Technology
What AI can do now.
Most of what a business hears about AI is one narrow slice of it, usually a chat box. The useful ground is much wider than that, and a good deal of it has nothing to do with chat. This page sets out the whole range, what each part is for, what it costs to run, and where the honest answer is still no.
Nothing here is a product we sell. It is the map we work from when somebody asks what is possible.
The shapes it comes in
AI is not one technology.
Six quite different things travel under the same word. They fail differently, they cost differently, and they need different evidence before you trust them. Picking the wrong shape for a problem is the most common and most expensive mistake we see, and it usually happens because somebody reached for the one they had heard of.
Assistants in the tools you already have
A general assistant sitting inside the suite the organisation already runs. Drafting, summarising, searching, explaining a spreadsheet. The lowest effort to start and the easiest to waste, because buying the licence is the easy half and building the habit is the other one.
Microsoft 365 Copilot · Gemini in Google Workspace · Claude · ChatGPT
Assistants that answer from your own material
The same idea, pointed at your handbooks, policies, contracts and records instead of the open internet. The model does not memorise your documents. It is handed the relevant passages at the moment of the question and answers from them, so it can cite what it used, and so access rules still apply. This is where most of the real value sits for an organisation with a lot written down and no good way to find it.
Retrieval over your own content · Copilot Studio · Microsoft Foundry · Gemini Enterprise agent tooling · open tooling where the content cannot leave
Agents that carry out steps
An assistant answers a question. An agent is given a goal, works out the steps and uses your systems to carry them out: raising the record, sending the request, updating the row, chasing the approval. That difference is the whole reason access limits have to be settled before anything is switched on, because an agent can do things a reader could only ask about.
Power Automate and Power Apps · Google Apps Script · n8n · the Model Context Protocol, the standard way a model is given access to a tool · direct APIs and webhooks
Prediction and pattern work
Ordinary machine learning, which predates the current wave and outperforms it on a large class of problems. Forecasting demand. Scoring which cases are likely to go wrong. Spotting the transaction that does not look like the others. Grouping customers by behaviour rather than by the categories somebody invented in a meeting. It runs on your own numbers, it is cheap to serve, it can be explained to an auditor, and it does not need a language model anywhere near it.
Forecasting · classification and scoring · anomaly and fraud detection · clustering and segmentation · recommendation · survival and churn models
Language work at volume
A production line rather than a conversation. Twenty thousand documents that need a field pulled out of each. Every inbound message routed to the right team. A decade of free-text notes turned into something countable. This is where a language model does the most work without anybody opening a chat window, and where the quality can be measured, because you can take a sample and count how often it was right.
Extraction into structured fields · classification and routing · summarising at scale · translation · redaction
Speech, images and paper
Turning things that were never data into data. Recorded calls into searchable text. Scanned paper into fields. Photographs into a condition assessment. Often the highest return of anything on this page, because the alternative is somebody retyping, and nobody has ever enjoyed that job.
Transcription and diarisation · document understanding and OCR · image classification and inspection · signature and form capture
Trained and tuned models
Most work does not need a trained model. Some does.
There is a strong instinct to train something, because training sounds like the serious version. Usually it is the expensive version of a problem that a good retrieval setup solves in a fortnight. Getting the distinction right matters, because it is the difference between a two week engagement and a two quarter one.
Use a general model when
The knowledge you need is in documents you can retrieve, the task is reading and writing, and the model only has to be good at English rather than good at you. This covers most of what organisations ask for.
Train or tune when
The task is yours: your grading scale, your fault codes, your house style, your definition of a good outcome. Or the volume is high enough that a smaller specialised model is cheaper per item by an order of magnitude. Or the work has to run somewhere a general model cannot go.
Fine-tuning
Taking a capable general model and teaching it your shape of the task from your own examples. What it reliably improves is form: the model learns your format, your tone, your categories. What it does not do is install facts you can keep current, which is what people most often hope for. Facts belong in retrieval, where you can change them without retraining anything.
Small and specialised models
A large model is a general instrument. If the job is one narrow thing done a million times, a small model trained for that one thing is faster, cheaper and often more accurate on it. It will also fit on hardware you already own, which changes where it can run.
Open-weight models
Models whose weights you can download, run and keep. That matters when the data cannot leave, when the cost per item has to be predictable, or when you need the thing you validated in March to still be exactly the thing running in November. You now own the running of it, which is a real cost and should be counted before the decision rather than after it.
Classical models on your own data
For the prediction and pattern work above, the model is usually trained from scratch on your figures and is small enough to be understood. Training it is straightforward. Assembling a clean, honest history to train it on is the work, and where these projects succeed or fail.
Where it runs
Where a model runs is a governance decision.
It gets treated as an infrastructure question and settled by whoever is nearest the budget. Where a system runs decides which laws reach the data in it, who can compel access to it, what happens when the supplier changes terms, and whether you can still run it if they withdraw the service. Those are board questions wearing a technical costume.
The vendor’s service
The assistant that comes with the suite. Least effort, least control. Fine for a great deal of ordinary work, and the right starting point more often than not. The terms, the model behind it and the region it runs in are theirs to change.
Your own tenant or cloud account
The same class of model, called through an interface, inside your own subscription and your own region, with your own logging. Meaningfully more control over retention, residency and audit, at a cost you can see per request. This is where most serious work lands.
