want something that moves the needle now: more bookings, faster payment, fewer lost leads. That usually points to a proven fix. Some just want the annoying jobs gone, and those are almost always proven fixes too.

A lot want all three, and that’s fine. The reason doesn’t change the level an area is at. It helps decide which job to start with.

Where are you?

Take each area of the business and write down three things: its level, the job it needs and the first move. Here’s how it looks for the restaurant group above.

Sales and orders run on a connected till system, so that area is connected. It needs a proven fix, and the first move is a daily summary for the owner straight from the till.

Finance lives in Excel, so it’s scattered tools. It needs a proven fix too, but only once the records are sorted. The first move is to pick the one system to trust for invoices and retire the side sheets.

People and HR run on WhatsApp, so that area has no system. Its records need to go into one place first. The first move is one place for staff records and the five key processes written down.

Now do the same for your own business: sales, operations, finance and people. If you can’t place an area, leave it blank and ask the person who runs it how a real case went last week.

The knowledge base, and the mistake in my first version

Every one of these jobs needs some kind of knowledge base, not only the whole-business one. A proven fix needs its small playbook, and custom work is where the knowledge gets written down in the first place.

When I first mapped this out, I had a vector database (a search index that finds text by meaning) swallowing everything: the processes and documents, plus the live data from the CRM and the accounts. It looked tidy, one place to ask anything. It was wrong.

A business has two different kinds of knowledge. The first is how you do things: processes, prices, policies, packages, who does what. It changes slowly, so it gets written down with an owner and a last-reviewed date.

The second is what’s true right now: customers, deal stages, bookings, stock, invoices, balances, the P&L. It changes all day, and it already lives in a real system.

Numbers and statuses should never be answered from a search index. A copy goes stale the moment it’s made, and a search that matches meaning can’t add up a column. So the assistant looks them up live in the CRM or the accounts, and says where the number came from and how recent it is.

The search index is for text: documents, emails, call notes, chats. Good search adds a line of context to each piece of text before storing it, then matches both the exact words and the meaning. When Anthropic tested this for its 2024 write-up on contextual retrieval, failed look-ups fell by 49%, and by 67% when the results were also re-ranked.

Two more things I’d have got wrong without checking. Most small businesses don’t need a search index at all: if the whole playbook is under about 500 pages, Anthropic’s own advice is to give all of it to the AI directly. And in the UAE the same name gets spelled several ways (Mohammed, Muhammad, Mohamed), and many documents are in Arabic. Exact-word search misses those, so search needs a model that handles both languages and a list of name variations.

What I do on every job

Drafts first. Anything past reading, like updating the CRM or messaging a customer, starts as a draft a person approves. Once the owner has seen how often it’s right, they decide what can go out without a check.

A test before handover: 20 to 50 real questions from the business, with the right answers and where they come from, re-run after every change. That’s how you know it works, and how you know a change didn’t break it.

Upkeep. Someone has to update the processes, watch the connections and re-run the test. A knowledge base nobody looks after ends up quoting last quarter’s prices, so it needs one named person, even if others help.

BCG’s 2026 report on AI transformation puts roughly 10% of the value of AI in the algorithms, 20% in the technology and 70% in people and process. That matches everything above. The model is the easy part. Knowing which job you’re doing, and where each part of the business is starting from, is most of it.

If you’re not sure where you are

That’s normal, especially for the parts of a business you don’t run day to day. Pick one area that isn’t working and one recent real example, like an enquiry from first message to final invoice, and bring them to a working session. We’ll place that area on a level, name the job it needs and agree the first move, even if the first move is to do nothing yet.