What AI actually does inside a business
Strip away the vocabulary and almost every useful business AI system does one of four things: it reads something unstructured, classifies it, summarises it, or generates a draft of something a person would otherwise have typed.
That is a narrower list than the marketing suggests, and it is the reason the successful projects look boring from the outside. An invoice arrives as a PDF, a system reads the supplier, date, line items and total, and writes them into your accounting package. Nobody retypes anything. That is the shape of a project that works.
What AI does not do is make a decision it was never given the information to make. It cannot tell you why margins fell last quarter if that answer is not somewhere in the data you gave it. Most disappointing AI projects are disappointing for this reason rather than any failure of the model.
The three kinds of AI worth knowing about
Three categories cover essentially everything a business will be sold. They solve different problems and cost very different amounts to build.
Generative AI
Models that produce new text, images or code from a prompt. In a business it is best at drafting, summarising, rewriting and answering questions from supplied material. See what generative AI is for how it works and where it breaks down.
AI agents
A generative model wrapped in the ability to use tools, take actions and work through a task in several steps. An agent can look something up, decide what to do with the answer, and then do it. See what AI agents are.
Classical machine learning
The unfashionable category, and frequently the right answer. Forecasting demand, scoring which leads are worth calling, spotting anomalies in transactions — these are prediction problems on structured data you already hold. They are cheaper to build than anything generative, easier to explain to an auditor, and they do not invent facts.
A useful rule: if the input is a spreadsheet, you probably want classical machine learning. If the input is a document, an email or a conversation, you probably want something generative.
Where AI pays off first
The best first projects share three properties. The work is repetitive, the volume is high enough that saving minutes matters, and a wrong answer is cheap to catch and correct.
| Process | Why it works as a first project | Typical effect |
|---|---|---|
| Document and invoice data entry | High volume, rigid output, errors are visible immediately | Removes most manual keying |
| Answering routine customer enquiries | The same forty questions repeat endlessly | Handles the routine tier, escalates the rest |
| Internal knowledge lookup | Staff already waste time hunting through documents | Cuts time-to-answer sharply |
| Drafting routine correspondence | A human still reviews before it is sent | Drafting time falls, quality stays controlled |
| Categorising and routing inbound work | Clear rules, measurable accuracy | Removes a triage step entirely |
For a longer catalogue of use cases, our article on how AI can automate your business in 2026 covers ten of them in detail.
What has to be true before you start
- The information has to exist somewhere a system can reach. If the knowledge lives in one person’s head, no model can use it. Digitising the source is the actual first project.
- The process has to be stable enough to describe. You cannot automate a process that is performed differently by every member of staff.
- Somebody has to own the output. A named person who checks results and handles the edge cases. Unowned AI systems quietly drift out of use.
What AI projects cost and how long they take
Nobody publishes this, so here are honest bands. They assume integration into systems you already run — the model itself is rarely the hard part.
| Scope | Typical build | What it looks like |
|---|---|---|
| Single narrow task | 3 to 6 weeks | Document extraction into one system, or a drafting assistant for one team |
| Assistant over your own content | 6 to 10 weeks | A knowledge assistant grounded in your documents, using retrieval-augmented generation |
| Multi-step agent | 3 to 6 months | A system that looks things up, decides, and writes back to line-of-business systems |
Two costs are routinely forgotten. Running cost: model usage is charged per request. Evaluation: deciding whether the thing is actually right often takes as long to build as the feature itself.
How to choose your first project
Pick the process your staff complain about most, then check four questions. Does it happen at least daily? Is the input already digital? Can a person tell within seconds whether an answer is wrong? Would anyone notice a saving of two hours a week?
If all four are yes, that is your project. Deliberately keep the first one small. A narrow system that saves six hours a week and is trusted is worth far more than an ambitious platform that nobody uses.
Why AI projects fail
- The scope was a category, not a task. “Use AI in customer service” is not a project. “Draft first-line replies to delivery-status enquiries” is.
- The data was worse than anyone admitted. This surfaces in week two, every time.
- Nothing was integrated. A system that produces an answer a human then copies somewhere else has moved the work, not removed it.
- There was no measure of correctness. Without one, the debate about whether to expand never resolves.
- Nobody owned it. This is the most common one.
What this looks like in a mid-sized business
Take a trading or services firm with thirty to eighty people, Tally or a small ERP for accounts, and a shared drive that has become the real filing system. Two or three people spend a surprising part of the week on the same four jobs: opening supplier PDFs, answering “where is my order”, hunting for the latest version of a policy, and writing the same email in slightly different words.
None of those jobs needs a strategy consultant. They need a system that can read a document, find a record, and draft a reply a human still sends. That is usually six to ten weeks of AI integration on top of software you already run, not a new platform with a new login.
We have built the document-assistant version of this — staff asking questions of case files and getting a grounded summary instead of rereading fifty pages. The lesson was not about the model. It was about which folders were allowed to be searched, and who was responsible when the answer was incomplete.
Fitting AI onto a system you already run
The useful installations sit beside an existing application: a client portal, a booking back-office, a custom .NET line-of-business system. A button that says “draft from this record”, or a watcher on an inbox, is worth more than a separate chat site nobody opens twice.
That is also why so many pilots die. The demo runs in a notebook. The production version has to respect your logins, your roles, your audit trail, and your hosting. If the application has no API and no way to add a button, you are in legacy modernisation first. See how to add AI to an existing application for the sequence we use.
What to ask before you buy
Four questions separate a serious proposal from a slide deck. Where do the facts come from — the model, or your files? What is the system not allowed to do without a person? How will we measure that it is right, in numbers, after thirty days? Who on our side owns the exceptions?
If those answers are vague, the quote is not for a project. It is for a workshop that has not been named yet. A short paid discovery is cheaper than a six-month build that discovers the same things in week two.
Bring one real document and one real question to that meeting. Ask them to say, on the spot, what the system would retrieve and what it would refuse to do. A partner who needs a week to “prepare a workshop” on a single invoice is not slow. They are telling you the work is not a model problem. That honesty is useful. Pay for it as discovery, not as a platform.
How you know it worked
Pick one number before you start and refuse to move it. Hours spent keying invoices. Time to first reply on a routine ticket. Number of times someone asks the same person for the same policy. After a month, either the number moved or it did not. “People like the chat” is not a result.
A first ninety days that does not waste a quarter
Days 1–10: pick one task from the table above. Write the definition of done. Confirm the files are reachable and name the owner of exceptions. Days 11–45: build read-only — a draft, an extract, a cited answer. Days 46–70: real users, every result logged, one number you refuse to move. Days 71–90: either add one write that creates a draft, or stop. A platform rollout is not a phase in this plan. It is how the quarter disappears.
If the application cannot accept a button, you are in legacy modernisation first. If the process cannot be described, you are in a workshop first. Neither is a failure. Naming them is how you avoid paying build rates for discovery.
If you are deciding what is possible, start with what generative AI is and what AI agents are. If a vendor has quoted you a chatbot and you are not sure that is the job, read agents versus chatbots.