Your own hardware
Open-weight models running on machines you own, in a building you control, with no request leaving the network. It is practical now for a large class of work: a single modern accelerator will serve a good open model to a department, and the smaller specialised models will run on far less than that. You take on capacity planning, updates and the electricity bill. Nothing about the arrangement can then change without you agreeing to it.
Self-hosted open-weight models · on-premise inference servers · private networking with no public egress · local transcription and document processing
On the device
Small models running on the laptop or the handset itself. Limited in what they can do, and unbeatable when the answer has to arrive with no network and no copy made anywhere. It removes a whole category of argument about where the data went.
These are not a ladder to climb. Most organisations end up with more than one, chosen per workload: the general assistant for everyday drafting, an account-hosted model for anything touching records, and something local for the one process where the material cannot leave the building. Deciding that per workload rather than once, for everything, is most of the skill.
Underneath all of it
The model is not the hard part.
Almost every AI project that disappoints does so for a reason that has nothing to do with the model. Three things carry the weight, and none of them demonstrate well.
Data you can stand behind
Where the figures come from, what they mean, who owns the definition and which copy is the real one. An assistant asked a question about a metric with three definitions will confidently give you one of them.
Retrieval that finds the right thing
Answer quality is mostly retrieval quality. If the right paragraph is not put in front of the model, no amount of model will rescue the answer, and the failure looks like the model being wrong rather than the search being wrong.
A way of knowing it works
A set of real questions with known good answers, run again after every change, so that improvement and regression are visible. Without it you have opinions about a demo. This is the single most skipped step in the field.
Anything that cannot be measured cannot be improved, and should not be trusted with a decision that matters.
The estate
Nobody runs one vendor. That is the whole problem.
Email is on Microsoft. Finance and HR are on Oracle. The line of business system came with the industry and nobody outside it has heard of it. One department runs on Google. And there are now three different AI assistants in the building because three different people each signed up for one. That is the normal state of a real organisation. A rule that only holds inside one vendor’s tools is not a rule.
So we work across all of it. Where we recommend something we will tell you what we rejected and why, and where one of your existing licences already covers it, we will say so.
Workplace suites
The suite the organisation already runs. We build inside it, not beside it, so the data stays where its governance already sits.
Microsoft 365 · Google Workspace · Teams and Slack · Outlook and Exchange · SharePoint and Google Drive
Business systems
Where the records live, and the oldest and least negotiable part of the estate. We do not replace these, and if someone opens with a proposal to, treat that as a reason for scepticism. The useful work is nearly always getting data out of them safely and putting the right thing back.
Oracle Fusion Cloud ERP and HCM · Oracle E-Business Suite · NetSuite · Dynamics 365 · Salesforce · HubSpot · Sage, Xero and QuickBooks · the industry-specific system nobody outside your field has heard of
Data and reporting
Where figures are derived and packs are assembled. Definitions drift at this layer. It is where one metric becomes three.
Oracle Database and Oracle Analytics · SQL Server · PostgreSQL · Microsoft Fabric and Power BI · BigQuery · Excel, which is still doing more of this than anyone admits
Identity, access and governance
Who someone is and what that entitles them to reach. Almost every AI problem turns out to be partly a question at this layer, whatever it looked like at the start.
Microsoft Entra ID · Google Cloud Identity · Microsoft Intune · Microsoft Purview · single sign-on and multi-factor · logging, backup and retention settings
Where it all runs
Underneath everything, and the layer most often assumed to be somebody else’s problem.
Microsoft Azure · Oracle Cloud Infrastructure · Google Cloud · AWS · your own servers · the machine in the cupboard that everyone forgot is still running something important
Governance is not one of these layers. It runs through all of them.
A rule that lives only in the assistant layer is bypassed at the file layer. A retention policy that lives only in a document is contradicted by a backup nobody configured. A consent recorded in one system means nothing if another holds a second copy of the same person under a different identifier. The work is making one set of decisions hold consistently everywhere they apply, which is not something any single vendor’s tooling can do for you.
The other half of the map
What is oversold, and what is not there yet.
A page about what is possible is worth very little without this section. These are the things we are asked for most often and decline or qualify most often.
Decisions about people, made automatically
Sifting applications, ranking staff, deciding who gets something. The technology will do it. The law is specific about automated decisions with a significant effect on someone, the failure modes are quiet and systematic, and the organisation carries the consequence. We will help you build the assistive version that puts evidence in front of a person, and we will say no to the version that replaces the person.
An agent let loose across the estate
The demonstrations are impressive. The failure mode is that it is confidently wrong at speed, in systems where being wrong is expensive. Agents belong inside a bounded process where the steps are known, the permissions are narrow and there is a record of what was done. That is a real and useful thing. It is not the same as handing one the keys.
Replacing a system of record
Nothing on this page is a reason to move off a working finance or case system. The gain is nearly always in getting information in and out of it more sensibly, and a proposal that starts with replacement should be read very carefully.
Accuracy nobody has defined
Asked how accurate a thing will be, the honest answer before any measurement is that we do not know, and neither does anyone else quoting you a figure. What we can say is how we will find out, what a good enough result looks like for your process, and what happens on the cases where it is not good enough. If nobody has answered that last question, the project is not ready.
None of this is caution for its own sake. The point of knowing where the edge is, is being able to work confidently everywhere inside it.
Not sure which of these fits your problem?
That is the usual position, and a reasonable one. Describe the process in your own words and we will tell you which shape it is, what it would take, and whether it is worth doing at all.