What AI Can Actually Do
for Your Business
Function by function, with the uses that work today, the ones that sound good and disappoint, and what still needs a person in the loop.
Quick Take
The pattern repeats across every function. AI is good at producing a first version of something a person then checks, and at pulling structure out of unstructured material. It is poor at anything that has to be right without anybody looking. Read the section for your function, pick one task, and test it properly using the ninety-day method.
Admin and Back Office
Where most SMEs find their first genuine win, and the least exciting section of this page.
Works today. Pulling data out of invoices, delivery notes, forms and PDFs into a spreadsheet or system. Drafting routine correspondence. Summarising long documents so somebody can decide faster. Turning meeting notes into actions. Filling in the first version of repetitive paperwork.
What it saves. The gains here are measured in hours per week per person on tasks nobody enjoys. That sounds modest until you count the roles doing it.
Still needs a person. Anything where the extracted number goes straight into an account, a payment or a filing without review. Build the check into the process and the saving survives; skip it and you have automated the production of plausible errors.
The catch nobody mentions. If your paperwork arrives as scans of scans, or lives in an old system that will not let anything out, the AI is not your problem. That is a legacy systems project wearing an AI costume, and it is worth knowing that before you buy a licence.
Customer Enquiries and Support
The most oversold function on this page, and still a real one if you are careful.
Works today. Sorting and routing incoming enquiries. Drafting replies for an agent to approve. Answering internal questions from your own manuals and policies so staff stop interrupting the one person who knows. Summarising a long ticket history before somebody picks it up.
What it saves. Response times, mostly, and the cognitive load of context-switching. Deflection rates quoted by vendors are usually measured on question sets that flatter the product.
Still needs a person. Anything a customer would escalate: complaints, refunds, contract terms, anything with a number in it that they might hold you to. A confident wrong answer to a customer costs more than the ten minutes it saved.
Be honest about the goal. If the plan is to reduce headcount, say so internally, because your staff will work it out within a fortnight and the pilot will quietly fail. If the plan is to make the same team faster, that is a much easier project to run and a much more reliable saving.
Sales and Proposals
Works today. First drafts of proposals and tender responses assembled from material you have already written. Researching a prospect before a call. Turning a call recording into notes and next steps. Rewriting the same case study for different audiences.
What it saves. Proposal turnaround, which in some businesses is the difference between bidding and not bidding.
Still needs a person. Pricing, commitments, dates, anything contractual. The judgement about whether to bid at all. And the last read before it goes out, because the failure mode here is fluent, generic and instantly recognisable to the reader.
Watch for. Outbound messaging generated at volume. It works until everybody does it, which happened some time ago. The businesses winning here use it to answer better, not to send more.
Finance and Reporting
Works today. Reading invoices and receipts into your accounting system. Flagging anomalies for a human to look at. Drafting the commentary that goes around a set of numbers. Answering questions about your own historic reports.
Still needs a person. The numbers themselves. Use AI to find what is worth looking at and to write about the results; do not use it to calculate them. Anything going to a bank, an investor, HMRC or a regulator gets checked by somebody accountable, every time.
Regulated sectors. If you are in financial services, the constraints change the design rather than the ambition. Our fintech technology advisory covers what that looks like in practice.
Operations and Scheduling
Works today. Drafting documentation and procedures. Making sense of free-text fields nobody has ever been able to report on. Summarising incident reports and site notes. First-pass classification of anything that arrives as a pile of text.
Usually oversold. Forecasting and optimisation. These are real disciplines with decades of established technique behind them, and the honest answer is that a well-built model or even a good spreadsheet often beats a language model at them. If a vendor pitches AI scheduling, ask what the underlying method is.
Still needs a person. Anything touching safety, compliance sign-off, or a commitment to a customer about when something will happen.
Engineering and Product
If you employ developers, this is the function with the most evidence behind it and the most nuance.
AI coding tools are close to universal in professional development now, and the gains are real but narrower than the marketing implies. The constraint has moved from writing code to reviewing and trusting it, which lands the cost on your senior engineers.
We have written this up properly rather than summarising it here: how to introduce AI coding tools to an engineering team, how to measure whether the tooling pays for itself, and how to review what comes out. The commercial version is our software development advisory.
If the question is instead whether AI should be a feature in your product, that is a build-or-buy decision before it is an AI decision. The build vs buy framework applies unchanged, and custom versus off-the-shelf covers the same ground for non-technical readers.
The Test That Applies to All of Them
Whatever the function, the candidate task should look like this: it involves reading, sorting, drafting or summarising rather than deciding; a competent person could do it with no special training given the hours; a wrong answer is annoying rather than dangerous and somebody would notice; it happens often enough that ten minutes matters; and the information needed already exists somewhere you control.
Score your candidates against those five and the shortlist writes itself. Then work out what it will actually cost, check that you are comfortable with where the data goes, and run one test properly.
Not Sure Which Function to Start With?
Tell us how the business runs and we will point at the two or three worth testing, or tell you there are none yet.